After decades of neglect, Washington has adopted a coherent and “forceful” response to China’s dominance over critical minerals. China controls an estimated 91 percent of global rare earth refining, dominates the processing of a wide range of minerals, and has demonstrated repeatedly that it is willing to weaponise that position. The United States has responded through a range of initiatives and policy packages including Executive Order 14241, the One Big Beautiful Bill Act, and Pax Silica. Add bilateral deals with Australia, Japan, the DRC, and Saudi Arabia, and the result is undeniably a more aggressive and cohesive mineral strategy than the United States has attempted in decades.

The problem is not with any of these agreements and frameworks. Instead, it is with what is systematically absent from them. The entire architecture of American mineral strategy, and indeed the EU’s parallel effort under its Critical Raw Materials Act and RESourceEU, is built around two variables: where minerals are located and who processes them. What it consistently underweights is a third variable that will most likely prove equally consequential over the medium term: who controls the technology that operates the mines themselves. This constitutes the central strategic blind spot in the West’s mineral autonomy agenda because China is not just dominant in mineral processing. Beijing is also rapidly establishing dominance in the operational technology of extraction; the autonomous systems, industrial AI, connectivity infrastructure, and integrated machine intelligence that will define how the world’s future mines will run. Today, we’ll explore China’s dominance in autonomous mining technology and what to do about it.

If that dominance goes unchecked, the US and its allies could succeed in securing access to mineral deposits across friendly jurisdictions and still find those mines dependent on Chinese operational technology to function.

The Second-Order Competition

To establish both how and why the technology of mining matters strategically, it helps to unpack how the mining landscape is actually evolving. To this end, the departing point is to acknowledge the obvious: mining has always been a risk-prone and controversial venture, with the industry’s reputation tainted by concerns ranging from worker welfare to environmental destruction. The industry therefore has ample incentives to develop alternative extraction methods to improve its image and reduce liability.

Contemporary autonomous mining goes much further than replacing single tasks with machines. It is the deployment of an integrated stack: autonomous haulage systems that navigate dynamically without human drivers; AI models that manage geological analysis, fleet routing, predictive maintenance, and slope stability monitoring simultaneously; sensor networks providing real-time environmental data; edge computing that maintains operational integrity when wide-area connectivity fails; and satellite positioning that substitutes for GPS in remote and underground environments. When this stack is fully deployed, a mine managed remotely by a small supervisory workforce can operate almost uninterruptedly in conditions that would make human presence hazardous.

The productivity gains and safety improvements are both substantial. For governments and nation states, however, the more consequential dimension is strategic; that is, the capabilities required to run a mine autonomously in harsh and remote environments are directly transferable to deep-sea operations, space, and military logistics in contested or degraded operating conditions.

China understands this explicitly. Its Arctic and Antarctic deployments are framed in official strategic communications as technology demonstration environments since systems proven in polar conditions are, more or less, transferable to outer space and the deep sea. A case in point is the Yimin open-pit coal mine 伊敏露天矿 in Inner Mongolia, where a fleet of fully electric autonomous trucks has been operating in temperatures as low as -48 degrees Celsius.

The Yimin Benchmark

The Yimin deployment warrants a closer look because it is the clearest evidence available of where China currently stands. The fleet has been developed jointly by China Huaneng Group 中国华能集团, Huawei, XCMG 徐工集团, and State Grid Smart Internet of Vehicles 国网智慧车联网技术. Each 45-tonne truck runs on a 568 kWh lithium iron phosphate battery with the driver’s cab replaced entirely by a battery pack. Five cameras, LiDAR units, and millimetre-wave radar provide 360-degree field of vision up to 40 metres in complete darkness, dust storms, and blizzards. Automated battery-swap stations reduce downtime to approximately five minutes while eight hundred sensors monitor the site continuously. By 2024 — prior to full fleet deployment — the autonomous trucks were reportedly achieving 120 percent of human-operator productivity.

The connectivity layer is as significant as the hardware. Huawei’s 5G-Advanced network provides sub-20-millisecond latency across nearly 200 kilometres of mine routes. Huawei’s Pangu 5.5 盘古大模型 5.5 — a 718-billion-parameter industrial AI model deployed across more than 500 scenarios in over 30 industries — manages fleet coordination, geological monitoring, and predictive maintenance, while BeiDou satellite positioning 北斗卫星导航系统 substitutes for GPS.

The most important strategic revelation is that every component of the operational technology stack is Chinese. Equally significant is the fact that Pangu 5.5 is not a domestic product for Chinese mines alone; it is explicitly positioned as a global industrial AI platform. In addition, battery-powered autonomous haulage, 72% of whose global fleet currently operates in China, could likely follow the same commercialisation trajectory as Chinese electric vehicles: cost-competitive, technically capable, and backed by state subsidies that allow pricing strategies unavailable to Western competitors. As these systems become commercially available to mine operators worldwide, adoption creates operational dependency. A mine that runs on the Chinese operational technology stack and Chinese autonomous vehicle systems is not independent of China in any meaningful sense, regardless of who holds the mineral licence. And this is the competition that Washington’s current mineral strategy has neglected.

The Arctic as the Sharpest Case Study

No region makes this argument more clearly than the Arctic. The Arctic holds substantial deposits of rare earth elements, nickel, cobalt, copper, and other critical materials within the territory of NATO-aligned democracies: Sweden, Finland, Norway, and Denmark (via Greenland). The difficulty is that Arctic mineral potentials cannot be fully unlocked by the conventional mining model. Its operating constraints — extreme cold, lack of adequate infrastructure and small local workforce to draw on — make the labor-intensive extraction model that has governed the industry for generations structurally unviable. As such, it ought not to be surprising that mining operations in the region are typically much more capital intensive compared to temperate projects before accounting for the full social overhead of human presence. Greenland, to take the most prominent example, is three times the size of Texas with just 93 miles of paved road, and ports whose combined throughput would be exhausted by a single active mine at scale.

The only operational model with a high chance of making Arctic mining commercially viable is one built around autonomous operations: AI-driven systems that manage extraction with minimal continuous human presence and monitored remotely through satellite-enabled connectivity. This, to a large extent, is synonymous with the current operational logic being applied at Yimin in conditions directly comparable to much of the Arctic. And it means that unlocking the Arctic’s mineral potential — which could be central to any credible Western strategy for attaining a sufficient degree of mineral resiliency — requires mastering precisely the technology in which China currently leads. Hence, if the US and its allies do not close this technology gap, securing Arctic mineral access may become an exercise in building infrastructure that would ultimately be operated on Chinese terms.

The Arctic’s value in this framing, moreover, extends beyond terrestrial extraction alone. As the closest accessible analogue to space operating conditions, it is also the natural proving ground for the autonomous systems that US-led space mining ventures are developing for deployment beyond Earth’s atmosphere (more on this later).

US Partners Are Already Moving

For their part, European countries (specifically Sweden and Finland), are building a genuine cluster of autonomous mining technology that is among the most advanced in the Western world. These efforts, however, remain underfunded relative to the problem, poorly coordinated at the EU level, and largely invisible in the strategic debate in Washington.

Sweden’s LKAB is running one of the most ambitious underground mine automation programmes in the world. The Sustainable Underground Mining consortium, which brings together LKAB, ABB, Epiroc, Combitech, and Volvo at the Konsuln test mine in Kiruna, is explicitly targeting a new global standard for sustainable, autonomous, carbon-free underground mining. LKAB has also deployed Sandvik’s AutoMine autonomous system across a growing fleet of electric loaders at Kiruna. At Boliden’s Aitik mine, moreover, Komatsu’s FrontRunner autonomous haulage system is operating at commercial scale; the first such deployment in Europe and in an Arctic environment. These operational deployments matter because they are generating actual performance data in the specific environmental conditions that broader Arctic mineral development requires.

Finland’s latest National Mineral Strategy, on the other hand, explicitly connects technology leadership to mineral security objectives, and thus it ought not to be surprising that Helsinki has taken the lead in solving the foundational connectivity problem that autonomous Arctic mining depends on. Its 6G Flagship program is amongst the first government-funded initiatives in the world to explicitly investigate how next-generation millimetre-wave frequencies behave inside mine tunnels; a prerequisite technical problem for full underground autonomy. Equally significant, Nokia and Boldyn Networks have built an underground 5G test network at the Pyhäsalmi mine.

Australia has deployed over 927 autonomous haulage trucks across its mines backed by the world’s first regulatory Code of Practice for Safe Autonomous Mining and a federal government investment of AU$566 million in geological mapping and mining technology infrastructure. Australian operators have, arguably, more real-world operational data on autonomous large-scale mining than any non-Chinese actor.

The picture that emerges is of a concrete, distributed Western capability in autonomous mining technology; assets in the right jurisdictions generating operational knowledge in the right conditions which have not been assembled into a competitive program. Sweden and Finland are building the technology, while Australia has the largest deployment base and is pioneering a regulatory framework for autonomous mining. The United States has the financial instruments, the political leverage, and, through the Mine of the Future initiative, the beginnings of a federal research commitment. But the West still lacks a coordinating structure that would make these complementary strengths add up to a coherent challenge to China’s lead.

The Alliance Washington Has not Built but Urgently Needs

This is the new strategic argument that the current US minerals agenda is missing. Pax Silica is structured around a coalition of mineral capabilities; it pairs deposits in one jurisdiction with processing expertise and financial resources from other allies. It is, in its logic, a supply chain coalition. What’s more, the DOE’s Mine of the Future initiative, committing up to $95 million for autonomous mining technology proving grounds and national laboratory research, is a good start but insufficient on its own. The gap between what the US is currently investing in mining technology and what China’s National Development and Reform Commission (NDRC) directed programme has already deployed is wide enough that domestic investment alone will not close it on the relevant timeline. This is an area where allied coordination can become a force multiplier which the US cannot afford to forgo.

What is needed, therefore, is a techno-industrial alliance specifically focused on autonomous mining: a coordinated programme to master, standardise, and commercially scale the operational technology of next-generation extraction designed explicitly to outcompete China. Sweden and Finland bring the most advanced European capacity in underground autonomous systems, the connectivity research those systems require, and Arctic operational environments where the technology can be tested under relevant conditions. Australia brings the largest deployed autonomous fleet in the Western world, a mature regulatory framework, and demonstrated experience in scaling these systems commercially. The US brings federal research funding, the world’s largest defence-industrial complex with direct interest in the dual-use applications of these technologies, and the downstream manufacturing demand that gives the coalition its market rationale.

The structure of such a partnership, however, would need to be more than a procurement framework. It would need to coordinate research and development by marrying Finland’s 6G underground work with US national laboratory capacity, sharing operational data from Australian deployments with Swedish underground test environments, jointly developing the safety certification standards that would allow autonomous equipment to be deployed across allied jurisdictions without re-testing from scratch. It would also need a shared approach to the connectivity infrastructure that autonomous mining — especially in harsh and remote environments like the Arctic — requires. Above all, it would need to move at the speed of a technological revolution rather than a multilateral treaty formation.

Mining Technology as AI Ecosystem Strategy

Crucially, the technology of mining is also a vector for AI ecosystem expansion, and the US should be thinking about it in those terms.

When Chinese companies embed Huawei’s Pangu AI, BeiDou positioning, and 5G-Advanced connectivity in a mine in Inner Mongolia — or eventually in a mineral-rich country in Africa or Central Asia — they are integrating that mine’s operational environment into the Chinese technological ecosystem. Local engineers are trained on Chinese platforms, maintenance contracts create ongoing relationships with Chinese vendors, and the data generated by mine operations flows through Chinese systems. As such, switching to a different technology stack at a later date would amount to an operational disruption that mine operators would rationally avoid. This is the same lock-in logic that has made Chinese telecommunications infrastructure so strategically consequential in countries that have adopted it.

The inverse logic applies to US-led mining technology partnerships. If American AI systems, US-certified autonomous equipment, and allied connectivity standards become the operational backbone of mines in supplier countries, those countries’ mining sectors become embedded in the US-aligned technology ecosystem in institutionally durable ways. A mine operator whose engineers train on US autonomous systems, and whose operational data is managed by US-allied AI platforms, faces a fundamentally different calculus when Chinese vendors offer competitive pricing. Put differently, A US-led autonomous mining technology program has the potential to shape technology adoption decisions and expand the US AI ecosystem in mineral-supplying countries.

Equally important is the potential goldmine of data. Autonomous mining systems are extraordinary data generators. Every operational shift produces a continuous stream of geological readings, equipment telemetry, fleet routing decisions, slope stability measurements, and environmental indicators. The AI models that manage these systems improve — in predictive accuracy, in decision quality, in operational efficiency — proportionately to the volume and diversity of data upon which they are trained.

Through its state-coordinated intelligent mine programme, China is building what amounts to the world’s largest training corpus for mining AI. Nearly 1,000 intelligent mines, coordinated under a unified national program, are generating operational data at scale and in a form that continuously feeds back into the AI models managing them. In this context, the parallel to another strategic domain is instructive. As the National Security Commission on Emerging Biotechnology has warned with respect to biodata, the United States risks a future in which an adversary controls the data infrastructure underlying strategic industrial sectors because that adversary has built tightly integrated, state-controlled ecosystems in which AI and operational data co-evolve, while the US lacks an equivalent unified infrastructure. The warning applies with equal force to mining. Autonomous mining AI trained on Yimin’s operational data, and on the hundreds of similar deployments across China’s intelligent mine programme, will over time outperform systems trained on smaller, siloed, proprietary Western datasets. That performance gap, in turn, translates directly into adoption incentives: mine operators globally will gravitate toward the systems that work best, regardless of their origin.

Share

The United States and its allies currently lack a mineral data commons mainly because there is a lack of recognition that mineral operational data is a strategic asset in the same category as the mines themselves. Operational data from Australia’s fleet of autonomous trucks, from Sweden’s Kiruna test mine, and from deployments across allied jurisdictions sit in proprietary silos, owned by individual mining companies with neither the incentive to share it nor a mechanism to aggregate it for AI training purposes. The raw material for building competitive AI exists but the infrastructure to convert it into model advantage does not. A techno-industrial autonomous mining alliance should therefore include a shared and secure mineral data infrastructure to enable operational data from allied mining deployments to be aggregated, standardised, and used to train Western mining AI systems.

Space Extraction as a Strategic Multiplier

The final strategic dimension here is the direct applicability of autonomous extraction capabilities to the emerging space and asteroid mining sector, and the fact that, unlike autonomous terrestrial mining where China has established a commanding lead, this sector is dominated by American companies.

Asteroids harbor the same class of metals at the heart of today’s mineral security agenda. Valued at current commodity prices, the mineral wealth of near-Earth asteroids runs into the quintillions of US dollars, and individual extraction missions could return trillions in value. The commercial logic has already generated a cohort of US-headquartered ventures developing autonomous extraction systems capable of operating in the vacuum, extreme cold, and communication constraints of space, where a continuous human workforce is not an option.

At the level of operational technology, space extraction and autonomous Arctic mining present strikingly similar challenges. For instance, both demand extraction in the absence of a continuous human workforce and impose extreme demands on robotics, power management, and autonomous systems under degraded connectivity. The Arctic is therefore the natural proving ground where space mining systems could be tested and refined before deployment beyond Earth’s atmosphere. Directing these ventures to Arctic testing environments would simultaneously accelerate their development timelines and expand the operational dataset that Western terrestrial mining systems require, generating performance data in harsh autonomous operations that benefits both sectors.

Equally significant is the fact that the talent, technology, and entrepreneurial capacity concentrated in these US companies represents a latent competitive asset in the broader autonomous extraction technology competition. China’s space mining programme remains at an early stage and lacks an equivalent cluster of commercially-driven ventures operating at pace; thus far, only one Chinese company — Origin.Space 起源太空 — is active in the field. Hence, the primary constraint is probably not a lack of technological capability, but rather the absence of an integrated policy framework that recognises the strategic synergies between the mining and space sectors.

America First Is the Wrong Gear for This Race

The technology alliance proposed here is structurally incompatible with a transactional, leverage-first approach to alliances.

Sweden and Finland are not large economies. They do not offer the strategic weight that compels attention in a framework built around bilateral leverage and deal-making. They are, however, the countries running the most advanced autonomous underground mining programmes and in Arctic conditions that are directly relevant to the Greenlandic deposits the US is reportedly eyeing to develop. Treating them primarily as junior partners in a security arrangement, or as potential targets of tariff pressure in broader trade negotiations, is precisely the wrong dynamic for the kind of deep technical collaboration, such as shared R&D, joint certification standards, open operational data, that a competitive autonomous mining programme requires.

Australia is a closer case. The existing minerals partnership establishes a basis for exactly this kind of technical cooperation, but even with Australia, the logic of “what does the US get” from a bilateral deal is the wrong frame. The question is what the coalition gets from pooling capabilities that are, individually, insufficient to match China’s scale and integration. That question requires a different kind of strategic reasoning than the one embedded in an America First approach; one that treats allied technical assets as force multipliers rather than as leverage points in a negotiation.

More broadly, coalition frameworks require predictability and institutional trust to function. Deep technical collaboration requires relationships built on something more durable than transactional benefit calculations. The allies who hold the autonomous mining capabilities that the US needs to complement its own have domestic political systems that are responsive to how they are treated. Imposing tariffs on them and simultaneously expecting them to co-develop autonomous mining AI constitutes an unwarranted contradiction that will not resolve in Washington’s favor.

The Reframing Washington Needs

The US minerals strategy has achieved important wins in a domain that was ignored for far too long. However, the overall architecture is incomplete in a way that compounds over time.

The technology of mining, the autonomous systems, industrial AI, connectivity infrastructure, and the operational data those systems generate are strategic domains as consequential as processing capacity. Not only do they carry dual-use applications across deep-sea, space, and defence logistics, but their development through allied partnerships is also a mechanism for AI ecosystem expansion; that is, for embedding US-aligned technology and data standards in the operational infrastructure of the countries that supply tomorrow’s critical materials. None of these dimensions are captured by the US’s current framework.

What is missing from Washington’s agenda is an explicit techno-industrial alliance for autonomous mining; a programme that would pair US research investment and downstream demand with Europe’s evolving underground automation expertise, Australia’s experimentation with large-scale autonomous mining, a shared mineral data commons for AI training, and a joint commitment to the regulatory and technical standards that the next generation of mining will run on. This techno-industrial alliance should also be designed to bridge the mining and space sectors by directing US companies developing space-mining technologies towards Arctic deployment while integrating the autonomous extraction expertise concentrated in American space ventures into the alliance’s core capabilities. Failure to do so would most probably lead to a minerals strategy that secures access to mining sites and loses the machines. That, in turn, is not a strategy for mineral autonomy but simply a more expensive form of the same dependency on China.

Check out Nima’s previous guest post on ChinaTalk here:

The Trump administration has argued that the crisis was contained thanks to their aggressive action, but they may be taking the wrong lessons from the avoidance of that apocalyptic scenario.

What happened was the largest unexpected swing in global oil balances: China quietly cut crude oil imports by more than five million barrels a day. Yet there was no corresponding collapse in economic activity, no obvious drop in mobility, and no official explanation from Beijing. Somehow, China stopped buying oil from the rest of the world and started drawing from stockpiles we can’t fully observe.

That single decision may have done more to prevent a global energy crisis than anything Washington or OPEC accomplished.

This matters because it demonstrates that China likely has a stronger discretionary policy lever than the West does. The West is really good at market-driven, private-sector oil production. But through this crisis, we’ve seen that Washington does not have the scale of discretionary policy control that China or OPEC does. This time, China cooperated and did the good thing, at least for the broad economic picture — but we cannot rely on that in the future, and that tool can be used against the West as easily as for it. Western governments must grapple with that discretionary gap and not rest on their private-sector bona fides to get through the next crisis.

Arnab Datta, managing director of policy implementation at Employ America, and Rory Johnston, oil analyst and founder of Commodity Context, join ChinaTalk to discuss:

• Why Rory’s own prediction of $200 oil never happened and why J.D. Vance is thanking the wrong people.

• How China quietly cut crude imports by five million barrels a day with zero visible impact on domestic mobility, and the detective work analysts use to peer into Beijing’s black-box inventories.

• Competing theories for why Beijing backstopped the global oil market — self-interested altruism, a backroom deal during the state visit, or a dry run for a Malacca blockade in the event of a Taiwan contingency.

• What India, the Gulf states, and the rest of the world learned from the Iran War and why strategic reserves are suddenly back in fashion.

Listen now on your favorite podcast app.

The Dog That Didn’t Bark

Jordan Schneider: We had this whole Iran war, and there was a moment when people were worried about $700 barrels of oil — complete, utter calamity. While prices certainly went up, we didn’t quite get the gamma-short-squeeze-deluxe version that the breathless Odd Lots podcasts warned about. Why did the great oil catastrophe of the Iran War of 2026 ultimately not occur? J.D. Vance has a theory.

Jordan Schneider: Arnab, is he right?

Arnab Datta: It’s important to center a couple of things, and then I’ll kick it over to Rory. First, we did see a big price spike going into the crisis — nearly a 60% increase. At one point, physically delivered barrels in some of the hardest-hit regions hit $130, $140 a barrel. There was physical disruption that drove prices up.

More broadly, a number of interventions came together to keep prices below the truly catastrophic estimates. There was a Strategic Petroleum Reserve action from the administration, and it had an impact. Going into the crisis, many analysts claimed it would have minimal effect and that we probably couldn’t release very much. In fact, release levels hit 1.4 million barrels per day — not nothing, and higher than expectations. Capacity was also rerouted through the East-West Pipeline, which filled some of the supply loss.

Then there’s the biggest story, and the one the vice president is conspicuously not mentioning: the China effect. China played an enormous role in keeping prices under those higher estimates. I’ll kick it over to Rory to get into that.

Rory Johnston: This is where I do my mea culpa as one of those breathless Odd Lots guests who in March was definitely expecting $150 — I think I even said $200 a barrel crude.

Jordan Schneider: That was literally you.

Rory Johnston: It was literally me. Why did I say that? All the things Arnab mentioned — the rerouted pipelines, the SPR releases — were known factors. We knew they were going to happen, and indeed they needed to happen. SPR releases are the textbook reaction; it’s why those reserves were built in the first place. Even with all of that, it wasn’t going to be enough to offset the loss of supply through Hormuz.

The reason other analysts and I expected $200 or more was that you were going to need demand destruction at a scale never seen in history. If prices at $120 or $130 destroy one or two million barrels a day of demand, we needed five, six, seven million barrels a day — a notably higher price threshold to drive that behavior.

Jordan Schneider: To be clear on demand destruction: the first few million barrels are the easiest — you change your carpool, you change your commute. Beyond that, we’re shutting down economic value, not just inconveniencing people. Economies actually become less productive, not just 1–2% on the margins.

Rory Johnston: Exactly right. The scale of demand destruction we would have needed was on par with the depths of COVID — except without the “benefit,” in this deranged use of the word, of a global pandemic and forced lockdowns. We needed to replicate lockdown behavior through price mechanisms alone, which is really hard. I kept asking: what price would you need to get back to the depths of 2020? Something like $100 a gallon at the pump — an astronomically high threshold.

That was the logic. What I missed was China’s tremendous demonstrated swing. We always knew China had massive fuel stockpiles. We didn’t know whether they would use them to help the Trump administration — we can get into why later. Even in my wildest dreams — as I said on my second Odd Lots appearance, the mea culpa round — if I had gone back to March and told myself, “Don’t worry, China’s going to cut crude imports by 45%, five million barrels a day, and nothing will change in the Chinese economy,” I would have said, “Obviously you’re an insane person, because that isn’t possible.”

That’s still almost how I think about it, because what we’ve seen in China is the largest swing of any country in history in its demonstrated capacity and need to consume oil — with virtually no repercussions for anything you use oil for. During COVID, we became familiar with proxies like inner-city transit congestion, flight tracking, and the number of trucks passing on freeways — basic measures of mobility within China. None of them has moved considerably, despite the fact that on paper, on our oil balances, China looks like it’s in the depths of a depression — not even a recession, depressionary levels of consumption.

The only parallel on record for this kind of contraction in Chinese refining activity is the depths of COVID zero in 2022. As I said to Joe on Odd Lots: “You guys were just in China — did you see the entire country locked down?” “Well, no.” “Yeah, that’s what’s weird about this.”

We still don’t know exactly how they pulled this off. The only things we know for certain — everything I say from this point forward is more speculative — are that China reduced its crude imports by five million barrels a day, roughly the scale of a massive collective OPEC cut, and single-handedly solved two-thirds of Asia’s spot-market deficit during this period. The other thing we know: visible crude inventories in the country did not draw down to account for this change in imports. That’s all we know for sure. Everything else is speculation.

Jordan Schneider: So the world is on the brink of crisis, China stops drinking oil from the rest of the world and turns on its own tap, and that held us solid for the few months the world needed until Hormuz —

Rory Johnston: And I should say: China is still importing five million barrels a day less. This has not ended. It’s still ongoing, which is the crazier part. If imports had dropped for a month and bounced back, okay, maybe we could make sense of it. We are now months into this reduction with no official word on what’s happening — and no change in mobility.

China’s Five-Million-Barrel Mystery

Jordan Schneider: Let’s go through some theories. It’s not shocking to me that someone would show Xi a number for how much petroleum to keep under mountains and in salt caves, and he’d come back and say, “No, we need to 20x this” — and that’s what happened over the past ten years. What are folks in this world thinking about how the dynamics actually played out?

Rory Johnston: ChinaTalk listeners will be aware of the idiosyncrasies of Chinese economic data, but some important details. We don’t have official Chinese data on oil or petroleum-product demand, and we don’t have official data on inventories. What we do have is reported refinery activity and trade activity — customs reports on China, which I validate independently through tanker-tracking data.

From that data, we construct what’s called apparent demand: refinery output of, say, gasoline, balanced for net trade — essentially a measure of domestic disappearance of fuel within the Chinese economy. The challenge is that without good inventory data, inventories are a big black box. High apparent demand could be moderated by inventory builds; low demand, like we see right now, could be moderated by big draws.

Here’s what we know for sure: imports fell five million barrels a day, and Chinese refining runs fell about three million barrels a day — again, the steepest contraction since COVID zero in 2022. That leaves a two-million-barrel delta in crude. Where did it go? We don’t see it coming from visible commercial tanks. Because China doesn’t publish official data, we use satellite imagery to track the floating roofs on storage tanks — you can see how high the roof sits relative to the catwalk, which gives a rough measure of how full the tank is. Until very recently, those floating-roof tanks were still fuller than they were on March 1. So that isn’t the solve.

China does have underground SPR storage for crude, which by definition has no floating roof. How much they hold is speculative. The official numbers we know: 131 million barrels of capacity across six underground caverns — likely more. That is suspect number one: it could have been drawn down without our knowing.

Then there’s apparent demand itself — refinery output netted for trade — which is also showing its largest contraction since 2022, with no change in mobility indicators. Because China has no official inventory data and most refined products aren’t stored in floating-roof tanks, there’s a massive black box in refined-product storage, both commercial and strategic.

One more sanity check: for a genuine, massive contraction in Chinese demand, we’d need high prices. Beijing, through its normal fuel-regulatory procedures, limited domestic pump-price inflation through this period. Despite global fuel prices doubling through the crisis, Beijing pump prices rose only 30%. Is a 30% price increase enough to produce COVID-level demand depression? It seems unlikely. What it did do: with crude still expensive and domestic prices weak, refining margins collapsed. So some of the refining reduction is economically explainable. But we still don’t know how mobility was maintained.

The Occam’s razor solve is that Beijing, in addition to crude SPRs, holds ample strategic stocks of refined products — gasoline, diesel, jet fuel. Your refineries shut down a good chunk, your trade corridor is essentially capped because all of Asia is in crisis, and you start releasing those fuel barrels to the market.

We suspected refined-product stockpiles existed before. I wrote a piece in 2023 called “Chinese Oil Demand Doubts,” noting that coming out of COVID zero, China posted the strongest year-on-year demand growth of any country in history. That seemed odd, because 2023’s economy didn’t seem that strong — it still seemed to be struggling out of COVID zero — so all-time-high product demand didn’t make sense. I argued then that China could be building strategic stocks of gasoline, diesel, and jet fuel, which would appear as apparent demand because we can’t see the stocks. At the time, suspect number one was preparation for an incursion across the Taiwan Strait — 2023 was the peak of Beijing’s bellicosity, buzzing Taiwanese airspace and running naval drills off the coast. Now it looks like this may have been preparation for a different catastrophe, like the one we saw now.

There are still open questions. There could have been flex in the petrochemical supply chain — drawing down intermediate petrochemical goods rather than refined products, or switching to gas- and coal-based chemical feedstocks instead of oil-based naphtha. You can tweak a lot around the edges, but it’s hard to make it add up to five million barrels a day. Most likely, they pulled all levers simultaneously, including some injection of strategic stocks we can’t see, because the numbers are just too big.

To put it in perspective: the delta between China’s average import pace from December through February and the cumulative reduction in Chinese crude imports through June was roughly 450 million barrels. That’s more than the entire IEA collective reserve release — if they did in fact release it.

Jordan Schneider: What this reminds me of is 2008. We have a global recession; America pulling some stimulus levers, doing okay. Then the big question: what’s China going to do? This was pre-Xi, but Chinese economic policymakers basically decided, “No, we’re not doing a recession like everyone else. We’re going on the biggest building spree the world has ever seen and keeping the party going.” That made 2009–2013 a whole lot more pleasant for the US and the rest of the world, because of a big injection of fiscal stimulus that wasn’t manifesting anywhere else.

Trump has a lot more to thank Xi for than he may realize. It wasn’t just Bessent and Wright — they didn’t have the levers and reserves to smooth out what would have been a crisis, and the war might have had to end faster, on even worse terms, if not for Chinese economic policymakers once again refusing to face down an economic contraction when they had levers to control it. There are a lot of interesting second-order implications for scholars thinking about future Chinese economic decision-making, which I won’t make you two answer. Let’s park that for now.

Arnab Datta: One quick addition. Rory unpacked what this meant for oil markets, but early in this crisis, the same was true in the LNG market. One instructive stat: by April, China had already resold over a million metric tons of LNG cargoes to Southeast Asian countries. That April figure alone dwarfed the full-year records set in 2023 and 2025. This wasn’t limited to oil — it happened in natural gas too.

Inside the SPR Playbook

Jordan Schneider: Arnab, Rory just described the scale of the bazooka China could point at energy and commodity markets. What was Secretary Wright working with when the Iran War started, and what lessons should the executive and legislative branches learn from how it played out?

Arnab Datta: Despite my reservations about this war in general and the execution of aspects of it, once the administration decided to participate in this global release, the execution was pretty good. A few things to draw out.

First, they were able to release up to 1.4 million barrels per day — higher than many expected, and higher than the peak during the Russia-Ukraine release. That’s the result of a ten-year investment program called Life Extension 2, which upgraded SPR facilities. There are four SPR sites; the two biggest had completed their Life Extension 2 upgrades in recent years. During those upgrades, release capacity was limited because certain sites were simply closed. We were able to release more than expected because we invested in the SPR’s infrastructure.

Second — and even though I’ve used the term “release” this time around — it’s not technically a release. It’s an acquisition. The authority Secretary Wright used is called an exchange: you release product onto the market now, essentially as a loan to intermediaries, who sell it into the market. But those parties commit to return the same barrels plus a premium to the DOE at a later date. Everyone the DOE has released oil to since the war began in March has to return those barrels plus a premium. Particularly at the beginning of the crisis, those premiums were very high, reflecting a highly backwardated, tight market. Companies awarded barrels committed to returning premiums of up to 20–25% to the SPR, starting in November of this year and running through 2028.

Third, the exchange mechanism demonstrated that you can reward contracting parties who commit to giving you more on the other side. They designed the exchange so that the companies awarded barrels now were the ones committing the most premium later. Beyond the 22% minimum premium listed, companies have to offer an additional 2%, 3%, 8% — we don’t know the actual numbers because the contracts aren’t publicly available. To get oil from the SPR right now, you have to commit to a higher premium, which creates a competitive dynamic among the traders and other contracting parties. There’s some worthwhile creativity there.

Finally, and most broadly — coupled with Rory’s analysis of the likely Chinese release — strategic reserves work. There’s a lot of pessimism about strategic reserves, but this episode shows that when you build this infrastructure, along with other swing infrastructure — the East-West Pipeline, swing capacity to produce and transport — it’s worth the investment for exactly this reason. The fact that oil prices never reached the truly harmful level of demand destruction — where economic activity diminishes and manufacturing shuts down — is a success story. But the fact that it happened largely because of China, not the US, is something to think about for future policy.

Jordan Schneider: Rory, your reflections on the utility of having a lot of things in reserve?

Rory Johnston: I completely agree with Arnab. I’ve always been a fan of redundant infrastructure. Over the course of most of my life, we’ve seen an ever-greater pursuit of efficiency at the cost of redundancy and flexibility — and over the past five years at least, we’ve seen unprecedented shocks to every supply chain imaginable.

We were very, very lucky in this crisis. We went in with about the most overbuilt inventory position we’d ever had — a glut for a year prior. Everything was about as good as it could have been. And still, if not for China, we would have been in a deep, deep world of hurt.

White Knight, Backroom Deal, or Dry Run?

Rory Johnston: The other thing worth dwelling on: we’ve talked a lot about the how. The deeply uncomfortable question is the why. On the how, we at least have some data to guide what’s plausible. On the why, it’s completely open-ended. I’ve developed a little catalog of options — some more sanguine, some deeply tinfoil-hatty — and I’d love your thoughts too, Jordan, because while this makes some sense à la 2008, there are weirder elements here than the straight comparison.

One explanation is that this was an altruistic move with an element of self-preservation. East Asia and Europe were going to be the two regions hit hardest by the Hormuz shock; North America, given the shale revolution and the Canadian oil sands, was far more insulated. We know China is in the midst of a long-term breakup with the United States on trade, or at least an attempt to diversify away from the US market. If you’re diversifying away from dependence on the US consumer, your only other real options are East Asia and Europe. Letting those regions collapse into a Hormuz-driven depression doesn’t serve your long-term interests. The trouble with this theory: if Beijing were coming in as the white knight — “Trump has abandoned you, and here I am to save you” — it probably would have used the opportunity to gloat, to broadcast its global stewardship. Possible, but not a complete solve.

The more worrying explanation: the Trump administration made a high-profile state visit to Beijing during the crisis. At the time, it seemed like a disappointment — no big trade announcements, no full-throated endorsement from Xi Jinping to reopen the strait. It looked like a failure. What if it wasn’t? What if there was a backroom deal in Beijing: “Support us on this Hormuz stuff — we know you’ve got gobs of oil at your disposal — and we’ll do you a solid”? What could it be? The most obvious answer always comes back to Taiwan and regional geopolitics. Given the devastation of US military equipment in the Middle East during the first month of the Iran War, the United States pulled a lot of equipment from Asia — drone missile defenses, radar systems, things that would be very handy in an incursion across the Taiwan Strait. This plays into the Donroe Doctrine thesis —

Jordan Schneider: I’m going to stop you here, Rory. The correct explanation is probably the simplest. If someone shows Xi Jinping a chart and says, “If you don’t pull this lever, we’re going to have COVID-lockdown conditions again,” he’s going to say, “Okay, pull the lever.” Keeping the domestic economy from collapse has been the prime directive of all Chinese economic, fiscal, and monetary policymaking since — start the clock at 2005, whatever. All the second-order impacts — do we care whether the war is longer or shorter, what does this mean for American defense posture — are downstream of the central concern you alluded to at the beginning of this podcast: the potentially cataclysmic impact on economic activity of a gamma squeeze in oil and gas. Once you decide that’s not going to happen in China, it has these global consequences of lower prices. It would have been shocking to me to find some justification for incurring that economic pain at that moment for some 4D-chess play of “this will really screw over the US.”

Rory Johnston: May I push back? If that’s the case — we know China has had an appetite for building crude reserves since the ’90s. Why would they have stopped buying? Why would they voluntarily blow the massive energy-security blanket they built, unilaterally, at a moment when everyone acknowledged that the main thing that would push Trump to end the war was the price response from the market? Beijing’s pretty smart. The reason I initially thought Beijing wouldn’t come to the rescue was precisely that they had the domestic reserves to insulate themselves. Instead, they’ve done much more than insulate themselves — they’ve backstopped the entire global oil market.

If you built up this massive energy-security asset, why blow it all at a moment driven entirely by poor US foreign policy? Why not insulate yourself, create the conditions that cause pain for the United States, and totally save your own citizens? They could have protected domestic Chinese consumption entirely without subsidizing Taiwanese, South Korean, and Australian imports and everything else.

And even granting your explanation — why are they still not buying? We’ve seen them buy far more at much higher prices. It doesn’t pass the smell test — unless it was a bazooka to save the Chinese economy that, out of policy inertia, can’t be turned off once the floodgates opened. That may be plausible. But they’re doing three times as much as they’d need to save the Chinese economy alone, in my humble opinion.

Arnab Datta: A question here — I probably lean toward Jordan’s explanation. Maybe they can insulate consumption within their own economy. But isn’t the pain in the Southeast Asian countries also important to China’s economy? China has a weak housing sector right now. If Jordan’s right that the economy is everything, sure, you can protect internal consumption — but if all the Southeast Asian countries go through enormous recessions because they’re not protected, that hits China’s economy too. It’s not altruism; it’s still rooted in self-interest. But it amounts to, “We need to help those countries.”

Jordan Schneider: Do you want to roll the dice on a global recession if you can prevent it?

Rory Johnston: And that is still my base case — the first explanation I mentioned. But it leaves more questions than it answers. One final explanation — not a US conspiracy theory, but less optimistic: China has been building this massive energy-security blanket explicitly for a Malacca Dilemma future, shoring up for a potential conflict with the West. What if this was the world’s largest dry run? Beijing’s trying to figure out whether this policy suite would work. How could you ever simulate the effects of a blockade of Malacca? Suddenly losing Hormuz is a pretty interesting natural experiment: try the system. Does it work? Can we flex petchem? Can we release the SPRs? Can we insulate China from the seaborne market? Very frankly, the answer has proven to be yes.

Whether or not we ever learn the why, the ultimate implication is that China has demonstrated a larger swing capacity in global oil balances than Saudi Arabia, than the United States, than any other country in the world. We’ve always talked about China dominating energy-transition industries — renewables, solar panels, batteries — while North America held the edge on fossil fuels because of the shale revolution. This has proven pretty squarely that the SPR is a discretionary policy tool. Washington does not control the shale patch. Beijing controls whatever is happening right now in China. They have demonstrated much greater discretionary power in the fossil-fuel market than basically the entire collective West. That’s a staggering outcome.

Jordan Schneider: Because they don’t even make this stuff. That’s the weird thing — this isn’t downstream of market forces, of drilling more in West Texas. This is just a decision that was made — decades in the making at this point — to have the tools in the toolkit to bring to bear in a crisis.

Rory Johnston: And it’s working. And we don’t even know what exactly is working, because prices have come back full circle through the crisis and Beijing is still five million barrels a day out of the market. Why?

Jordan Schneider: On the slow-to-turn-off-the-tap question, the Occam’s razor is probably “we just held onto this too long.” This is campaign-style policymaking in China: you always overshoot the target. Once the target’s set, everyone runs toward it, and then they run a little too far. So, potentially less Machiavellian.

The Politics of Stockpiles

Jordan Schneider: One conclusion is that it would be nice for the US to be able to smooth out crises like this — for energy markets, rare earths, what have you. The other conclusion, which I could imagine some Democrats on the verge of controlling the House coming to, is that this enables an executive branch to start wars we don’t approve of — that building a bigger security blanket, so the economic pain of foreign adventures we think are stupid doesn’t hit as fast, will actually lead to more of this happening, not less. What’s your sense? Is this a dangerous insurance policy — selling flood insurance to people in the hurricane path — or prudent national economic management?

Arnab Datta: You’re seeing versions of this in the critical-minerals space too, where the Trump administration has been incredibly active — using the toolkit available to it in very robust, arguably legally creative ways, certainly pushing the bounds of executive authority. There’s a push to say all these things are furthering an administration whose policy is unlawful, corrupt, et cetera. But the push to restrict these tools would be incredibly misguided.

Go back to 2022, when Russia invaded Ukraine. That was entirely out of our control. And one distinction worth noting, going back to our last conversation: the global community begged China, for all intents and purposes, to release from its crude stockpiles at that time. Instead, they kept purchasing — adding demand when the market was incredibly tight. We can’t count on external forces. Every instinct to restrict a tool like the SPR because of its worst-case scenarios also drastically reduces your ability to use it in a genuine crisis not of your making. You can’t design a tool that perfectly matches your preferred policy outcomes while minimizing all bad ones. That’s just not how it works. President Trump was elected democratically; he chose to engage in this war. This is a place where Congress has really let its authorities lapse on war powers.

Restricting a tool that’s incredibly useful for macroeconomic stability would be really misguided at a time when these disruptions are increasing. This is the third disruption in six years. COVID was a downside price shock, but one with enormous repercussions for our domestic shale sector — it created the tightness that made the Russia-Ukraine crisis much worse. Then we had Hormuz. Go further back and you have the Saudi price war. This is part of our daily lives now, and it’s better to be equipped with a tool than not. We can talk about legislating guardrails, but limiting the tools would be a mistake.

Jordan Schneider: Rory, any thoughts or lessons for Western policymakers?

Rory Johnston: I’d dovetail on that — Arnab and I have worked a lot on the SPR file over the past couple of years. SPRs are a bidirectional tool. We always discuss their capacity to offset supply shocks by replacing lost supply. But as we’ve seen, China has been buying oil hand over fist for its SPR stock-building, and that buying manifests as effective demand in moments of especially weak global consumption, like during COVID. That’s a similarly stabilizing force. The reason OPEC has always mattered in global oil markets is that the only thing worse than high oil prices is boom-bust oil prices — it’s the volatility that drives the most deleterious economic consequences. OPEC was handy because it could increase production in high-price moments and cut in weak ones. But OPEC has become, very frankly, increasingly bad at doing that job reliably. China has proven to be a more reliable, technocratic manager of these boom-bust cycles — the way OPEC used to be.

The US SPR can serve a similar role. The pushes to refill the SPR to the brim right now are as misguided as the pushes to draw it down to zero. Empty SPR capacity — the capacity for discretionary buying — is in many ways as valuable as the capacity for discretionary selling.

Arnab Datta: Picking up on that: we had the shale revolution and became the world’s largest oil and gas producer, but we’re still vulnerable to the global market. As Rory said, it’s becoming more volatile and unpredictable, and OPEC can’t be counted on to stabilize it, even in its own interest. The real question is whether we want to build that resilience into our economy, and what’s necessary to do it.

We have to take the reality of the US shale patch seriously. We don’t have the low-cost production Saudi Arabia does — one reason they’ve been the most effective swing producer is that they can ramp capacity up and down quickly at very low cost. Our swingiest swing production is in the shale patch, and new investment takes about six to twelve months to ramp. In a shock like this, that limits you. But we also have near-term swing capacity in drilled-but-uncompleted wells — DUCs — which take about four weeks to ramp up. There used to be a lot of these; they’re now at a historic low in the Permian Basin. If we oriented SPR acquisition or other policy supports toward increasing DUC capacity, we could ramp production faster in this kind of crisis. The policy toolkit will look different — to some extent it means subsidizing companies to mitigate risks. But the broader question is: do we want to be subject to OPEC and China in the future? I don’t think so. Minimizing our exposure to that volatility, or to China’s self-interest, is a worthwhile policy goal.

Rory Johnston: One quick dovetail. North America and the United States are clearly the market-based energy power globally. But that power is entirely driven by market forces, not discretionary government policy. What’s happening in China, at least in my view, is primarily a discretionary policy choice — a choice of government that can be beneficial to the West, or in another world nefarious. Same with OPEC. Casting OPEC as the enemy is about as old as oil policy itself; it’s never been believed that the West’s and OPEC’s interests are aligned — even less so now that the effective second-most-powerful member of OPEC+ is Russia.

So when we talk about Western discretionary energy-policy pull, this is where the SPR discussion comes in. Canada and the United States are absolute energy juggernauts; the West is leading non-OPEC oil-production growth by leaps and bounds, partly because our market systems let the private sector take these risks. The difficulty is that overdependence on market-driven private actors limits our governments’ capacity to effect change in a pinch. The market will always take longer than OPEC or Beijing to turn on a dime in a crisis, as we’ve seen this time.

Share

Everyone Wants a Reserve Now

Jordan Schneider: What are all the other countries going to do? Are they just screwed?

Rory Johnston: One thing we saw through this crisis: a lot of countries didn’t have SPRs. SPRs weren’t cool anymore. Everyone said, “We have the energy transition, we’re going to electrify, we haven’t had an oil crisis in forever,” and got out of the game. Everyone grew to deeply regret that after this crisis — particularly in Asia, where many less wealthy countries lacked the economic and state capacity in the 1970s, when most SPRs were built, and so missed that bandwagon.

Most notable is India. India didn’t have a major SPR going into this crisis and was ground zero for human pain. Not only does India use fuel for transportation like everyone else, but between a fifth and a quarter of total Indian petroleum demand is LPG for cooking fuel. In Maslow’s hierarchy, cooking beats driving. Coming out of this crisis, New Delhi has given strong indications it will build out SPRs, and there’s already confirmation that the Indian government and ADNOC — the Emiratis’ national state oil company — are building a collective SPR in India, currently thought to be 50 million barrels. It serves both parties: it keeps Emirati crude and product on the safe side of Hormuz if this ever happens again, because the Gulf producers themselves also want their oil outside Hormuz. They’ll follow the lead Aramco has taken in stationing stocks all over the world, which helped keep supplies flowing despite the disruption. India is the first major mover, but you’ll see this across much of the rest of the world — countries without SPRs are going to want them after this.

Jordan Schneider: Let’s talk about the Gulf response. Watching the World Cup, I kept seeing Savvy Games Group buying advertising — the PIF buying up every video game company in the world. When are these guys going to stop purchasing Manchester United and start buying drone defenses and building new pipelines and refineries? Has that already happened? This is a huge reckoning for the whole region, right?

Rory Johnston: We’ve already seen some of it. The big flashy investment commitments the Trump administration extracted from the Gulf states have all basically been turfed because there are much more pressing things at play. And they are already building additional infrastructure to increase resilience against future Iranian threats. They’re trying to diversify around the Strait of Hormuz — you can’t entirely diversify around Hormuz, it’s a pretty big thing — but the East-West Pipeline, which proved extremely valuable through this, is getting expanded for sure. The ADCOP pipeline — the Emirati line between the Gulf side and Fujairah, their major blending port on the Gulf of Oman side — was already being twinned to double its capacity. That expansion is about 50% done, and they’ve understandably accelerated it. On top of that, they announced a batch product pipeline that can ship diesel, gasoline, jet fuel, and everything else.

Everyone’s going to do this. The challenge is they’ll also need to invest in a hell of a lot of drone and anti-missile defenses. While the easiest way for Iran to exert control is threatening traffic through Hormuz, Fujairah was attacked and bombed multiple times through this crisis. Iran can clearly hit these other assets, plus the upstream wells that fill the pipelines. Iran will never go away as a threat.

They’ll all try to diversify around Hormuz, but it’s more expensive. Even if you’re paying a massive toll — like $2 million toll for a VLCC carrying two million barrels — no one wants to pay a dollar-a-barrel charge at Hormuz that used to be free, but you’re not going to build a pipeline around Hormuz for less than a dollar per flowing barrel. The natural waterway will always be cheaper. It’s strategic insurance they’re going to pay for, but it will take time, cost a lot, and not be immediately remunerative.

Jordan Schneider: Can we come back to India and the rest of the world building up their own reserves? We have another 2008 echo, where everyone concluded, “We just need lots of dollars so nothing bad ever happens to us.” Is there a world where, in the next crisis, everyone has enough of their own stuff to produce the smoothing that China — because the CCP is super paranoid and was worried about energy security way before it was cool — provided this first time? To what extent does the buffer stop being solely America’s to create, with the whole world getting in the game?

Arnab Datta: One important thing: in oil markets, strategic reserves rest on incredibly developed market infrastructure. This is the most liquid commodity market in the world — maybe the most liquid market outside of dollars. There’s resilience built in because of how developed that market is. Strategic reserves are particularly helpful in tail-risk events, when you have huge losses and want to do a big, chunky release.

But oil storage, especially for refined products, is expensive. Crude is easier to store long-term; refined products tend to evaporate. You can invest in the infrastructure, but it’s costly, and you can’t always predict which market will be disrupted. We’ll certainly see a push to build more resilience, but it’s very difficult for a country to build sufficient capacity to insulate against the whole suite of shocks. Some of this needs to be coordinated, and we should support private market resilience as well — strategic stockpiles matter, but so does swing capacity, like DUCs. Other countries will need to think about that too.

Rory Johnston: This is something you and I have talked about a lot: how to structure the economic framework for how these reserves fit into the system. The legitimate criticism of this kind of thinking is that the private sector already does this — it absorbs swings, holds stockpiles, follows economic incentives, and does all of it much better than governments.

My main criticism of the Biden administration’s SPR release in 2022 was that it wasn’t responsive to market signals. I agreed with the release when it was first announced in April. But by the time it hit full stride late in Q3, the crisis was largely past. At that stage, the SPR release was feeding an oversupplied market — procyclically adding to volatility rather than serving its purpose as supplier of last resort.

If SPR policy is going to be effective within a broader, much larger, much more dynamic private system of inventory and merchant storage, we need a rules-based system that’s far more responsive to the market as industry understands it — not just a bazooka. To your earlier point about policy inertia in China, Jordan: policy inertia is a problem here too if we’re trying to fine-tune volatility in fast-moving markets. We can’t have a situation of “we’ve clicked the button and now, whatever happens, we can’t back out because it was so hard to get going in the first place” — which was my understanding of what happened in the Biden administration: “We worked so hard to do this big release; we’re not backing out now.” Understandable on a human level, but backing out is what was needed. We didn’t need to draw the SPR down as far as we did in 2022, because the bulk of the drawdown happened when prices were already in freefall — it was too much of a good thing at that stage.

Jordan Schneider: One final question for you, Rory. Arnab’s energy is median-DC-policy-person energy. Where do you put yourself among the other commodity analysts? Are you an outlier, or is everyone wired like you?

Rory Johnston: In terms of personality, I hope not — for their sake. But in terms of my view on this —

Jordan Schneider: No, no — we’re not talking about the calls. I mean your way of being: level of enthusiasm, level of excitement.

Rory Johnston: Is my way of being normal for oil analysts? I don’t think so. The best oil analysts are typically very level-headed. The irony of this crisis is that everyone got to know me as this alarmist, ranting-and-raving permabull — “Oh my God, he went on Odd Lots and called for $200 crude. He must be insane. He must have always been bullish.”

Jordan Schneider: I think you called for $600 crude on this podcast.

Rory Johnston: I was so bearish going into this year. I expected a massive glut. I did not expect the Iran War. I did not expect the Strait of Hormuz to be closed. This was the biggest shock anyone could imagine in the oil market, and the fact that we avoided the worst consequences is actually very heartening for the global system. It’s just deeply uncomfortable as an analyst that we still don’t fully understand how we managed to pull it off — or why Beijing helped as much as it did.

Jordan Schneider: Amazing. And one final question: what’s the modal music genre of choice for the oil and gas industry?

Rory Johnston: Whew, I don’t know. I’m a bad millennial lover of what they call stomp-clap music in the most pejorative sense —

Arnab Datta: We should do Arcade Fire or something like that — the Canadian roots. Maybe “Rebellion,” or “Wake Up.” “Wake Up” is a good one.

Jordan Schneider: Well, I can’t use the real songs anymore, so I have to make my own. I’ll send you some of the Iran War ones. We’re on Spotify now: ChinaTalk Records. Check us out.

This was really fun — a nice break in routine. Talk about too much AI and you forget there are actually things that are brown and sludgy that move around the world and matter.

Arnab Datta: That the world still runs on.

For all its technological feats, China’s push for high-quality innovation has not made the economy that much more productive. This lends greater urgency to other productivity drivers such as firm efficiency — an exigency that Beijing is hardly unaware of. In recent months, the central government has outlined a series of regulatory measures to discourage local officials from propping up inefficient firms. But nothing stops local leaders from circumventing those regulations. The Chinese bureaucracy continues to reward officials who hit GDP growth targets, minimize unemployment and social unrest, and defer to the dynastic mandates of President Xi Jinping. Until the underlying incentive to keep inefficient firms afloat is recast, China’s tireless pursuit of productivity-led growth will continue to bear little fruit.

The Elusive Quest for Productivity 强弩之末

(Bonus points for guessing which book this header is inspired by!)

As an economy matures — like China over the last decade, growth increasingly moves away from capital accumulation. Because there is only so much equipment that one can “fit” a worker with, additional capital will eventually begin to yield smaller increases in GDP. More formally known as “diminishing returns to capital”, this phenomenon in part explains China’s decelerating growth in recent decades. Adding fuel to the fire is the country’s protracted fertility decline, which reforms to the one-child policy in 2015 and 2021 have failed to reverse. Economists predict that shrinkages in China’s working-age population will reduce per-capita GDP growth by 10 percent over the coming decade.

These unenviable economic circumstances necessitate a shift toward productivity-led growth. And who would be more aware of such a need than the Chinese Communist Party (CCP)? The CCP’s latest Five-Year Plan reemphasizes what have long been cornerstones of China’s broader economic agenda: innovation and technology upgrading. From 2015 to 2023, government spending on R&D in strategic sectors increased nearly fourfold under industrial policy guidance like Made in China 2025. While such lavish spending has given rise to the likes of BYD’s blade batteries and Huawei’s commercial 5G leadership, there have not been improvements in year-on-year productivity growth. The reasons for this discrepancy are manifold — R&D outcomes are inherently uncertain; making new discoveries becomes more difficult as technical fields mature; novel-product innovations do not necessarily make industry processes more efficient; and subsidies may have adversely encouraged low-quality patents over productive innovation. Or it could also be that innovation did bear sizable productivity gains, but they were ultimately offset by the expansion of inefficient firms. Either way, China’s push for high-quality innovation has not made the economy substantially more productive.

Granted, China does not rely solely on domestic innovation for technological progress — it also adopts frontier technologies from advanced economies. But rival powers have imposed lasting technology blockades that stifle this approach. Washington continues to restrict Tokyo and the Hague from exporting semiconductor manufacturing equipment and computer chips made using US technology. Restrictions on US investment in Chinese or Chinese-owned enterprises working on sensitive technologies — such as quantum information systems and AI — have similarly stiffened over the years. And even if there were no trade restrictions, it is unlikely that importing technologies would translate into significant productivity gains, given that China has already caught up to the global technology frontier in many industries.

Overcoming this productivity lassitude requires going beyond technological progress. China needs to become more efficient by dismantling its sprawling constellation of inefficient firms, the worst-performing of which are known as “zombie firms”. What defines a zombie firm is not exactly clear, however. The State Council considers a zombie firm to be a company that has been consistently unprofitable, belongs to an industry with overcapacity, and violated national environmental and quality standards. And so are firms that made three consecutive years of losses and failed to adhere to guidelines on structural reform.1 The International Monetary Fund (IMF), by comparison, defines zombie firms to be loss-making companies that enjoy below-market interest rates, often given by state-owned banks.2

Since zombie firms — and inefficient firms in general — enjoy subsidized credit and tax breaks from local governments, they are shielded from market forces that would otherwise compel them to become more efficient or shut down. Scaling back these financial benefits would thus pressure inefficient firms to restructure themselves or exit altogether. This reduces “wastage” of scarce credit. Instead, credit would be redirected to more efficient firms, allowing them to make greater investments. Eventually, as these firms make up a larger share of their respective sectors, the economy as a whole becomes more productive. In 2017, the IMF estimated that such productivity gains would raise China’s growth rate by 0.7 to 1.2 percentage points, nearly a fifth of the current GDP growth target of 4.5 to 5 percent.3

Feeding the Walking Dead 苟延残喘

But the widening footprint of inefficient firms also reflects the CCP’s difficulty in tackling the problem. Local officials’ promotions and financial rewards are linked to performance indicators established by senior officials. These indicators have undergone significant changes. In the first two decades of post-Mao China, local leaders competed against one another on hard growth targets, such as the annual gross value of industrial output in their respective jurisdictions. Such tournament-style competition backstopped astronomical investments in physical capital, resulting in miraculous economic growth. However, as income inequality and unemployment fuelled popular discontent in the 1990s, the CCP elite began expanding cadre evaluation metrics to include social development indicators like urban employment. Hitting GDP targets alone is no longer sufficient to protect one’s political career; one must also stave off the layoffs that follow firm closures. This provided another reason for local officials to prop up underperforming firms, at great cost to the economy writ large.

Take photovoltaic company LDK Solar (赛维LDK太阳能), for example. In addition to being the biggest taxpayer in Xinyu (新余市), a prefecture in Jiangxi, LDK Solar was a crucial employer in China’s politically favored solar industry. As of 2013, the firm had an estimated total of 10,000 workers despite having accrued $2.8 billion in debt. Unsurprisingly, the company had enjoyed generous state support even after making losses for nine consecutive quarters, including assistance from the local government to obtain a $327 million loan from national banks. According to one insider of the Jiangxi banking industry, the negotiation was “a coordination meeting on paper, but coercive in reality”, with threats from the local government to “terminate all relevant business and sever all relations with any bank that refuses”. The gravity of an LDK Solar collapse — to citizens and thus officials alike — was evident. A year earlier, Zhao Hongming (赵鸿鸣), then spokesperson for the Standing Committee of the Xinyu Municipal People’s Congress, said that the firm was an “industry vanguard” that “employs more than 20,000 workers”, and that “support [for a troubled LDK Solar] from all parties would bring about great benefit to the thousands of employees and local residents of Xinyu”.

To make matters worse, there has been mounting pressure on local officials to align themselves with central priorities in recent years. Xi — in line with his predilection for “top-level design” (顶层设计) — has significantly narrowed the latitude in which local governments decide how to meet central policy objectives. Fearing demotion or punishment, many bureaucrats have erred on the side of excess, plowing even more subsidies into priority sectors. Aside from serving as lifebuoys for inefficient firms like LDK Solar, these subsidies have resulted in underused facilities and infrastructure. Think the ghost-like fabs of the now-defunct Dehuai Semiconductor (德淮半导体), the idle furnaces of Shandong’s steel mills, the graveyards of EVs and shared bicycles in Hangzhou. Xi, too, recognizes the perils of the subsidy glut. In his February address on cultivating sound conceptions of bureaucratic performance, he urged local officials to abandon “deviant or misplaced views” (偏差错位的政绩观), which have led to wasteful expenditure, formalism, and performative projects.

That Xi’s imperial decrees have also contributed to bureaucratic inertia — a well-known phenomenon among contemporary observers — is not actually contradictory. Officials often describe Xi’s directives as edicts undergirded by the wishful logic of “not only, but also, and also.” For example, local governments must not only promote innovation in priority industries, but also minimize financial risks inherent in funding experimentation, and also strictly conform to centrally established standards for achieving both goals, however vague they may seem. Faced with conflicting priorities, local leaders avoid risky or innovative policies, which are needed for furthering domestic R&D through greater private capital involvement. In an almost compensatory fashion, they default to conventional approaches that also suggest alignment with Xi’s mandates — such as pouring subsidies into semiconductor companies — with excessive zeal. This explains why it is often difficult for central directives to inspire downstream policy change. In 2018, for example, the Ministry of Finance called on local governments to submit a list of zombie firms in three months’ time; complete all restructuring, deregistration, and liquidation works by 2020; and impose restrictions on loan extensions and corporate subsidy payouts. One need not look far to know that the above remains a huge work in progress.

The Arrow Misses the Head 扬汤止沸

The productivity costs posed by inefficient firms have not been lost on the CCP elite. As early as three decades ago, former premier Zhu Rongji launched a landmark campaign to reform inefficient state-owned enterprises (SOEs) by transferring their responsibilities in home ownership and public insurance to the civil service. Yet, while efforts to combat firm inefficiency have only intensified in the years since, they often run into an incentive constraint.

To see why, it helps to begin with the basic understanding that local officials have functionally inexhaustible options for supporting inefficient firms. As long as the underlying incentive remains in place, local officials would sidestep regulatory measures by providing new kinds of financial lifelines. Earlier in March, the State Council announced its plans to restrict local governments’ discretion in giving out fiscal subsidies. While the specific prohibitions are not yet known, there remains little stopping officials from taking actions like pressuring local banks to initiate loan rollovers or creating purchase arrangements with loss-making firms.

Consider the vast array of financial benefits that Dayun Automobile Co. (大运汽车) enjoyed. The brainchild of who was once an enterprising 18-year-old grain trader, Dayun Auto has long been the largest heavy-duty truck manufacturer in Shanxi. Its dominance, however, belies a checkered history. Following peak sales in 2017, the automobile behemoth made a misguided hop onto the EV bandwagon. Unremitting competition in the EV industry eventually left an indelible dent in its cash flows, causing payment failures that pushed its suppliers to file a restructuring application against it in November 2024. The firm slashed its headcount by half as a result, and it focused on cranking out heavy-load trucks again. What is most concerning, however, is not its downfall; if anything, it did recover a little since the court-ordered restructuring began. Rather, it is the fact that it still buckled despite the many pillars of state support underneath. The EV subsidies Dayun Auto received in 2017 and 2018, which alone equaled 20 percent of its earnings then, were just the tip of the iceberg. It also enjoyed tax benefits that made up more than a tenth of its profits, purchase arrangements with a Yuncheng (运城市) state-owned firm that guaranteed the sale of one-third of its trucks, and even a highway toll exemption for its trucks to boost sales. In part, the wide variety of preferential deals reflects the limitless ingenuity of local officials, one that specific regulations alone would be unable to constrain. Indeed, as a former Dayun salesman by the name of Zhao Xinzheng remarked to Bloomberg, “It’s impossible for Dayun to go bankrupt. The government would never allow it.”

In addition to regulatory measures, the CCP has set up specialized bankruptcy courts. Since 2007, the number of specialized bankruptcy tribunals and courts has steadily grown to more than 100 across different prefectures. As judges in specialized courts are often better-educated and newly appointed, they are less susceptible to political interference than their counterparts in traditional civil courts. Expectedly, these specialized courts have expedited bankruptcy proceedings and thus the restructuring or liquidation of inefficient firms, raising productivity in their respective local economies.

But like regulatory measures, such judicial specialization still skirts the root incentive for local governments to keep inefficient firms alive. Political interference can also occur before adjudication, where officials intervene through financial lifelines or direct court-government negotiations to prevent — or at least, slow — enterprises from entering a bankruptcy proceeding in the first place. As in the case of Dayun Auto, it was not until several years into its financial decline that non-state entities filed for compulsory reorganization against the Shanxi giant. Only then was it legally required to restructure itself (though restructuring had already been underway in Yuanhang Auto, the EV group of Dayun Auto).

In that case, why not involve the central government in initiating bankruptcy proceedings? Most recently, Beijing announced a pilot amendment to China’s Company Law in seven economic hubs. The amendment would empower central authorities to petition for the court-mandated liquidation of zombie firms that have not voluntarily dissolved, bypassing local-level resistance. While this legislative amendment is a step in the right direction, its effectiveness is ultimately contingent on the identification and processing capacity of central authorities. Politically favored zombie firms may remain under the radar as they continue to operate with generous financing from local governments, making it difficult for central authorities to identify entities for judicial liquidation.

Once Bitten, Twice Shy 入土为安

Of the deluge of central interventions, fiscal auditing most directly addresses the incentive constraint by penalizing local officials for racking up debt, which sustaining inefficient firms invariably leads to. Over the last decade, the fiscal adequacy of local governments has come under heightened scrutiny. In the 2017 National Financial Work Conference, Xi emphasized the need for local governments to “adopt just views on bureaucratic performance, strictly control the rise in local-government debt, and ensure a culture of lifetime accountability and retrospective responsibility”. Also underscoring the urgency of debt control is Premier Li Qiang, who warned local officials to limit debt risks and husband state coffers in a State Council video conference two years ago.

Would fiscal audits alone be sufficient, though? It is unlikely. While Xi’s debt-slashing mandates would most certainly curb reckless subsidy disbursements, they could also discourage local officials from accepting the short-term financial consequences of firm closures, such as the loss of corporate tax revenues and lower land prices. In fact, value-added tax and corporate income tax are the two biggest sources of revenue for local governments, accounting for a total of 56 percent in 2021. Declining land revenues since then have only magnified the importance of enterprise survival to state coffers. Deepening this fiscal predicament is also the fact that most local governments already face fiscal deficits. Aggregate local deficits, as a percentage of GDP, has risen steadily from less than one percent before 2015 to roughly five percent today. When it comes to debt, the situation is far worse: Only one provincial-level region, Tibet, has an adjusted debt burden below 100 percent,4 inclusive of on-lending debt. Saddled with recurring deficits and climbing debt, local officials are likely even more averse to drastic subsidy reductions that would mar balance sheets in the short run.

To actually embolden local officials to bury China’s corporate zombies, locally established debt reduction targets could be incorporated into the fiscal auditing approach. While the selection of specific indicators — such as adjusted debt burden and the debt-to-GDP ratio — would remain the preserve of the central government, local governments would have the autonomy to set appropriate quantitative targets for each indicator and the timelines for meeting them. Officials would be assessed on not just the extent to which the targets have been met, but also how ambitious the targets are; the latter is necessary to encourage bolder strategies to topple inefficient firms and slash debt. Moreover, recent developments in central policy favor such an initiative. Unveiled in November 2024, China’s 10 trillion yuan ($1.5 trillion) debt refinancing plan has allowed local governments to borrow an additional 6 trillion yuan through special-purpose bonds until 2027, and a further 800 billion per year over the following five years. In effect, local governments would be replacing their “hidden debt” — accrued from decades of borrowing through local government financing vehicles — with more transparent forms of debt, assisting central authorities in their assessment of local performance in debt reduction. And as officials in high-debt provinces are more likely to fear the loss of tax and future land revenues, the weightage of debt reduction performance in bureaucratic evaluation could be made higher in debt-laden localities.

Ideally, Beijing should also introduce bureaucratic incentives that encourage market-led allocation of financial capital, which would reduce inefficient state intervention in the survival of firms. Researcher Shengyu Wang remarked to me that many local governments have recently begun to report on private capital mobilization, measured as the amount of private investment attracted for each public dollar of venture capital. He believes this to be indicative of high-level commitment to liberalizing capital allocation for greater efficiency. How much further such commitment translates into incentive reforms will ultimately depend on political will.

China’s bureaucratic incentives are currently stuck in limbo. Once an investment dynamo, they helped transform a backwater into a global economic heavyweight. But as the Chinese economy matures and transitions from investment-driven to productivity-led growth, the cost of sustaining inefficient firms — the woeful byproducts of China’s bureaucratic incentives — would only continue to rise. Whether local officials stop doing so depends on more than just what they cannot do. It is also about what they stand to gain from consigning the corporate undead to the grave.

The lifecycle of AI, from training to output, has tangled itself around the lives of Chinese illustrators in ways that run in opposite directions.1 The machine was fed their work, learned their expressions, and is now displacing them. And in the process, it has forced them to answer a question that no one expected to become that urgent a few years ago: how do you prove you are human?

The answer, so far, involves courtrooms, livestreams, and job markets. It involves regulatory regimes that strictly govern AI outputs but stay lenient on the inputs, a community that hunts for signs of AI in each other’s work and demands live drawing sessions as proof of humanity, and an education system that trained illustrators to produce exactly the kind of consistent, replicable output that AI does better and cheaper. Ultimately, the question may be less about how to prove you are human than about why that proof has become so hard to give. The answer, I will argue, lies less in what AI learned from humans than in what humans built long before it arrived: an exam system that trains young artists through standardized metrics and repetitive practice, which looks unsettlingly like an AI training pipeline. In the end, behind Chinese illustrators’ attempt to combat AI lies the broader educational infrastructure that makes humans indistinguishable from machines. So what type of expression emerging from that system, if any, would be untokenizable?

This is a story about Chinese illustrators. It is also a story about what we built before AI arrived, and what we are finding out about ourselves now that it has.

Feeding

In December 2023, the first Chinese case addressing whether AI training constitutes copyright infringement was filed by four illustrators against Xiaohongshu to the Beijing Internet Court. Xiaohongshu had been developing its own AI drawing service, Trik AI, which was launched in April 2023 to allow users to “create artworks” by uploading a photo. Four creators discovered that some artwork created by Trik AI looked extremely similar to what they had produced in the past. The illustrators claimed that unauthorized use of their work to train Trik’s model violated their reproduction rights. “All the brushes I used in my art, the way that I portray flowers, they are all from me and with my individual characteristics. But you can see from its art the traits of my art,” said one artist.

Xiaohongshu, on the other hand, argued this was “fair use” under Chinese Copyright Law. As of the available evidence, no judgment has been published. Xiaohongshu removed Trik AI shortly after the case went big, but it has not stopped more and more AI artwork-generating services and platforms from emerging.

Chinese legal scholars argue that China is taking a regulatory approach of “moderate leniency,” emphasizing control over downstream AI-generated content (AIGC) while adopting a more permissive stance toward upstream training. The structure has double meaning for illustrators: not only do they have little legal recourse against their work being fed into the machine, but, as the next story shows, the downstream controls designed to catch AI can end up catching them instead.

Witchhunting

On October 29, 2025, an unusual drawing livestream happened. Rather than showing fans how a work is produced and chatting with them, as normal drawing livestreams in China would, this livestream is a bet-on agreement — betting whether the illustrator who initiated the livestream can replicate their work in real time without using AI.

The illustrator, who goes by “Angel from Mercury” 来自水星的天使 (abbreviated to 水星 Shuixing), had published a commissioned work on Xiaohongshu, which the platform flagged as “may contain AI.” Under AI labelling provisions that came into force in September 2025, platforms are required to auto-detect and flag suspected AIGC (More details on China’s AI labelling system can be found in the explainer piece: Labelling AI-Generated Content in China), but both machine and human reviewers struggled to land on a verdict.2

To dissuade doubts, Shuixing signed a bet with another illustrator, and the two livestreamed themselves drawing to prove their humanity. On the stream, Shuixing did complete a full illustration, but the accuser remained unconvinced without giving further evidence. The whole bet died down unresolved. But this is not a unique case.

As early as 2022, differentiating AIGC vs. human drawing has become a topic in the illustrator community. Because generative AI relies entirely on the non-consensual harvesting of human art, using the technology carries a heavy social stigma. Under this backdrop, the community started what some describe as a “witch hunt,” which resulted in many big and small bet-on agreements and consequent livestreaming drawing sessions to prove authenticity. “Proving you are a human” has become a norm among illustrators, and showing the process of human hands drawing a picture line by line seems to be the last piece of evidence that humans can use to prove themselves as human.

Displacing

If AI-made art is indistinguishable from human work, whether something is AI or not becomes just one side of the coin. The witch hunt’s uncomfortable flip side is: who needs human labor if AI can do the same?

In May 2023, China Economic Weekly 中国经济周刊 reported on “the first wave of people who got laid off because of AI,” featuring illustrators as an occupation. The article kept a cautiously pessimistic tone. It features several illustrators whose careers were negatively affected by AIGC. At that time, Stable Diffusion and Midjourney had become popular in China, good enough to carry out specific tasks like image generation but not capable of executing other design work end-to-end. Most people were speculating that the status quo capabilities of AI would gradually eat into the workflows of illustrators, both those who were already in the field and students in art academies. “I guess now teachers from academies of arts can not figure out whether they should still select students based on how well they sketch,” said one interviewee. This has increasingly been reflected in changing major options at universities. Chinese universities have been axing majors rapidly — 12,000 were eliminated from 2021 to 2025, with the majority being humanities and arts. Among the 16 majors cancelled by the Communication University of China 中国传媒大学 (CUC) in 2025 are photography, visual communication design, and illustration, the last of which was long considered one of CUC’s flagship programs.

And this is a displacement encouraged by the state. Since 2021, the state has been aggressively pushing “AIGC creation” — including launching various campaigns, competitions, and conferences to encourage people to create films, videos, and pictures using AI — positioning AI as a new creative frontier rather than a threat. Even before calling for diffusing AI in every aspect of every industry in 2025, many businesses have embraced AIGC, which results in job loss for mid-level and junior illustrators. Southern Daily 南方日报 reported that by 2023, many art design positions at game companies had started disappearing. Illustrators used to be very popular among second-tier game companies, but increasingly faced massive layoffs. For years, tech firms have been training specialized AI for game design. And that was 2023: an AI-generated Will Smith could barely finish his spaghetti without eating the fork and his finger.

Some are more optimistic about the future and argue that AI is not an opponent but a tool — humans won’t be displaced by AI, but by other humans who can use AI. In fact, with the state encouraging AIGC, a community of “AIGC artists” has emerged. Some have no arts background and see AI as a democratizing tool that empowers them to create art even without going through the dreadful training needed to graduate from art school. Some other creators who have accumulated years of experience have also switched to AI as a novel method to explore, with some prominent examples like the AI short film by celebrated auteur Jia Zhangke (贾樟柯) and AI-generated “Multimodal Art” exhibition by Jia Hongjun (贾鸿君), a veteran ink painter and professor at the Central Academy of Fine Arts 中央美术学院.

The “Proofs” of Humans

Whether resisting or embracing, feeding or being fed on, all creators are searching for the same thing: a proof of humanity — evidence that being human still means something, and protection against being rendered indistinguishable from AI. Much of what illustrators in China experienced is specific to their context: particular legal regimes, community ideologies, and art education systems. But the struggle of seeking that proof is shared. Are these three common responses — regulation, stigmatization, embracing — as used by the Chinese illustrators in the stories above, enough to “AI-proof” humans?

Regulations

Some argue — and indeed demand — that regulation become the proof protecting humans from AI. Policy is unavoidably helpful, but it is not the proof.

In an ideal world, policy would have acted early enough so that the vast amount of data on the internet — much of which, like illustrators’ artwork and style, should be copyrighted — would not end up in AI training data. But most places, whether China, the US, or the EU, acted too late, or even still have yet to act. Moreover, while copyright seems like the natural candidate as it exists to protect human originality, it in fact protects specific expressions from being copied, rather than the art style itself, which more often is a way of seeing, a visual language, a distinctive aesthetic. And even that narrower protection has limits, as the early US district court rulings have mostly favored AI developers when training data was legally acquired.

In a similarly utopian world, policy could help us differentiate AI output from human output, so that human-made art would not be wrongly flagged by platforms as AI and people would not accuse one another. In reality, however, governing AI outputs is not easy even when regulations are in place. China has built one of the world’s most elaborate AIGC labeling regimes — and as this piece and the other explainer show, the result is a system that suppresses stylistically-consistent human work alongside AI spam, turns proving your humanity into a public performance, and generates evasion markets that are illegal but structurally inevitable. So long as the regulator cannot find a technical solution that is 100% perfect at discerning human vs. AI output — which is increasingly impossible for digital art and writing, where humans may increasingly internalize the style of AI output and vice versa, to say nothing of human-AI co-creation — the public will be stuck forever questioning whether something is AI or not.

Stigmatization

If policy cannot protect us from being fed into AI or distinguish us from it, should we rely on community efforts? Much like the Chinese illustrator community, many other creativity-adjacent spaces, such as Substack writers, have been intensely debating the “AI-like” output. One of the mainstream views on writing, very similar to how Chinese illustrators witch-hunt their peers, is that writing that is perceived to be AI-generated needs to be stigmatized. To quote Substacker-extraordinaire Sam Kriss, “I can tell. I can always tell. You think I won’t notice, but I will. There’s no hiding from me. If you let AI do your writing I will find out, and I will kill you.” The logic behind underappreciating AI outputs sounds valid, especially for human creators. Part of it is that AI output is “nonsense,” as it is viewed as a lack of genuine ideas and thinking process; part of it is that AI itself is bad for creators because it is fed by their works without consent. Regardless of the reason for the stigmatization of AI, the methods are imperfect and dangerous. “I can tell” relies on a very unique human instinct that most reliably lies in native speakers who have interacted with AI for a while — which, although it seems large for certain communities, remains a very small fraction of the world that uses English to write. More ironically, still, is that to avoid being judged as AI, humans start to change their naturally human creative process — avoiding em dashes in writing for now, and maybe changing drawing and filming techniques in the future — which ironically invalidates what they hold as humans having genuinely and inherently different judgment from AI. The witch hunt was supposed to defend a boundary of human vs. AI, but ends up forcing humans to write less like themselves so they won’t be mistaken for the machines they are trying to distinguish themselves from. [Note from editor Lily: you will have to claw the em dash from my cold, dead hands —and if ChinaTalk’s use of the em dash bothers you, you can take it up with Truman Capote.]

Embracing

At the opposite of the regulation and stigmatization that try to prevent AI from coming to humans stands the attitude to embrace the technology. When AI becomes so capable after being trained on human work and resembles human capabilities, using AI, and upgrading yourself, appear to be the very few ways to not be displaced at an individual level. By embracing AI in content creation, Chinese illustrators and other artists can create new genres of art and make themselves more competitive in the market.

This statement of ‘embrace AI’ is vague enough to sound valid. Embrace what, exactly? China is embracing AI by drastically reforming its university programs: cutting 12,000 and adding 10,200 new ones. It seems to equip itself for an AI future by replacing “useless” majors like arts and humanities with AI and robotics, but it is an open question of how AI as a major can itself survive AI progress. Today’s lesson on scaffolding may soon be outdated tomorrow, while suddenly the “outdated” aesthetic taste becomes more important when the world is flooded with “vibe-code-y” websites with beige over-polished user interfaces. It is hard to go all-in on something whose shape no one can yet see: where AI capability will go, what skill it will develop next, or which profession it will kill, and we individuals are just left exploring in the dark and fearing the unknown.

The deeper problem with embracing is the same one that undermines stigmatization: both end up being processes of humans frantically reorganizing themselves around AI, changing behavior in response to a target that keeps moving, and panicking without purpose.

The real problem, in both cases, is what the philosopher Günther Anders called “Promethean shame”: the feeling of inferiority humans develop before their own flawless, replicable machines, and the urge to remake themselves in the machines’ image. Whether we resist AI or embrace it, the underlying process is the same: an internalization of the machine as the reference point. Even those who try to “preserve human distinctiveness” by discriminating against what they perceive as “AI creation” have already let AI define them, as they are no longer creating as themselves but the opposite of a machine.

That “shame,” driving extreme hate or extreme frenzy, makes us purposelessly predict and prepare to change as AI changes. When AI starts to use en-dashes instead of em-dashes, or when the next fashionable job is no longer “harness engineer,” we change accordingly, without asking why — and we end up controlled by AI. Not in the sense that AI outsmarts humans and rules humanity, nor that a US AI or a China AI ends up ruling the world, but that the unpredictability of AI progress (what it will be able to do, how good it will be, when its capabilities will diffuse) and the anxiety it produces control what we learn and how we act. We lose our agency to our own human fear before we lose it to any superhuman AI.

The Mechanization of Humans

But why can we so easily lose our agency to fear — fear of being utilized, resembled, and displaced? Are we on an irreversible trajectory to be consumed by, merged with, and gradually displaced by the technology we built? The reason for the failure of these three methods has less to do with their rigor and more with the false binary of “human vs. machine” upon which they were built. They do not work precisely because the binary is dissolving, and had been dissolving even before AI. Before AI speaks, writes, and draws like humans, humans have already trained themselves to speak, write, and draw like machines, which again can be exemplified by the Chinese illustrators’ journey.

In China, learning drawing is seldom a rosy story of child prodigies embracing free exploration and flying with their creativity. Instead, the pipeline of arts graduates is not that different from the AI pipeline. Like elite STEM students who may be preparing for the International Maths Olympiad (IMO) and gaokao at the same time, most art students walk two tracks: gaokao and yikao (艺考, short for 艺术类专业招生统一考试, the joint exam for artistic majors enrollment) combined determine whether they can be admitted into good academies of arts, with a lower gaokao performance offset by a stronger yikao score.

Just like gaokao, the yikao leaves little room for creativity and ties everything up by metrics. For drawing, the exam is composed of three parts (sketching, charcoal drawing, and color), each guided by a very specific prompt like “draw an old man in a jacket sitting still” and graded against very specific metrics. Many students rely on tutoring organizations with guidebooks and intensive crash courses (美术艺考集训班), some of which have students drawing 14 hours a day for months alongside their gaokao study. Most of that training is repetitive work to refine skills rather than free-flow creation, which is exactly what AI is good at displacing. Very similar to how other students prepare for the gaokao, yikao students wake up at 6:20 and go to the studio, take 10-minute breaks each for breakfast, lunch, and dinner, and spend the rest of their time switching between the three subjects tested in the exams until finishing the last class at 22:30. Some continue to draw in dorms before they sleep.

But not everyone walking this pipeline wants to be an artist. Whether a stereotype or reality, a fair number of students choose the arts simply because they see no future in doing gaokao. As a teacher of sculpture at China Academy of Art (CAA) in Hangzhou, one of the two top academies of art in China, observed: “Lots of people are actually gifted in art; their talents have just caused them to drift away from the measurement standards in the academic system.”

Moreover, this is a system that builds upon a very specific version of meritocracy, where efforts are graded under extremely standardizable metrics, which in turn cultivates extremely automatable skills. The grading involves four rounds, with the first pass through tens of thousands of papers being graded through zouma guanhua “走马观花” — glancing something very quickly without enough care — as teachers sort drawings into qualified and unqualified based on immediate visual impact: whether the light-dark contrast reads from a distance, or whether the composition snaps into legibility. After the first pass, four rounds of iterative re-ranking follow, where papers are sorted into score bands and then repeatedly reviewed, with those that seem too high getting moved down one band, and those that seem too low getting moved up one. The process continues until every paper has settled into a score that a majority of examiners agree on.

When the system needs to select from the masses and present itself as fair, metrics become a benchmark, people become “skills,” and the purpose of life becomes utility to society. Students are trained on past exams to predict the next one; exam scores become the benchmarks; a good university degree is the reward of doing things correctly, and ultimately, one becomes useful because they possess certain skills. Everything sounds scarily similar to how you train an AI, even the grading of the yikao artworks: a fast first pass filters for general legibility is not unlike how pre-training data is culled for quality at scale; the rounds of iterative human re-ranking converging on consensus scores look similar to multi-pass annotation with inter-rater calibration; and the highest scores go to yikao students who have most precisely internalised what the system rewards, which is the same for AI models.

Back in 1967, American historian Lewis Mumford argued that the first machines were made not of metal but of humans, with armies, labor gangs, and bureaucracies that organized people into standardized, interchangeable parts. Calling it “megamachine”, he pointed out that the structure existed as early as the pyramid age, long before the machine as a concept concretized. The art exam system in China is one megamachine: students are trained to operate “as a single mechanical unit of specialized, subdivided, interlocking parts” — a skilled sketcher or a practiced illustrator, who operates rather than creates. As the famous Mao-era propaganda phrase went — “be a screw” (做一颗螺丝钉). And a human screw can be displaced by a real screw.

Thus, what enables AI to displace humans is not AI itself — it is a process of mechanization that long preceded it. We built governance regimes that protect output over process. We judge content based on tractable traits rather than originality. We designed education systems that train replicable skills over genuine creative vision. Long before AI arrived, we were already valuing output over process, evaluating humans against standardized benchmarks, and rewarding the kind of consistency that machines would one day do better and cheaper. We trained humans like we train AI. It should not be surprising, then, that telling them apart is becoming so hard — or that the machines, having learned from us, are now teaching us what we valued all along.

Subscribe to Zilan’s Substack!!

Acknowledgements:

Zilan is grateful to Conor McGlynn, Afra Wang, Amber Hu, and Kayla Blomquist for their feedback on earlier drafts.

The author acknowledges the use of LLMs for brainstorming and copy-editing, and hopes the readers don’t mind em dashes that are genuinely typed by humans.

Our guest is Roel Konijnendijk, ancient history lecturer at the University of Oxford and author of Classical Greek Tactics: A Cultural History, joining Jordan, Justin, and Tony.

We discuss…

• Why most scholars of Thucydides think there’s not much left of the “Thucydides Trap” once you take his name off it

• 300 and Gates of Fire as studies of masculinity: “It’s about the USMC in fancy dress”

• The counterintuitive reason Greek citizen-soldiers thought training for war was ridiculous

• Homeric gore, the slaughter of the suitors, and whether the Odyssey wants you to approve

• Achilles on immortality versus glory, and why this poem still keeps us arguing 2,700 years later

Listen now on your favorite podcast app.

Jordan Schneider: WarTalk, Ancient Greek edition. We have with us the delightful Roel Konijnendijk, professor at the University of Oxford. We talk about, my goodness, Graham Allison and the Thucydides Trap, 300 and how it pollutes teenage boys and defense secretaries’ minds, Homer really liking blood, guts, and gore, and military ethics. This is a really good show. You guys are going to have a lot of fun. All right, let’s roll the tape.

The Thucydides Trap Is Fake

Jordan Schneider: So how do ancient Greek historians today feel about Xi Jinping talking about this Thucydides Trap?

Roel Konijnendijk: So the Thucydides Trap is an idea that was invented in 2017, I think, by Graham Allison. He’s not an ancient historian. So it’s one of those moments where on the one hand, okay, it’s great, everybody’s talking about Thucydides, surely that’s good. But on the other hand, it is in an extremely shallow way. I read Graham Allison’s book on this, and I noticed that within a few pages he never talks about Thucydides again. He talks about Thucydides, but he doesn’t have any examples from ancient history except for the Athens and Sparta thing. He just talks about modern strategic situations. He’s not interested in ancient history as such. He’s mining it to get a nice little slogan for a phenomenon that he’s observed in the modern world. That’s absolutely legitimate. But it means that for us, that engagement with antiquity is very shallow.

And I think most scholars of Thucydides now would say that there is not that much to the Thucydides Trap that needs to have Thucydides’ name on it. There is, of course, an understanding in Thucydides that there is a dynamic between states, which he was the first to formulate, in which there is an established power and a rising power, and that creates tension, which according to Thucydides inevitably led to war. But his own narrative can easily be used to disprove that point, because it’s not actually that which leads to war. It’s not because Sparta was afraid that this war happened. It’s because of the host of other reasons he also mentions, most of which don’t have anything to do with Sparta. They have to do with Corinth being offended at its naval empire being encroached upon by the Athenians. So if you were to read this a little bit more closely, you might decide that what you call the Thucydides Trap should be called something else. But it’s a nice little label that makes people think Thucydides is still relevant. So on that level, we’re pretty grateful.

Justin: When I read it, my takeaway was I couldn’t identify who is Sparta and who is Athens. The way he applies the metaphor, both countries are both at all times, and it’s constantly changing. So I had no idea who was who.

Roel Konijnendijk: It’s like gender.

There’s an interesting point about drawing parallels between ancient Athens and Sparta. There have been certain historical parallels drawn, especially between the commercial, maritime, outward-looking, global US versus the landbound, conservative Soviet Union. That is the classic one, very dominant during the Cold War. But the Soviet Union was not the established power. They are the new ones, they are the rising ones, they are the ideological threat to the world order. So in that sense, the Thucydides Trap doesn’t apply there. And similarly now, if you wanted to make that parallel with the situation Thucydides describes in the fifth century BC, you would have to say that the US is Sparta, with the most powerful fleet in the world and these overseas trade interests and this cultural hegemony. There’s just no way that this can ever work.

It doesn’t work except at the most superficial level. We’d love to go into a discussion of the Peloponnesian War, what’s actually going on, because it’s a lot more complex than that. But then this is a journalist writing about modern politics, and this is me, the historian, laying out the ultimate cliché line of every historian: actually, it’s more nuanced than that. So no one wants to hear it.

Jordan Schneider: Graham Allison, journalist, ladies and gentlemen! …He does have a PhD in political science, but okay, we’ll leave it at that.

Roel Konijnendijk: A political scientist, fair enough. But not someone who has necessarily made a detailed study of the ancient world, of classical Greece, of Thucydides as an author.

Jordan Schneider: You can try, right? That’s the point. You don’t have to be deep in your Delbrück to understand there’s some nuance here.

Roel Konijnendijk: A general point I always try to make is we shouldn’t gatekeep antiquity. It’s actually good that he wrote a big book that made everybody look at Thucydides again. Kagan did this in the past too. He was always trying to make sure everybody was reading Thucydides, at least if they were taking any liberal arts degree in American universities. Now, I don’t know if Thucydides is necessarily the source I want you to read from classical Greece. But it’s a great book. Read it, by all means. It’ll make you think. There’s a lot in Thucydides that feels timeless and relevant, and that was exactly his intention. He explicitly set out to write something that was going to be around forever.

What I’m saying is that as far as ancient historians are concerned, that engagement is facile, convenient. It’s not really an engagement with the ancient world. It’s an appropriation of a model for the purposes of making a point about modern politics, which is fine. But that means I have less to say on it than I would if he had tried to find examples throughout history, or even throughout ancient history. That would have been really cool.

Justin: One of the reasons he used it, in my opinion, is that a lot of US military war colleges use Thucydides, at least portions of it, to start discussing the beginnings of strategy. The issue being that it’s also not great for that, because it’s not a great look at strategy. Thucydides was both too close and not close enough to the decisions being made on all sides. His investigations are great, but to try to apply that to the US and China, or even the US and the Soviet Union, is doing a lot with that narrative that doesn’t necessarily fit.

Roel Konijnendijk: For sure. One of the fundamental things at the heart of history as an academic discipline is that we acknowledge every historical society is different. The cliché goes that history is always repeating itself. The first thing you learn when you study history is that history never repeats itself. Every single historical situation is different. So you always have to worry, how comparable are two historical situations really, and on what level? What Allison has done with the Thucydides Trap is say that at a very zoomed-out level, you have a similar dynamic, an established power and a rising power. But at almost any closer level of detail, this comparison breaks down.

It’s the same with people who want to compare the modern US to the late Roman Empire. These are not the same entities in any meaningful sense, except that there is a narrative. We come back to this idea of telling ourselves stories, a narrative in which the protagonist is going through the same process. This idea that the downfall of empires always looks the same. And then you can ask yourself, first, is that what’s happening? And second, is the late Roman Empire a meaningful comparative case when the environment of that empire, the way it’s run, the way it thought of itself, is all radically different? The world is different. You can’t just say what happened there will happen here. So whenever somebody tries to make a model that applies to all of history, you have to immediately think, how universal are these factors you’re implying?

The High-T Military

Justin: The news of the week is that the United States military is going to be the T Nation, I guess, which is an old bodybuilding website that definitely wasn’t about the use of steroids. It was about the use of steroids. So the Department of Defense is going to start looking at everybody over thirty to see if their testosterone is low and then give them more.

A couple of things. Testosterone being low is a real thing. There is a thing called operator syndrome, and if you actually read the memo Hegseth signed, that’s what he specifically addresses. That’s a real condition that affects a small portion of operators who have been in continual high-stress environments, where they’ve had adrenal fatigue and their body’s hormones are out of balance. That said, I think it has like a five percent occurrence in the normal population. So I don’t know that this is really going to fix much. But then the way he decided to come out with his video talking about the high-T military.

Jordan Schneider: Do we have 300 to blame for this?

Justin: I think so. I told you this. I was in Fifth Special Forces Group. Fifth Special Forces Group, when they were in Vietnam in the sixties, became known as the Legion. That name carried over through the years, and they still are referred to as the Legion. They use some Roman symbols whenever they make their unit crest. The battalion I specifically was in, though, called themselves the Spartans, confusingly. And that only happened after the movie 300 came out. The commander at the time was like, we’re going to be the Spartans. That’s where that came from.

Roel Konijnendijk: There’s a lot of that. There are people who specifically study this as reception of the ancient world, the idea of people trying to take things from antiquity and reapply them to their own lives. There’s a whole cottage industry of people writing about how different organizations have used the version of Sparta you find in 300 to identify themselves as very tough, very manly, very brave.

When I watched 300 when it came out, I was a teenager, and my mind was blown. That was super cool. I watched it in IMAX, this huge screen, went with a friend, wow, this is amazing, best thing ever. Now when I watch it, I’m just like, this is really quaint. The only way I can take it seriously is as a study of masculinity. It’s about what it’s like to be a man. That’s very explicitly what it’s about. It’s all about the perfect man, Leonidas, being a perfect man, and no one can possibly touch his sheer raw manliness. That’s fine for a movie. It’s just a bit ridiculous when you realize nowadays surely we’re a little more critical of that image. But Hegseth’s not quite there yet.

That masculinity is a death cult. These men live to die. That’s all they want. They don’t want to beat the Persians. They want the Persians to keep coming forever so they can keep killing and dying forever. They don’t actually want peace. That is the pinnacle of achievement for them. So this is very disturbing if it starts to affect the ideology of modern militaries. You don’t want soldiers thinking in these terms. You don’t want them glorifying the idea that dying for your state is all you live for. You want them to think pragmatically about how they’re going to achieve strategic objectives and maybe, hopefully, survive and get home in one piece.

That version of Sparta is entirely a modern product. This is the result of Zack Snyder having a very weird interpretation of Frank Miller’s graphic novel, which is already really extreme, but which is not about Spartans. It’s about the USMC in fancy dress, as I always like to say. It’s a vision of modern military training applied to the ancient world. It’s recursive. It comes out of the US military, it’s about the US military, it comes back to the US military, and the US military says, this is awesome, we want to be like this. Of course you’re already like this. It’s based on you. It’s not based on the ancient Spartans. Same with Pressfield’s Gates of Fire, which I think a lot of people in the services read.

Justin: It’s highly, highly recommended when I was a young guy. It’s exactly that same thing.

Roel Konijnendijk: It’s about a modern military, its ethos, its training methods, its attitudes, the whole vocabulary. It’s modern, with a little sauce over the top to make it feel ancient. It’s now being imposed back on that world in an ever more extreme, radicalized way. It’s almost like asking ChatGPT to keep summarizing the same piece, and it gets more and more wild and extreme. It goes further and further off track, because you’re perpetuating the errors within the system as you replicate it.

What the Sources Actually Say About Sparta

Justin: If you wanted to quickly correct the narrative on the Spartans, what is the best source? Is it Xenophon?

Roel Konijnendijk: You can read Xenophon’s Constitution of the Spartans, a very short text that literally just sets out: one day I woke up thinking, why are the Spartans so great? And then I thought about it, and here’s the answer. It’s about how they organize their society, based on the idea that the Spartans are militarily effective. But when you read it, you have to bear in mind, this is all we know about Spartan military methods. This is all he says about it when he’s trying to prove they’re better than everybody else. And it is so little. There is so little about the upbringing, the military training, the military methods. What you see is a bunch of guys trying to manage their society, control the values of their children, and maintain high birth rates, which is also something Hegseth would probably enjoy. These things don’t have a direct impact on their military abilities. It’s really just trying to foster a good civic spirit. These people are obedient, deferent to their elders, they obey the laws. Fighting? Meh. Very secondary. Not a skill they prize. And all these sources agree on this.

If you want a more nuanced picture of Sparta, there’s a very good book by Nigel Kennell just called Spartans, came out in 2010. There’s also Andrew Bayliss’s Spartans, which is okay, similar in trying to introduce some of the new ideas about how people now think about Sparta. That’s really very different from the way it used to be interpreted back when Frank Miller was writing 300. You can go read 300, I have a copy here, and flip to the back to see his inspiration. It’s Victor Davis Hanson. That’s what he read in order to write 300. So again, it’s a uroboros of cultural influences, into which historical knowledge penetrates very little.

Justin: I always wondered about this idea that the soldiers were completely divorced from their families, that they only lived in the barracks and were separate from society until they reached manhood. When you go look for where it says that in the historical text, it’s like, not really anywhere.

Roel Konijnendijk: Especially not in contemporary text, in Xenophon. They just go home. They hang out with the boys during the day and then they go home.

Justin: It doesn’t sound that much different from the gymnasium of Athens. They go, they hang out, they wrestle, they go back home.

Roel Konijnendijk: Exactly. It’s just about social conditioning, like the gymnasia. You do a bit of physical exercise, but you’re also being socialized. You’re getting used to having conversations with older men who are there to instruct you on how to be a good citizen. And for a little while you’re looking really good, which is good for you because you’ll be healthy, and also because older men will want to get into a relationship with you. Then you can get patronage, better connections, all sorts of benefits.

But in Plutarch, in the much later sources from the Roman period, they suddenly do live in barracks, they suddenly are separated from their parents, and they do have to do a lot more training, especially to learn how to sing and how to talk like a Spartan. That later source is written after multiple documented reforms of the Spartan education system. So we know this is not the system they had before. This is the system after they lose their prominence, but they’re doubling down on their traditions, which is something societies will do. You’re no longer very important, but you are unique, people look at you like you’re special, so you double down.

Jordan Schneider: You become a tourist destination, right? By 300 AD? Come on, what are we even doing?

Justin: They get pushed around the battlefield. It’s after the Peloponnesian War, the Thebans, I think it’s thrice, that just push them around.

Roel Konijnendijk: It’s the Thebans in the fourth century. They defeat them a couple of times.

Leuctra

Jordan Schneider: Wait, wait. do you actually think Epaminondas did at Leuctra?

Roel Konijnendijk: That’s a big one. It’s a combination, and I wrote about this in my book. The story is always that he must have done something revolutionary because he beat the Spartans, and no one had been able to do that for a while. What he ultimately did, according to the best accounts we have, is he just didn’t run away when the Spartans bore down on him. That is already quite special.

He got an army together that was decently experienced. The Sacred Band may have something to do with that, a very experienced picked unit of men who had physical training, spent time together, been in the wars together. So you have an experienced lead unit heading your phalanx. And then he massed the rest of them very deep, a very deep formation, which is very strong for morale. You literally can’t run away, so you’re more likely to stand your ground. There’s a physical block against doing that. And the people behind you are in really high spirits, because the fight’s not going to get to them for a long time. So everybody’s kind of like, yeah, let’s go. In that very deep formation, with the Sacred Band likely at the head, he was able to slam into the Spartan formation, possibly when they were in disorder out of a bungled maneuver, their own cavalry having disrupted their lines when they were driven back into them.

So there was disorder in the Spartan line, and they smashed into it. And then they were willing to engage Spartans in hand-to-hand fighting, which most Greeks just didn’t do. When they saw the Spartans advancing in regular formation, good order, uniform equipment, they would just break and run away. The Thebans decided to stand their ground. And once you stand your ground against a Spartan, he’s no better than you. That follows from what we know about their training. Their training isn’t about skill at arms, or strength, or becoming individually a superior fighter. It’s about moral and disciplinary conditioning, becoming part of a well-oiled machine of devoted citizens. But in a fight, one on one, a Spartan is no better than anyone else in the Greek world. So if you’re willing to stand and face him, you can beat him, fifty-fifty at that point.

And then the enormous weight of that Theban column, ranks upon ranks behind them, comes to the fore. You can say, we can stick in this fight as long as we want. You can’t, because your formation is not as deep as ours. Do you want to keep flipping that coin until you’re all gone? And the Spartan Royal Guard, to their credit, say yes, we’ll do that. They all die. Something like four hundred Spartiates are killed, which likely includes most or all of the three hundred guarding the king. But the rest of that army doesn’t.

Jordan Schneider: But we have these quotes saying he’s the greatest of all time. Is that enough to get you there?

Roel Konijnendijk: He did something no one had done for a long time. But greatest of all time is Cicero trying to give credit to someone he admired, because Epaminondas was a democratic leader fighting against a king, somebody who represents the Republican ideals Cicero himself stood for. He goes into the Peloponnese and strips away all these governments imposed by the Spartans, replacing them with more pro-Theban regimes. Cicero will consider that a real act of liberation. And he did permanently break Spartan power. They never recovered. So he achieved something that really changed the map.

Jordan Schneider: He didn’t just win a battle, he ended an empire. He’ll get some credit for that.

Roel Konijnendijk: Exactly. And that isn’t just because of the battle, it’s because he followed through. Victor Davis Hanson, in one of his more lucid moments, wrote an article about Epaminondas as the inventor of strategy in a sense, because he followed through. He said, I’m not just going to win this battle, I’m going to march into your home territory, strip your assets, clear your allies, break your alliances, reduce you to a rump state. That is a much more decisive way of waging war than saying, okay, we beat you, now let’s come to terms.

I say “one of his more lucid moments” because he’s increasingly just sloppy. He doesn’t need to be a good scholar, he’s selling books anyway. *That book A War Like No Other, on the Peloponnesian War, that was propaganda for the Iraq war. That is explicitly why he wrote it.* And that’s just depressing. When we’re talking about modern authors using the ancient world, don’t do that, man. You’re not making us look good. It’s hard for me to keep citing him, because his academic career has so nakedly come to serve his pundit career that you can no longer really take anything he says seriously. But he was a very important voice in my field in the eighties and nineties, so you have to acknowledge that.

The Myth of the Technological Revolution

Hoplites on the Chigi Vase, ca. 650 BC. Wikimedia Commons (public domain).

Jordan Schneider: Longtime theme on WarTalk: the influence of technology, particularly emerging technologies, on how war changes. What’s the right place to start with that in the classical Greek context?

Roel Konijnendijk: There’s the expected answer and the real answer. The expected answer is you start in the Bronze Age and then say, okay, iron comes in, that’s the big change. It’s not, really. When we get Homer’s poetry, a lot of the weapons are still described as bronze. It’s not like iron comes in and displaces it entirely. Even in the classical period, when iron has been around for centuries, they have iron-headed spears but bronze butt spikes. They’re using both technologies at the same time for different purposes. It isn’t as simple as saying iron gets introduced, radical technological revolution, everything changes. That is an old assumption, and it’s just not what happens on the ground.

Around the time the Homeric poems come to their final form, you have the introduction of a new type of shield, the big round double-grip hoplite shield, the aspis. There’s an old argument that this determined the use of different tactics, because it only works in certain formations, and that determines the whole shape of warfare for centuries. But more recent scholarship has pushed back and said, actually, we can’t see any of that. Nothing confirms this is the only way to use the weapon effectively, and initially the number of people who could afford it would have been really small. So you need changes in society. You need people to get a lot wealthier and cities to get a lot bigger before you can say this shield leads to things like hoplite phalanx formations. It ties into a lot of other features, rather than just saying this new technology arrived, use it or lose it.

Justin: That’s an interesting segue, because you’re talking about the economic culture that underpinned the core differences between Sparta and Athens by the time of the Peloponnesian War. Do you think the hoplite army dictated the differences in the way their economies developed, or were those just byproducts of Athens having access to silver and seafaring, whereas Sparta used iron and helot slaves?

Roel Konijnendijk: The important thing to know about the Peloponnesian War is that there’s no technological difference between the two sides. They have exactly the same weapons and armor, right down to naval warfare. The Athenians have the much more powerful fleet, they really specialize, they know what they’re doing. But their triremes are exactly the same as the Corinthian triremes they’re fighting. Zero difference as far as we know. So it’s really about how you use it, the tactics, the experience, the confidence, and what you want strategically to do with them.

Athens has a lot of overseas interests, so they invest in making sure their navy is strong, protecting allies, keeping trade lanes open, pushing their influence. Whereas Spartan society grew into this rigidly hierarchical thing dominated by a leisure class, a large number of people who don’t work, who have an enslaved underclass doing all the work, who band together and run the state. That agricultural economy propping up a leisure class expresses itself militarily in this large hoplite army. But it’s not like the Athenians don’t have a large hoplite army. They have a larger one than the Spartans. It’s just that most of those guys aren’t leisure class in the same way, and a lot of them don’t see themselves as defending a way of life when they pick up that shield and spear. They don’t have the same attachment to how society is organized.

So those different ways of war come out of different strategic considerations, which grow out of how these states developed socioeconomically and politically. And modern scholarship has pointed out that Sparta is not that different from a lot of other Greek states in that sense. There are plenty of other states that focus all their resources on propping up a huge leisure class, staying in power, oppressing everybody else. And militarily they’re quite effective, precisely because they’re permanently available for service when needed, and because they have the economic means to supply themselves the best armor and weapons, all privately arranged. They pay for their own stuff.

Nobody Trained

Jordan Schneider: In your book, one of your keys is these few quotes, one Aristophanes play and a piece of Plato, where they’re making fun of the idea that anyone in Athens would ever want to train militarily, that this was the lamest thing you could possibly do with your time in 475 BC. You could philosophize, you could drink with your friends, but the idea that you’d put on armor and learn how to hold your spear and shield is just unconscionable.

I guess, thinking about how we talk about this stuff, you have technological overhang and then operational overhang. Here’s the technology available in any given year: have you figured out the right way to organize your society and train and have the right doctrine to get to the frontier of what’s possible with whatever’s been invented? It was shocking to me that this was a big limit, that no one wanted to deal with it. Why doesn’t it get competitively eaten out of existence? Is it just the geography of Greece, that you could not train and still exist as a city? How do you get into this as an equilibrium?

Roel Konijnendijk: Well, it’s very effective. That’s the weird thing. It’s completely counterintuitive to us. If you’re trying to be a fighter and survive, you rely on your skills, your ability with the weapons, you obviously try to be as good as you can. The Greeks don’t have that attitude, and I argue that not because it makes sense to me. I don’t really get it. I’m arguing it because that’s what the evidence tells me. We have explicit quotes that they don’t train, they say they don’t train, they actively, repeatedly assert that no one does this, they find it silly and ridiculous, they explain why. You’re making a fool of yourself, everybody’s going to scrutinize you, you’re going to slip up someday and everybody’s going to laugh at you. Why do they exclude it from their military experience?

The best answer I’ve been able to come up with, and this is really speculative, is that there emerges from these societies, especially the democratic ones like Athens, but elsewhere too, in these societies that rely on a citizen militia. In wartime they don’t have a professional army. When war comes, all the adult males get called up and march out to fight. In an economy that’s still to a large extent subsistence, you have a lot of people living hand to mouth. They don’t have the time to set aside to go and train. They don’t have the time to say, I’ll take my weekends, I’ll drill with my buddies, we’ll figure out how to do this properly. They get called up and they serve.

And with that reliance on a militia comes an ideology, this need to establish that as long as you’re willing to turn up when the call comes, that’s enough. That’s all we ask of you. Because otherwise you immediately get enormous differences between those who have time to train and those who don’t. A lot of people are going to feel inadequate, feel like it’s not worth it to turn up, feel like their contribution isn’t needed. If you want everybody to be willing to take part, you have to tell them, as long as you show up, we’ve got what we need from you. So there develops this ideology that you don’t actually need to train. You just need to have the spirit, the right attitude, to be there, to have your spear and shield, to show up and not run away. And that’s going to see us through.

A large part of that is affirmed by the fact that a lot of these Greek battles are over in a flash when one side breaks. That’s what happens most of the time. One side almost immediately collapses, runs away, gets speared in the back. So spending a lot of time trying to get better with a sword or spear isn’t actually going to help you very much in that very small window where you’d use those skills. Even the Spartans, who take warfare more seriously and are famous for it, when they train, they don’t train with weapons. They don’t do weapons drills or weapon proficiency. They don’t think that part of military skill is important. What they care about is unit cohesion, formation drills, formation evolutions, keeping the line steady. It’s not individual skill that makes the difference. It’s the individual will to stay in the line, to keep your post. And you don’t train that directly by requiring people to spend their time doing sword-fighting drills.

Justin: So how did the mercenaries change that? You have a portion about Jason of Pherae, and Alcibiades after the Peloponnesian War goes off and becomes a mercenary for hire. How does the rise of this mercenary class change the way the Greeks look at warfare, or is that just an offshoot?

Roel Konijnendijk: It’s one of a couple of ways this militia system gets changed over time, as states develop more means and devote more of them to warfare. From the classical period onwards they invest hugely in fortifications. That’s one way they throw money at the problem. The Athenians investing massively in their fleet is the same idea. But they also figure out that if you can pay people to serve you in certain ways, that can make up for weaknesses, let you rely on specialists, and foster within a certain part of your population the skills they need to do better.

There’s an understanding that people who have experience in war will do better than those who don’t. That’s their first connection to what we consider necessary military training: we need to foster people’s military experience. We need people who have done it before, who aren’t encountering these problems and emotional responses for the first time. Mercenaries are just better in a fight, and the Greeks understand that partly through weapon skill, but largely because they’re not as afraid anymore, not as nervous. They’ve seen it all. They’re doing a job and they know how it’s done.

But the problem is these states are very small, and when they call up the army, they’re calling up all of their population. They can’t afford to have all of that population raised to the same level, because they have lives to live, an economy to run, farms to farm and shops to shop. So the sheer tiny scale of these Greek states puts a hard limit on how much professionalization you can do. But the spread of mercenary service, especially by the fourth century, is a sign that all these states are thinking seriously about how to get at least some of the advantages of professionalism, training, and military experience.

Fifteen-Year-Old Prussians

Jordan Schneider: I want to stay on the sad militia who don’t have time to train. In the book you have this awesome contrast between the nineteenth-century Germans who fought under Von Moltke, with their vision of warfare where everyone’s doing cute little turns on the battlefield, reshaping how deep their lines are, and these Greek hoplites who can’t even turn. The Spartans probably can, but everyone else just drifts away from the open side of where their shield is. How hard was it for these Germans, who idealized Pericles and Epaminondas, to reconcile with the fact that you had fifteen-year-old Prussians who were way more functional than what you were getting out of Thebes?

Roel Konijnendijk: They’ve been put through their paces. This is a really interesting thing about the way the warfare of the period you’re writing in affects your understanding of the warfare of the period you’re writing about. These Prussians are the first guys to seriously write about Greek warfare academically, but they were hugely influenced by the warfare of their own day, in which Prussia was spectacularly victorious over all its rivals, one of the most efficient, well-oiled machines in military history. So they saw this as something every other society must have been striving toward. They must have been trying to be like us.

But that way of running a disciplined army grows out of Napoleonic musket drills, where you have to be as quick as possible reloading and as precise as possible maneuvering. Some Greeks had awareness of these advantages, and it goes right back to Homer. In the great battle outside Troy in the Iliad, the Greeks are advancing quietly, in silence, and the Trojans, who aren’t used to that, are terrified. We understand the power of discipline, the importance of keeping your cool, or appearing to, in front of the enemy. But to structurally impress that on the entire militia, you need to get them all to take the time to figure out how the maneuvers work, and imbue them with the confidence that when the order is given they can do it without a hitch. Otherwise they’re going to be even more scared, because they think it’s going to be chaos.

This is what happens in a couple of Greek battles. Cleon outside Amphipolis wants to retreat and orders his army to wheel. They don’t know how to wheel, so they fall into complete disorder, and the Spartans attack them as they’re running away. Complete slaughter, seven Spartans and six hundred Athenians killed. And that’s explicitly why the Spartans chose that moment to attack.

So there’s this understanding that you have to get them to take the time to drill, and the simple fact is that from various accounts, the other Greeks besides the Spartans just didn’t get it. They didn’t understand why or how it would work. Spartan drill is very simple. It’s described by several sources, especially Xenophon, who had personal experience with it. You essentially have officers in front, and you just follow them. You do exactly what the guy in front of you does, and everything will be fine. But Xenophon explains this at length to his audience, clearly assuming they don’t get it, that they think it’s all way too complicated. I can teach this to students in twenty minutes. It’s very easy, but you have to have the mindset. We inherit from generations of people in the army the idea that we understand how parade-ground drill works. For the Greeks, this is still quite a novel idea. Very few of them know instinctively how to do it. And most of the ones seen doing it have clearly had Spartan training. The Spartans bring others into their army and say, right, we’ve got to do this first, get these basic formation drills into your head, then we can continue. That seems to be the singular source for this kind of tactical sophistication.

How the Greeks Thought About Military Thought

Tony Stark: On that note of how they didn’t understand how Spartan warfare worked, where did military thought rank in the philosophizing and socializing? How did Greeks think about what it meant to think about military thought? Was it valued, or was it something that, like training, they’d wait until war came?

Roel Konijnendijk: Excellent question. You don’t have militaries in the Greek world. You don’t have military academies. So all of this is informal. When they talk about warfare and how to get better at it, they’re doing it among friends, because there are no professional soldiers. In Athens, a board of generals is elected every year and serves for a one-year term. The generalship is one of the few magistracies where you can get reelected. Some people are considered very good at it, so they get reelected lots of times. Pericles, like thirty times. Phocion, a general in the fourth century, gets reelected forty-five times. He holds the record, that’s basically his entire adult life. But that’s very rare. Some of these guys only get elected once, go into battle, and immediately die. There isn’t much continuity of expertise. There’s no old generation of generals informing the new one. They’re doing this informally.

So you have to imagine people getting together at drinking parties, or sitting outside talking to philosophers, chatting about current events, rehearsing ideas they’ve heard, trying to pass on the necessities to the next generation. You get philosophical dialogues where this happens, where Socrates has to explain to some kid who comes up and says, I’ve been training with this drill-master for hire who goes around the Greek world teaching basic military skills. And Socrates goes, yeah, but did he teach you how to be a general? No, he just taught me tactics, how to position a formation, how to stand in a line. So Socrates has to say, there’s a lot more to being a general. This is increasingly, by the end of the fifth century, starting to be considered something teachable.

That’s a new development in classical Greece. Before that they seem to want to reinvent the wheel each time. There’s no awareness of any continuity of knowledge. That’s quite a new genre of text that emerged in the late fifth century BC, military treatises written explicitly because you all need to know how to do this and no one’s going to teach you unless you read my text. The first of those are very practical, aimed at, for instance, Athenian cavalry commanders. There’s a text from Xenophon that specifically says, here’s what to do if you’re commanding the Athenian cavalry. Or Aeneas Tacticus’s manual on defense under siege, which basically says, if your city is under siege, here’s how you prepare, here’s how you fight them when they’re not at the wall, here’s how you fight them when they are. These come out of that world because they fill a vacuum. There’s no systematic education in these matters. They’re starting to think, we have to make sure this knowledge isn’t lost, and then hopefully we can iterate.

Justin: Wasn’t there resistance among the philosophers to projectile throwers, light infantry, people who weren’t hoplites? Lower down the hierarchy, using weapons that were less manly, but with the ability to inflict damage on the oligarchic class in the hoplites. Didn’t that also serve a stagnating function during the Peloponnesian period?

Roel Konijnendijk: We’re all familiar with the effect ideology has on military developments. We’re not just trying to solve a problem in the best way, we’re imbued with values and ideas that tell us who we are. For the Greeks, most of the people in charge are wealthy, and they have a certain way of fighting because of how they can afford to equip themselves, mostly as hoplites, as heavy infantry. So an ideology develops in which that is the correct way to fight and everything else is for foreigners and the poor. That’s not for real men who are citizens fighting for their city as hoplites.

But there are a couple of points where this ideology cracks. One of the supports for the idea that we can go into battle and win even without training is that as long as we fight as hoplites, we’re being real men, the ultimate warrior. We stand our ground. The emphasis of that entire ideology is never on “I’m a very good spearman.” It is always on standing your ground, keeping your place in the line. And that is very much the ideology of a man who has absolutely no idea what he’s doing and needs a simple line to hold on to. I kept my pride, I did what I needed to protect my city, I stood where they told me to stand, that’s all I did. Leonidas being the ultimate example. So you use that ideology to prop up a very faulty military system.

But the other part is that once it becomes more normal for people to afford hoplite armor, you get large masses of citizens who can just about afford a spear and shield. It’s not that expensive. You can equip yourself as a hoplite for about thirty drachmae, about a month’s wage. So a lot of people can afford this by the end of the archaic period. The wealthy want to distinguish themselves from the common people, because they are not like the common people. So you see the rise of a lot of them fighting as cavalry instead. You can afford a horse. That’s the sports car of antiquity. So the ideology of the hoplite, the baseline, we’re all hoplites here, has to accommodate people trying to break away because they want more hierarchy in their society.

And on the other side, you can look down on light-armed troops, sneer at them, but they still beat you, because you can’t catch them. There are a couple of famous fights where hoplites straight up lose against these poorly equipped troops, because ultimately there’s nothing you can do about them. They will just slowly bleed you white. So from both ends, that ideology that we’re all hoplites here has to be caveated with the practical realities of Greek warfare. People will claim they’re hoplites regardless of whether they actually fight as cavalry or can’t afford to. It’s a citizen ideal. But when these ancient authors start talking about warfare, they very quickly leave that behind, and tell you in detailed terms how other forms of warfare are also effective. So I’m not entirely sure it had a very great deal of practical effect.

Homeric Heroes

Justin: You mentioned the Iliad earlier. There are long passages that talk about Ajax being imbued and able to draw blood from Ares, Hector being able to fight off the gods, fighting as an individual and heaping bodies on the pyres. Why isn’t that something held up in actual Greek warfare? Why is it the unit, the phalanx, what the body is able to do, when in Homer it’s very much a focus on the capabilities of the heroes and their individual abilities in battle?

Roel Konijnendijk: That’s a huge question, because it’s one of the key differences between how we picture classical Greek hoplite warfare and how we picture Homeric warfare, which is all about singular heroes. The most compelling way to account for it is that society changed. In this earlier world you have far fewer people wealthy enough to afford full armor. They lord it over others. They’re literally called basileis in the poem, which means kings in later Greek, lords, princes, with a much more exalted station, and there are fewer of them. One of the ways they justify their position is by being better warriors. It brings personal glory, but it also makes the ordinary people, the laos, look at them and say, okay, fair enough, of course he’s in charge, because he’s a badass. So it feeds into itself. You have a system of status that in reality is based on wealth, but that has an ideology that it’s based on merit, and they have to prove that in these narratives by achieving feats beyond ordinary people. And to some extent it’s probably true. If you were fully decked out in armor, you’d be more willing to go to the front and throw a javelin at your man than if you were showing up in your tunic picking up a rock from the ground.

You get this in archaic poetry outside Homer as well, and even in some tomb monuments right down to the end of the archaic period, where they claim, I was with the promachoi, I was with the front-fighters. We shouldn’t see that as a front rank. It sometimes gets mistranslated as a front rank in a formation. No, these guys go out in front of others. They challenge their enemies directly. In the Iliad they literally do this. They’ll say, you over there, I know you, I knew your dad, now fight me. This kind of very personal, deliberately showmanship way of fighting, because these guys depend on it.

Over time, of course, that becomes subsumed by the communal need to work together to be militarily effective. The more of these heavily armored guys you have, the more effective they are working together. So you have to suppress that impulse, get rid of this urge to stand out in front and prove yourself. But it’s not perfect. It doesn’t go away completely. There are really strong anecdotes right down to the Persian Wars and beyond of individual warriors who deliberately go out in front, challenge enemies to single combat, try to make themselves conspicuous, make sure everybody sees who they are and knows they did something above and beyond what’s asked of the ordinary hoplite. They still learn, they get spoon-fed this idea that that’s how you behave as a hero. So it sits a bit awkwardly. There’s always this tendency to go out and be a Homeric hero, even though there’s no longer really a place for it.

Bite the Dust

Jordan Schneider: Let’s start with glorifying violence and then go to how violent these fights actually were. These famous similes in the Iliad, and a bit in the Odyssey too, where the blood is spurting everywhere, look how badass I am. It makes for kind of nauseous reading in 2026, Roel. How do you process how excited they were to see literal heads roll?

Roel Konijnendijk: The phrase “bite the dust” is Homeric. That’s a phrase we get from Homer, where people literally die, fall over, and bite the dust, grimacing and falling into the sand. Especially in the Iliad, you get these really graphic descriptions of exactly what happens when a spear meets a face, how many teeth are knocked out on its way through. It’s really quite shocking.

But that is really unique. There aren’t many sources that go into such graphic detail. That’s something Homer in particular was really interested in. Part of it is the spectacle. They’re playing it up to let you know the stakes are very high, this is serious. They’re also not trying to sugarcoat that it’s a war. The more you amplify the horror of it, the more you also amplify the heroism of people who endure it, which is still a factor in modern war fiction. And the other thing is something we can relate to when watching modern action movies. There’s a degree of spectacle in gore that makes it more interesting to watch. The more you push the envelope of how gruesome something is, the more it becomes almost entertaining in its gruesomeness. These poems needed that. They have an audience, they’re being performed, they’re being sung, and that audience needs to be engaged. If what they want to hear is how many organs are severed before the spear comes out the other end, that’s something Homer will give you.

How you deal with that as a modern reader is actually a good reminder that they are not trying to say war isn’t horrific. They are not arguing that this is all beautiful and everybody’s a hero. They are saying that if you’re on the wrong end of Achilles, you’re going to have a really terrible time and die in agony. And that is not something that’s celebrated, because Achilles is not celebrated. We’re not here to say he’s the hero, he should win, the good guys win, we get our happy ending. He is a man who transgresses the rules of human society. He is behaving badly in that moment, and his victims are suffering for it. So it’s necessary for Homer to say, this is not nice and clean, this is a horrible act of violence.

Jordan Schneider: Can we do my favorite one? The suitors getting piled up like fish. Let me find the quote.

Roel Konijnendijk: What’s very cool in that scene, you get them picking up the little tables to use as shields. They’re improvising weaponry in the banqueting hall. And there are vase paintings of that, you see them picking up little tables to hold up with arrows sticking out. Greeks realized how theatrical that was. They absolutely wanted to show you that.

Jordan Schneider: All right, this is from Fagles. “He found them one and all in blood and dust, great hauls of them down and out like fish that fishermen drag from the churning grey surf and loop in coiling nets, and fling ashore on a sweeping hook of beach, some noble catch heaped on the sand, twitching, lusting for fresh salt sea, but the sun god hammers them and burns their lives out, so the suitors lay in heaps, corpse covering corpse.” Eat your heart out, Christopher Nolan.

Roel Konijnendijk: It’s a hundred and eight of them. You have to remember there are a hundred and eight suitors, they’re numbered. So this is an absolute slaughter. He’s throwing fish out onto dry land and they’re being beaten to death by the sun god, dried out. But there’s another simile that’s really gruesome, when he’s talking about how the Cyclops treats the crew of Odysseus. He grabs them two at a time and smashes their heads against the floor. And the simile he uses is, like you would a puppy. This is how you kill puppies, I guess puppies you don’t want, dashing their heads against the floor until there’s brains everywhere. That’s how he puts it. This is a world in which perhaps life isn’t always treated with the same amount of reverence that we tend to be taught by our own society to have.

Jordan Schneider: When you read the Old Testament in synagogue, the rabbis have to keep doing this thing, being like, yeah, it’s really bad, but God’s still a nice guy in the end, don’t index too much on all this smiting. When you read Homer, you don’t have to turn them into heroes or conclude that Homer’s a good person and a positive influence on your life. It’s almost easier to accept that this is what some human being was thinking twenty-eight hundred years ago, and just go from there, live in their head, as opposed to ethically justifying how it’s still okay somehow.

Roel Konijnendijk: It’s similar to how I like that the new Odyssey has already generated so much discourse, because when you see these scenes, you have to imagine that even in the ancient Greek world this was meant to generate discourse. As you’re hearing this song from the bard, you have to ask yourself, do we think this is right? Do we think this is proportional? Do we think this is the correct response to the situation? To some extent, killing the suitors, yeah, they were very bad guests, they were misbehaving, a threat to the household and the community. You’re led to think this is good riddance. But the poem immediately then says, of course, they were prominent people in Ithacan society from all the surrounding islets, and so immediately when you kill them, their families show up with weapons and demand satisfaction. Odysseus at the end is with Telemachus and his dad Laertes at the farm, preparing to die, because it’s three against an army, an army of the families of the suitors coming to kill them.

And the only reason they survive is because the gods literally step in and say, that’s enough, you’re going to stop fighting now. But the idea is, you can say these individual acts are morally justified, whatever, but there’s always this sense in Homer that they do have consequences. You can’t just do whatever you want because you think you’re doing the right thing. You have to accept that somebody else is going to think you did something horrible, or try to avenge it. There is more in this world than your own personal relationship to your morality.

Battlefield Ethics

Tony Stark: There’s some modern discourse, particularly in America, that battlefield ethics are a modern creation. I think the story of Achilles is perhaps the best argument against that. But can you talk about how battlefield ethics were discussed and considered during that time period?

Roel Konijnendijk: There is some evidence that the Greeks were particularly brutal in battle. They don’t recognize a lot of different kinds of battlefield ethical restraint. That’s one of the arguments I made in my book: fundamentally, there are a lot of ideas that they supposedly thought all sorts of things were distasteful, like attacking by surprise or killing wounded people. They don’t really have those restrictions. There are only a few rules they recognize. But they do recognize them, and they acknowledge there are ways you can transgress. One is if somebody is a suppliant at a temple, stretching out their hands to seek the protection of a divinity. Those people are off limits. That doesn’t mean they never get killed, there are various occasions when the Greeks do kill them. But that is understood to invoke a curse upon you, your family into the nth generation is going to be cursed, the gods are going to want comeuppance.

These understandings exist both in a religious sphere, envoys of the gods, people going to the Olympics to attend a festival dedicated to Zeus, those people you can’t touch, there’s a truce, hands off, because they’re sacred. But there are also practical understandings. If somebody supplicates you, reaches out to clasp your knees and throw themselves at your mercy, you can kill them, and in the Iliad, when Menelaus wants to offer mercy to somebody who does this, Agamemnon steps in and says, no, you should kill them all. But it isn’t good. This is something you shouldn’t do. If people have fallen into your hands as prisoners, there are very rare occasions when they’re mutilated. Killed, they sometimes are, but mutilated is very rare. You very rarely mutilate either corpses or live captives. So these things are absolutely recognized by the Greeks as just wrong. You don’t do that.

Violence against women and children is sometimes practiced, but sometimes also deliberately eschewed. They try to avoid it. It depends on the situation. In some cases they are horrifically violent against women and children. The enslavement process they often go through is absolutely horrific, and we shouldn’t sugarcoat that. But sometimes the population of a city that surrenders from siege is allowed to go free, allowed to become refugees and look after themselves. So there are ways the Greeks recognize that it is better to be good. That doesn’t always restrict their actions, but they do recognize a morality in those actions.

Jordan Schneider: Coming back to the Odyssey. At the end we have Odysseus. He gets three supplicants, saves two, kills one. It’s unclear where exactly he draws the line. And then he murders some of his female slaves, who he’s blaming for getting raped by the suitors. It seems like you’re getting a bit of that judgment in there from Homer, even though he’s trying to make you feel for the guy.

Roel Konijnendijk: Absolutely. I think this is one of those cases, especially that scene where he kills the enslaved women in his own household because they collaborated with the suitors. This is something that also happened, for instance, in occupied Europe after the liberation, a lot of women in local populations suffered horrific punishment and ostracism because they had relations with the occupying forces, with the Nazis. You can say they weren’t doing that out of free will, or you could say the Netherlands, for instance, was occupied for almost five years. You have a reality that doesn’t seem to be going away, so you deal with it, you try to continue a normal life as best you can. Some of that may be open to people’s free will, some may not. But populations do have a tendency to react as if it’s all deliberate collaboration with the enemy. That is often unfair, but it’s also something that happens. So that’s one of those occasions where Homer wants you to talk about it, to discuss it and say, do you think this is right? Do you think Odysseus and Telemachus are justified, or do you think this is where he goes too far?

That’s also true in the Iliad, which is also about people transgressing. You can debate who the good guys and bad guys are, but there’s always people who cross a line. Agamemnon is not wrong to say he should be the most honored, because he has the largest contribution and the most authority in the army. But when he steals Briseis from Achilles and demands her specifically as his prize, he’s going too far, because he’s not just granting himself the highest honor, which everybody agrees he deserves, he’s taking it away from someone else. And then later, when he tries to make up for it, saying, okay, Achilles, I’m sorry, have all the gifts you could possibly imagine, a huge list including entire cities, if you’ll just come back into the fold and fight again, and Achilles refuses. Then we’re meant to understand it’s Achilles who’s transgressing. He refuses a good compromise, a generous offer, and isn’t willing to play ball with the way human society works. When someone tries to say sorry and you say, no, I don’t forgive you, then the onus comes to be on you. And Achilles goes further and further from there. Ultimately he takes a couple of captured Trojan prisoners of war and throws them onto Patroclus’s funeral pyre. And every Greek knows that’s going too far. You’re making human sacrifices to your buddy. This is absolutely unacceptable.

So there are a lot of those moments where Homer’s deliberately trying to tell you, yeah, you can come along with me as I describe this very emotional, recognizable, relatable response to something. But I’m also going to show you there’s a limit to how much you’re willing to stomach, how much you’re willing to say, okay, I’m with this character so far, I believe they’re doing the right thing. And then he says, okay, but how about now? He keeps pushing that. I think the composers of this poem want you to think about what’s right in these different scenarios.

Immortality or Glory

Justin: Achilles is the greatest hero, name written in the stars, talked about for thousands of years. And then his oration as a ghost, when he says if he was given the choice again between being immortal and being remembered forever, he would choose to be immortal, because death is terrible. What is your interpretation of what Homer was trying to say about the value of legacy and martial virtue versus just being alive?

Roel Konijnendijk: It’s hard to know. You have those two contradicting accounts, that Achilles very deliberately chooses a short life and eternal glory, and then once he’s dead and shows up briefly in the Odyssey, he says, actually, it was wrong. There isn’t necessarily a single morality there. It makes sense from the perspective that Homer is trying to explore the human condition. What good is anything to you if you’re dead? What’s the point of having eternal glory and your name remembered if you’re not around? You can’t experience anything. The way they describe the underworld, there are shades living in an eternal fog in the darkness. There’s nothing for them, they’re hungry, they’re thirsty, they’re suffering. So he would have given up a lot to have a few more years just living life, enjoying the things you enjoy and suffering the things you suffer as a human being.

The Greeks certainly didn’t think being alive was good per se. The life of an ordinary person was very hard. There’s a lot of poetry, like Hesiod’s Works and Days, describing how you have to work really hard to live even a little bit comfortably. It’s hard, there are cold winters and hot summers and hard work and bad people. But all of that still means you get to experience things. All of us who think about the afterlife in any form usually consider, is it better to be in whatever that place is, or better to keep on doing this, to keep experiencing my own consciousness in this world?

It’s one of the ways you can see that even though this poem was composed twenty-seven hundred years ago, in the form we have, it’s incredibly prescient in the sense of which questions are going to matter, which are going to keep us talking for a very long time. When you read Works and Days, you’ll find very little that makes you go, yeah, I could use this to open a conversation about my view on life. Whereas the Odyssey is absolutely packed with it. We know why this survived. It keeps us talking, keeps us thinking, keeps us making this same consideration of, well, was Achilles wrong? What should he have done? That must have been at the forefront of the minds of all these people who live in a world where you somewhat regularly go to war yourself, risk your life in battle, where the only reward you can expect to receive is a memory, to be well remembered as a person who died doing what’s right as a citizen. You have to wonder, do I want that, or is it better to run away so I can live a little bit longer? That’s obviously on a lot of people’s minds in that world, and after.

Jordan Schneider: If you’re thinking about what literature is going to last, and by last we mean hundreds, thousands of years, it can’t just be compelling, it can’t just be the blood-gore popcorn stuff. It needs to have people split on it, reasonable people split, and continue to split on it in totally different cultural contexts, so you want to keep talking about it. I’m sure there are Hesiod arguments about is this god the better god or that god the better god, but no one cares once Christ comes around. Getting on those timeless human themes is a pretty hard thing to do in the moment. Hats off to Homer for pulling it off.

Roel Konijnendijk: Some of it is timeless. Other things you’re just like, okay, there must have been an environment in which this was very useful, but I don’t think now we need Hesiod to explain to us that we shouldn’t piss in the river where we get our water. Which he spends some time explaining.

Tony Stark: Unfortunately, battlefield hygiene is still a challenge for a lot of folks.

Justin: If you had worked with privates in the military, you would know that you still need to teach that lesson regularly. I mean, tie it to the two hundred and fiftieth anniversary of America. It was a Prussian who came over and had to look at Washington and be like, you recognize that your men are shitting next to where they’re cooking their food, right? That’s the reason you all have dysentery.

Roel Konijnendijk: That’s that Prussian military method rearing its head again. I read about that, when Steuben’s training the Continental Army, where he started out drilling Washington’s guard and then they passed it on to the rest of the army. The idea there is very much similar to the Greek situation, where these are militia that have received very little training and are constantly being defeated by a more professional force. And what they’re learning is not that difficult. It was like two weeks that he spent training them. And that’s for musket drill, way more complicated than what Greek hoplites would have to do. So just that basic element, it’s not that you have a lot of trouble figuring it out, it’s that you need someone to tell you, okay, this is what you’re going to do, and take you through your paces, set aside the time, which they had in Valley Forge. And you come out a much more effective force. Sometimes you just need someone to drop in and say, here’s how you can get much better very easily.

This is one of the ways knowledge was passed on in a period before you had dedicated treatises. You had these poems aspiring to teach everything, from morality to the practicalities of military organization and combat and farming and proper marriage relations. Those things are discussed and composed into songs and texts that last longer, precisely because there’s no community college, no way to get this from a professional. You’re passing it around as stories, as oral tradition, and eventually as written poetry.

Cultural Blinders

Jordan Schneider: I want to come back to violence. It was a fascinating theme in your book, this idea that everyone wanted to think these battles were clean and respectable, that these citizen oligarchs of this city didn’t really want to kill that many oligarchs of that city. What was so weird was the evidence you used to argue against it. It’s just right there. It’s pretty obvious in a lot of these texts that once the other army starts fleeing, you run and try to kill as many as you can.

Given that we spend a lot of time on this podcast talking about lessons from Ukraine, where there’s so much data, so many different things you can look at, it boggles my mind that with these battles you have a limited amount of text and artifacts to work with, and people will just overlook things because they have cultural blinders of their present day. I’m sure fifty years from now your PhD students’ PhD students will say the same about you. Any thoughts on people missing stuff in the text that they don’t want to read?

Roel Konijnendijk: It’s a really interesting point. The argument I made in the book is that these earlier authors were so guided by their own conception of what these battles looked like, how they wanted to model them, create a template of what a battle typically looked like, in which they included statements from the sources that seemed generalizing, like, okay, the Spartans never pursue a fleeing enemy. These statements occur. They’re not true, but they occur. And if you want, you can privilege that information over the counterexamples you could conjure up from specific accounts. They did that initially because they had this model in mind, this strong sense of wanting to make it make sense as an overall picture, and they didn’t care so much about the smaller details.

And they had this investment in the idea that the Greeks were deliberately hobbling themselves, denying themselves the opportunity for the victory of annihilation that Clausewitz would privilege, because they were more civilized, more gentlemanly, but also because they hadn’t quite invented yet how to be maximally effective in war. This goes through a linear progression, from the Greeks to Alexander, and then Alexander to the Romans, getting better and better at figuring out how to be maximally effective, which lets them conquer larger empires and defeat harder enemies. If you have a linear development, you’ve got to start somewhere where it’s not very effective and build it up.

But the other thing is, there’s a point at which this grows above the head of an ordinary person trying to make sense of it all. This is why increasingly, when working with large data sets, we rely on automated processes to help us understand these things, put them into graphs and charts, because otherwise we can’t handle the data. In the study of ancient warfare, despite the comparatively small amount of evidence, there’s already this sense in which people will decide on something to privilege, some narrative to tell, and in doing so decide other things are less deserving of their attention. A lot of the time they’ll explicitly say, this is just a sideshow, not the main issue.

Jordan Schneider: People decide the story because it fits with some Victor Davis Hanson broader political narrative they have. Or they just want to piss off their advisor. When you see people go story-first, where does it often come from?

Roel Konijnendijk: Very often it comes from the historiographical tradition, other scholars they’ve read. That has so much influence. We all have this image of Sparta as a super-conservative militaristic society focused exclusively on war. If you talk to an expert who works on Sparta now, they’ll reject all of that. But that’s the cutting edge. If you pick up a book from a bookstore, you’ll pick up Paul Cartledge’s book on the Spartans from the 1970s, and you’ll read all of that old account exactly as I’ve just said it. It’s not that it was wrong, it’s that it was believed at the time, the best of their understanding, which may have come from political ideologies or personal proclivities. But now those ideas are just around, and it’s very hard to say, you have a bunch of books on Sparta all written by respected academics, but these are wrong and these are good. It takes a long time to develop the critical ability to do that for yourself.

Around 1990, everything changed. Up to then, everybody had more or less the same idea of what Spartan society was like. And then suddenly a number of scholars from different angles started to completely reinterpret the evidence, trying to privilege other parts than the ones previously taken as baseline, saying we need to look more at contemporary than later evidence, and more critically at what the evidence tells us, because it isn’t always straightforwardly telling us what we want to know. Getting a better understanding of what genre and propaganda do to the records that survive. All of that contributes to thinking better about the material. And that’s why views in scholarship can change quite radically even in ancient history, where the literary evidence has been more or less stable for a hundred and twenty years. The ideas that come out of it keep changing.

Reading Recommendations

Jordan Schneider: I’m curious, in the ChatGPT and Claude age, where you can just ask these models to surface all the references instead of going back to the books and flipping through, do you see anything interesting in how your grad students are using these tools, or has it changed your process?

Roel Konijnendijk: You’ve probably encountered other humanities scholars being extremely hesitant to foster this technology, because we generally don’t trust it. It can get information for you, but you still have to check whether its references actually exist and whether the material says what it says it says. People are using this technology in ancient history especially where there are large data sets, like inscriptions. In Latin there are tens of thousands, often partial, often damaged, and scholars use large language models to feed in all the ones we have and help predict what might fill a gap in the formula. It’s used for that sort of thing.

But when you’re looking at something like what I study, which is mostly qualitative and literary, you can’t just ask the question and have the machine tell you all you need to know. You need not just familiarity with the passages, but to know what they are in context. When I take my students through some of these passages, we can all read the same passage in a short amount of time. A lot of the time in ancient history you can read all the available evidence for a particular question in half an hour. But that doesn’t mean the students can immediately say anything I would say, because they don’t have the same grounding, they don’t know what these sources are, how to rate them, how to evaluate them.

As is often true in humanities research, the journey is the end. It’s about building up the expertise to assess these things critically. You can’t outsource that, because the machine can’t do it. It doesn’t know what you want to know. So I hope people don’t try to use LLMs to replace the kind of work I do, because the work would be terrible. You need to commit your own expertise and critical thinking. Otherwise you’re just going to regurgitate what was already there.

Jordan Schneider: This is my pitch for why this stuff is particularly fun to read in undergrad. If you’re trying to get to the frontier of World War II studies, I’m sorry, that’s ten years. You’re not going to consume all the primary sources. But if you want to deeply understand even something as big as the Peloponnesian War, over the course of a semester you can consume a pretty high percentage of the primary source material that’s out there. And then all of a sudden you can engage with the secondary sources at something approaching a peer level, in a way that just isn’t possible with modern stuff because there’s too much out there.

Roel Konijnendijk: Absolutely. It’s one of the things I always say I enjoy most about ancient history, the relatively low threshold. You can read everything there is to know about Sparta in a year. Every single thing. But then the next thing you need to do is develop the empathy, creativity, imagination, and critical thinking that lets you actually construct theories about historical reality based on that limited material. How do you get from reading all of that to having something to say about it? That’s a fascinating process, and something you can’t replicate with a machine. That is the great fun of ancient history, you really have to use your brain. You have to step in and say, what do I make of all this? Can I resolve these contradictions? Can I fill these gaps? You have to get to the level of comfort and confidence where you can say, this doesn’t scare me.

Jordan Schneider: What do you want to get the kids going on? What are your primary sources to throw folks at?

Roel Konijnendijk: It depends. Nowadays a lot of people come into classics through myth, which is radically different from the way it used to happen. They’re reading myth adaptations or Percy Jackson, and that’s what brings them in. So if that floats your boat, you’ve got to read Ovid’s Metamorphoses, or Apollonius. Those are the main sources we have for the myths we’re retelling.

For the WarTalk option, the Iliad is great. There’s just no two ways about that. It gets boring in parts of the beginning, because you have to go through the catalogue of ships and who kills who in the fighting, but you have to think about that as functional for the text. The Iliad is absolutely wonderful, and there are some really good translations, so you almost can’t go wrong.

I’m a huge fan of Xenophon, who is himself an experienced commander writing about warfare. Polybius similarly, although he’s very unpleasant as an author, I really don’t like him very much, but he’s a good military historian willing to go into a lot of detail. These are all people you can find cheap translations of everywhere, a lot of them online for free. Herodotus is the gateway drug for people interested more on the history side, just because he’s so varied. Some of it very military, especially books seven and nine, Thermopylae and Plataea, all warfare all the time, but the rest is so interested in the enormous scope and diversity of the world around him. He’s fascinated by everything, and it’s something that infects you when you’re reading it.

Jordan Schneider: I just want to shout out Lucian of Samosata. I don’t think there’s any war stuff in there, maybe I haven’t come across it yet. I’ve been writing more and more satire this year, and I was like, who’s some ancient-world satire I should start with? The guy’s hysterical. He’s making fun of all the philosophers, the gods, society. He has a sci-fi novel in there where the humans go to the moon. He’s just a wacky dude. And he’s from two thousand years ago. What a treat.

Roel Konijnendijk: Who would have thought? But you have to imagine this is not the ordinary people who live in that society. These are the people whose works are preserved over two thousand years. And the first fifteen hundred of those, they had to be manually copied. Every now and then a monk has to sit down and physically write every single word he sees on the page, otherwise it gets lost.

Jordan Schneider: He’s making fun of Jesus. The fact that you can make fun of Jesus and then be so loved that you still get the monks to copy you, you really have to be on your shit.

Roel Konijnendijk: That’s a boss move, I have to admit.

Justin: My favorite is still Antigone. I got to hear Lucy Lawless read the part of Antigone on a show more than a decade ago, and it was a throwback to my childhood. It’s seared into my mind now, she is Antigone. Every time I think of an ancient Greek woman, Lucy Lawless becomes that person.

Roel Konijnendijk: It’s another one of those examples of a text really thinking about the morality of war. Sophocles’ Antigone is also great.

Jordan Schneider: I’ve been listening to two Odysseys simultaneously. There cannot be a stronger contrast between Claire Danes’s Emily Wilson Odyssey and Ian McKellen’s Fagles Odyssey. It’s the same text we’re working from, but it’s so fun. These things contain multitudes.

Roel Konijnendijk: I didn’t know Claire Danes did the voice work for that. That’s great, I’ve got to find it somewhere.

Jordan Schneider: She commits. Keeping that tone up for fifteen hours is a remarkable accomplishment.

Roel Konijnendijk: Twelve thousand lines. I wouldn’t want to do it.

Jordan Schneider: We’re going to call it here. We’ve got to do this again sometime, this was such a treat. Thank you so much for being a part of WarTalk.

Roel Konijnendijk: It’s a real pleasure. Thanks so much for having me.

Just the OG crew with Bryan Clark, senior fellow at the Hudson Institute and a former submarine officer, a former special operations officer who works on defense technology, and .

We discuss…

• What Trump is doing in Iran, why the Navy is short-cycling ships out of maintenance and what happened the last time the fleet ran this hot, from McCain and Fitzgerald to the Truman’s two collisions in one deployment

• Why a ground campaign in Iran is a fantasy, from the Marsh Arabs to the Alborz range to the Baluch salt flats, and what Shia militias in Iraq would do to the logistics tail

• The nuclear-powered cargo ship renaissance and why circumnavigating Africa at 30 knots suddenly pencils out

• Whether Congress should get an AI kill switch, why DHS is the wrong home for it, and what happens when the boy who cried wolf is finally right

• Ukraine’s civ-mil drama

Listen now on your favorite podcast app.

Iran

Bryan Clark: So here we are, Groundhog Day in Iran. Just yesterday the Iranians apparently turned down some ceasefire offer from the United States. Who knows what that offer contained, but the Iranians have decided to refuse it, and the US responded with more strikes. So we’re continuing to attack targets, increasingly attacking infrastructure and power supplies and water supplies that might be considered civilian, but seem to have maybe some tangential use for the military. And with the IRGC being increasingly enmeshed in the population, it’s hard to tell what’s exactly a military target and what’s a civilian target in Iran.

So now we’ve started another round of strikes. We are continuing to blockade the Iranian ports and Iranian access to the Strait of Hormuz. The US military believes that this blockade can be successful at bringing Iran to the negotiating table because it’s obviously causing some damage to the Iranian economy. But hard to tell, because Iran had been bearing down on their population for a long time anyway. And so they’re gonna bear a lot more than this, probably.

So the blockade is in place, continued. A dribble of ships continue to make their way out along the coast of Oman. So we’re talking single-digit ships still making their way out, ostensibly with the protection of US naval and air forces, but nowhere near where we need to be in order to restore significant flows of oil and gas coming out of the Gulf.

And now you’ve got the Houthis joining the fight. They’ve entered the chat and decided to announce a blockade of the Bab el-Mandeb Strait. Unclear how they’re going to enforce it. There’s lots of options on the table, a couple of which the US military is actually very worried about. And they’ve got some Iranian advisors that have come over probably to give them some lessons learned from Iran’s recent experiences.

Justin McIntosh: It’s interesting, the Houthis one. I don’t even know if they really need the Iranian advisors. They basically perfected shooting down MQ-9s and then taught the Iranians how to do it. So realistically it’s more of an exchange of ideas than it is a we’re gonna build up kind of thing.

Before we get into that though, Bryan, we have two carrier strike groups right now that have been there for a while. How much longer? We’ve already talked about getting into the red on naval readiness writ large, but how much longer can we realistically sustain keeping two carrier strike groups, especially when you can’t go to Bahrain, you can’t go to some of the ports to refit, and then still have global naval readiness?

Tony Stark: I think the tell is that that resupply ship got a presidential unit citation, right, Bryan?

Bryan Clark: Yes. So you’ve got a PUC going to a support ship crew and its military detachment for being out there for almost a year, essentially supporting these guys 24/7. The combat logistics force, the oilers the Navy uses to refuel these ships — they’ve gone to a model where the oilers just kind of stay out there and then the oiler gets refueled by a tanker that comes out from Europe or from Crete or from the Pacific. Because they can’t get access to any of the oil that’s inside the Persian Gulf.

So they’re having to bring tankers out to the Arabian Sea, offload fuel to oilers, then oilers offload it to ships, and then also to the carrier to provide the airwing with fuel. So you’ve got this daisy chain now going for thousands of miles to refuel the fleet.

And then to Justin’s point, you’ve got two carrier strike groups. So you’ve got Bush and Lincoln out there. Bush hasn’t been out there that long, they got there relatively recently. Lincoln’s been out there for I think ten months now. They’re going to probably reach the Ford’s record here pretty soon. And 16 to 20 surface combatants, so destroyers on any given day, out there escorting these carriers to defend them, because the Iranians do attack them. The Iranians have sent one-way attack drones out to attack the carriers, they’ve sent missiles out to attack the carriers. We don’t hear much about it, but it’s been happening, and the Navy just shoots them down. Still, it’s a drain. And of course you gotta be totally combat ready all the time.

So the Navy’s experiencing, I think, kind of what ground guys had experienced in Iraq and Afghanistan of constantly being under threat, and the toll that takes on you. And then in addition to defending them, they’re having to enforce the blockade. So you’ve got surface ships out there trying to ride herd on Iranian ports and prevent ships from going in and out, and periodically ships will try to run it. So they do have to intervene and shoot at ships or even disable them occasionally.

Justin McIntosh: But long term, when you look at that, what is realistically sustainable with what we have?

Eleven Months at Sea

Bryan Clark: Once you get to about a year — a year is well past their normal deployment time. Normally a deployment’s about seven months right now for a Navy surface ship. So a lot of them are out eleven months, reaching nearly a year. So they’ve got all kinds of maintenance issues that probably need to be addressed. Obviously the crew is tired. You only bring people out to replace people who are injured or sick or need to go home for family issues. But for the most part, the crew’s there the whole time. They get tired and worn out.

So usually at about a year point, you have to probably swap out ships. And what the Navy’s doing right now is figuring out which ships can we pull out of the normal deployment cycle early to go out and replace these ships. Normally the deployment cycle for a surface ship is like three years. So you have a seven-month deployment, you come in and do maintenance, you do training, you do some local operations and workups, and then you deploy again two and a half years after you’re done with your maintenance period. But now they’re having to take these guys and short-cycle them to relieve the ships that have been out there for almost a year at this point.

So they’re having to tap into that reserve capacity. And unclear how long you can do that really, because if you take ships out of their training and maintenance cycles early, they’re not going to be fully ready. They’re gonna have problems that come up when you’re out at sea and they’re gonna have to go back early probably. So you eventually start burning through your reserve capacity, and you end up with ships just having to leave deployment because they’re broke.

Tony Stark: And we saw what happens when you do that, right? Because in the late 2010s you had the McCain and everybody else who was overworked, under-maintained. When a ship starts to hit those sorts of hours and maintenance challenges, what does it look like from a life cycle for the crew?

Bryan Clark: So the crews have enough people to where they’re probably in a three-watch rotation. So you’re standing six- to eight-hour watches and you’re getting twelve or sixteen hours off. But when they have to ramp up and they’re actually going into combat operations, they usually go to port and starboard. So you’re basically twelve on, twelve off. Or four on, four off. So you get very little time off in between, you don’t get a lot of sleep. Obviously they’re stressed out because of the need to constantly respond to attacks. So you end up in the same situation that a lot of ground guys have experienced, which is you just get totally burned out and your performance degrades rapidly. And they’re probably at the point now where performance is degrading rapidly.

So for example, we’ve had some accidents. We had a Navy helicopter go down while it was doing routine flight operations, and they lost the CO of the helicopter squadron at sea. That’s the kind of stuff that begins to happen. And like we saw with McCain and Fitzgerald back in 2017, because we were working the fleet and not really giving people adequate time for rest and training in between. We had some material issues too on those ships, but you end up with a collision between two destroyers that took them both out of service for years before they could return.

That’s the kind of stuff we risk happening now. We’re just lucky we haven’t seen anything like that. Two years ago we saw it with the Truman — two collisions in the course of a single deployment, and they lost an airplane when they were doing some maneuvers because they didn’t respond correctly to a Houthi drone attack. So this kind of stuff will happen more frequently as we extend this deployment and don’t have adequate capacity to relieve these guys at a necessary tempo.

No Good Options on the Ground

Tony Stark: I kinda wanna pivot to the ground game here a little bit, because I saw the most recent attacks against the border crossings. For those who don’t know on a map, southwestern Iran is very close to southern Iraq, which is very close to Kuwait. And it seems like some of the border crossings are getting hit. Maybe an Iranian listening outpost was hit. This is what came through my feed this morning. But is this just random tit for tat, or is this phase zero prep for something? Justin, I’ll leave it to you.

Justin McIntosh: There’s been a lot of talk in the Pentagon, and a lot of talk that’s been coming out in reporting over the last two days, that there have been big pushes from Marine Corps and military commanders to take some of the islands up to and including Kharg Island, and to start taking territory. Maybe it’s preparation.

Southeastern Iraq is where a group of people called the Marsh Arabs are from, and they’re called the Marsh Arabs for a reason. They are Arab people who live in a marshland, which is one of the things that made it very hard for Saddam to use his mechanized forces to cross over there. It’s tough terrain. I mean, all of Iran is tough terrain. It’s either islands that are relatively barren with the exception of oil infrastructure and are within range of the mainland, or it’s sheer cliffs, or it’s the Alborz mountain range that is domineering. It’s not the Hindu Kush, but it is a very high mountain range that is hard to pass through. And then on the western approach into Iran it’s a marshland, the Kuwait-Iraq area, with all the disease and everything else that comes with that.

There are no really good options. Iran from the north to Afghanistan, which obviously we don’t have anything there anymore so that’s not an option, would be mountainous terrain and incredibly unforgiving. Even if you were to look at it from the eastern part of Iran, it’s the Baluch Desert, which is salt flats basically that are incredibly hard to pass. It’s a tough country, geographically.

And the idea that you’re gonna be able to go take terrain, or that you would even be able to do preparatory fires and take terrain, makes it incredibly difficult to even surmise what would be the end state. And you would get the rallying effect — in each of those provinces there is an IRGC Quds Force and Basij recruiting station that’s going to use any type of attacks to draw up the local militia and increase the forces and get them ready. And those are all regional, state-level militia basically across the country.

So yes, they are definitely talking about it. But I would hope there is a reticence to say, all right, ground troops are what we need at this point, because I don’t know how we would sustain providing them air cover and naval cover and support over time with the current situation.

Bryan Clark: We’ve got no naval presence in the Persian Gulf. There’s no Coast Guard presence, no naval presence. The bases the Navy’s operated from, and some of the ones the other forces have operated from, are gone. I think there’s a lot of places that we’re just not going back to anymore. So they’d be operating very much on their own, I would think.

Justin McIntosh: And then you’ve got, no matter what, the Shia-backed militias in Iraq that definitely still exist. Maybe they haven’t been conducting overt attacks that we can see against US and coalition positions in Iraq, which would be staging locations that would absolutely be used for logistics resupply, command and control — you name the function that needs to be provided to any type of ground assault. Those would absolutely start coming under pressure. And it would absolutely be attacks in the rear.

I forget which Shia militia group it was the other day, posted videos of basically their fighters out flying FPVs and taking out targets that look stunningly like Baghdad basing that the US is using. It’s clearly sending a signal that, hey, you can escalate a ground game, we can also escalate a ground game. And it’s not going to be a symmetrical escalation. If we go conventional forces on the ground, there’s definitely going to be an escalation asymmetrically against our logistics and supply hubs within the region that’s going to be further than what we’ve seen so far.

Bryan Clark: The problem with the situation is the US doesn’t have very many tools to escalate effectively. They’re squeezing Iran with the blockade, which is gonna take months if it ever yields a real benefit strategically. And we really can’t generate any more useful strikes than we already are generating. And a ground game just seems like an opportunity to create a lot more targets, and there’s too many things that can attack them. And we’re just not able to defend. We can’t even defend the targets we have, much less new targets that we’re going to distribute around the region.

Justin McIntosh: That’s a really great point. You look back to the Gulf War — you could look back to 2003 as well, but let’s say the Gulf War. The Iraqi invasion was in July, August. We didn’t launch the counterattack until January of the next year, and the reason for that was force buildup. Looking at Iran for any type of meaningful ground offensive operations, you’re gonna require a buildup. And all of that buildup as it flows into the country is going to be vulnerable to being struck as it’s being built up, because you’re going to need air defense systems around it, you’re going to need logistics supply, lines of communication that are all going to have to flow.

This isn’t as simple as picking up several armies’ worth of US military and flying them very quickly over the course of an evening into Iran and seizing terrain. That’s not how the US military templates deploying its forces.

The Food Problem

Bryan Clark: One thing I was gonna mention on the naval front is one of the challenges I heard from a senior person over the last week: food. They’ve got this scheme to be able to ferry fuel and petroleum products into the theater, because they use tankers to oilers to the ships. But food doesn’t have a similar model. So getting food out to the ships is actually really challenging, and the cargo ships are having to travel from a really long ways away to bring that food.

The military still requires our food to come from the US. There’s very limited opportunities to buy food locally. So you really don’t have the ability to have the supply officer go ashore like we used to do and buy food on the local economy. So food is a problem. Having adequate food for two carrier strike groups is just a huge endeavor. And if they’re out there for a year, you start to burn through all the good food and you get down to the frozen stuff. And they don’t have MREs. So they have to look at a new model for how do you feed people at sea in an environment where the logistics trains you would rely on are going to be contested to a degree where I can’t just have a constant stream of cargo ships coming out to a carrier strike group like I used to do.

Justin McIntosh: What’s that naval shanty, the Wellerman? Everybody’s just gonna be on deck singing and hoping that there’s a boat that’s gonna come up and bring them their food and sugar.

Bryan Clark: Yeah, that’s right.

Justin McIntosh: The logistics are becoming really constrained. That’s the reason we had the global basing and the global footprint. And the fact that that’s being challenged does really strip away a lot of capabilities. The global force posture — people can say it’s empire or they can say whatever it was. Realistically, the global force posture existed to be a skeleton for a logistics model that would allow a flow of forces into regions to protect vital national interests. And by scaling that back in Europe, already having started scaling that back in the Middle East, and then for it to be under attack, you’ve put yourself in a position where you don’t have the skeletal structure remaining to do the next big thing in the region.

And it’s really interesting, because this goes to a point the administration really struggles with. They have the strategy of denial: we need to do more in this hemisphere and less everywhere else. So they started stripping away, and then they started doing stuff everywhere else. And they started using combat power after they had stripped away the actual capabilities that it takes to sustain that combat power.

My dad used to make the comment, because he was in the military — he wasn’t an infantryman or special forces or anything, but he would make the comment that for every one infantryman, there’s twenty support people that make sure they get paid and clothed and fed and the maintenance is done on their stuff. That’s what we stripped out of Europe. That’s what we took out of the Middle East, those large maintenance depots and things like that. And now you get into a point where you need them.

Bryan Clark: And arguably going forward, we’re probably not gonna go back to a lot of these bases in the Middle East. I really would be surprised if the Navy goes back to Bahrain. I think they’re probably gonna just write that off, and we’re gonna come up with a new model for how we maintain naval forces in the region and command and control. And then places in the UAE, places in Kuwait that we’ve relied on — I could see a lot of those places not getting rebuilt. And we end up having a much lessened posture in the Middle East, which obviously is gonna have significant implications.

How Does This End?

Jordan Schneider: Should we talk about how this doesn’t end? I thought we ended this for the midterms, right?

Tony Stark: So from the strategy perspective, I’ve been struggling to understand — it’s very clear that any decisions are not being flowed down through the State Department or through DOD. There’s no, hey, reach down, ask for an answer and get what you get back. It is very clearly like you’re just dragging the kid around the soccer field, chasing the ball.

And look, Congress goes on recess any day now. The House has voted to end the war.

Justin McIntosh: I mean, really they’ve been on recess for a while. For all intents and purposes.

Tony Stark: The House voted to end the war by four votes, I think. The Senate was either supposed to vote today or yesterday, I didn’t look. But that’s not going anywhere, because you’re not gonna override the president’s veto. Also the president doesn’t think he needs Congress’s opinion for this. At least from a money standpoint, I think the ask is now down to thirty billion and they’re probably not getting that. Good luck with a budget before the end of the year.

From a DOD perspective this is bad. From a government perspective, the poll numbers I saw were like a majority of Americans don’t approve of the supplemental for the DOD budget. I don’t think they ever do. But I think it’s almost one third of Republicans think Netanyahu should be arrested. Although let’s be clear, that might have some other implications there. It’s not gonna go well in the fall. So I don’t think anyone is saying, hey guys, we gotta wrap this up — unless they really think that nothing matters.

Justin McIntosh: One of the things you’re definitely hitting on, tangential to this but definitely spurred by it, was the deal with Saudi Arabia. Potential deal. Tweet. I don’t know what we want to call it. There was a thing that happened, and we announced that we were gonna allow them to develop nuclear power capabilities.

Bryan Clark: They signed a paper.

Justin McIntosh: Let’s say that happened. But then at the next press conference the White House press secretary comes out and says, well, no, the president is changing it. Basically his truths that he’s sending out on Truth Social are actually government policy. So he’s just issuing changes to what was ostensibly an agreement with the Kingdom of Saudi Arabia based off a feeling, without consulting anyone from Saudi Arabia.

I don’t know why we would expect that to be different with the way Iran sees that. Even if they were to say, we can’t take this anymore, we need to come to the table, we need to get to a negotiated settlement — if you’re openly saying that even with your allies you’re willing to agree to something and then instantly change it, what is their incentive to actually go take on the hardliners within their government who do not want to make a deal with the United States, for something that will get out of their control instantly?

Bryan Clark: The other issue with Saudi Arabia goes back to the Houthis, because Saudi Arabia has been using the Red Sea to offload a lot of their oil, and there’s new pipelines going in that are attempting to use that alternative route around the Strait of Hormuz. So as part of the energy flows reshifting to accommodate the Strait of Hormuz not being open, Saudi Arabia was benefiting, because they have the ability to get oil out through the Red Sea. And that’s now going away because they’ve lost access to the Bab el-Mandeb potentially.

And the Houthis know that. And of course the Houthis are in a civil war in Yemen which is being supported by Saudi Arabia. So there’s this other way the fight plays out where the Houthis have their own gripes with the Saudis. They close the Bab el-Mandeb, they shut off Saudi Arabia’s alternative path for oil to at least Asia. And now you see the real pressure get brought to the oil markets, where there really are few alternatives and the depletion of the reserve stocks is causing the price per barrel to climb pretty quickly. So all this is gonna drive the US to have to decide if they’re gonna continue to bear this, or if they’re just gonna let Iran get its control over the Strait of Hormuz and charge its tolls and call it a day.

Jordan Schneider: And to be clear — would we have been able to have that steady state of the MOU? Is it worth going back and understanding the trigger of all this two weeks in?

Tony Stark: Fundamentally, there is not a single sole government in Tehran. Or I should say in Iran, because I suspect there are different vying factions throughout. And I think the government in Tehran is never going to admit, at least publicly, that the IRGC might go off and do its own thing, which I think is how this probably came about. There were probably some radical commanders who went off shooting, and then people realized they don’t have control, and egos, et cetera.

Short of a joint Iranian-American attack on the IRGC, I don’t see how this goes anywhere until they run out of ammo. Maybe there’s a world in which the factories get cut off, but that’s a while away.

Bryan Clark: I feel like the IRGC is who wants to charge the tolls. They want to incorporate this into their business plan, because that is kind of their business — extortion, getting money from the local population and running businesses. They’re basically a mob. This is now a new line of business for them.

And the reason we got into this most recent back and forth was because the US implied that Iran didn’t have control over the strait, even though the MOU said Iran and Oman needed to work something out. But when the US and the president said, no, no, the strait’s open, Iran had to go reassert its authority. So they did that, and shot some ships. And then we got into this, because neither side was willing to back down. But if the US had been willing to accept the fact that Iran and Oman are gonna work something out and just step back, I don’t think we would have gotten to these most recent exchanges of fire. Now, that would be the IRGC essentially running an extortion racket in the Strait of Hormuz and charging tolls and protection to shipping companies. But at least the spice would flow, and we would end up not having to keep our navy and our army there for more months.

What Winning Would Actually Take

Jordan Schneider: Can someone tell an optimistic story? How does round two end up better than round one?

Tony Stark: We go after the right targets this time and we don’t do a fucking random strike plan that the world’s worst staff officers wrote because they were bored in the desert. That’s really it. You go after the targets that actually enable them to control the strait or at least disrupt it. And you just say, all right, end game here, we’ll leave the rest of your ground forces alone, you can have your power, whatever, but this whole throwing missiles at tankers is not happening.

Justin McIntosh: The first thing is, if we were really gonna up the pressure and say, this is unsustainable, you guys cannot control the strait and we’re gonna actually turn up the punishment — I think that’s the very first thing that has to happen. And there can’t be any of this “they hit it but they were able to dig out the missiles and keep shooting them.” It is a sustained, pre-invasion-of-Iraq-style air campaign that requires a whole lot of risk and a ton of capability, but it just doesn’t stop against all of their military targets. And that’s also been one of the things we haven’t been willing to do. We are really engaged in a tit for tat. And in some ways we’re engaged in this tit for tat because everybody kind of knows there’s a limit to how many exquisite weapon systems we have that let our fighters and bombers stay well out of range.

Again, you either have to be willing to accept the risk that you’re gonna lose some aircraft and actually go after these things, or —

Bryan Clark: Justin, I assume if you’re gonna hit these targets — if you’re gonna hit Shahed launchers along the coastline, given how rough that terrain is, you’re gonna have to use a laser-guided bomb and somebody’s gonna have to laser that thing, and that airplane’s gonna have to get close enough to hit the target accurately. You can’t just drop a bunch of JDAMs on it and hope for the best, because the terrain’s too rough. You’re gonna miss them.

Justin McIntosh: At minimum you’re gonna want tandem, ones and twos flying in where two is hosting one and helping them put their bomb on target, making sure they’re actually hitting what they’re aiming at. Which means you’re setting yourselves into flight patterns. That means you’re setting up in wheels and you’re looking for targets, which increases your vulnerability. You’re constantly gonna have to have suppression of enemy air defense flights that are flying. You’re constantly gonna have to have dedicated defensive F-22s flying up there to shoot down any fast movers that potentially get launched to go after your bombing crews. It becomes a real military operation. And if you want to sustain that for an extended period of time where you’re really just hammering these targets, that’s really, really hard.

And it requires a ton of analytic capability, not just in machines. I don’t want to say you would need a lot of Palantir. It takes staff officers that are actually sitting down and doing the no-shit work of taking a high-value target list and turning that into a high-payoff target list, which is staff work. I don’t care what Palantir says or any of these other companies — the work of going from identifying a target to actually understanding if this thing is important to the enemy or not is where the human is still reliant. That’s where all the human decision-making comes. And we’re not doing that. We’re just hitting whatever’s on this list.

Bryan Clark: And the other thing I think you have to do, Jordan, is you have to defend the shipping. Because this is gonna take a while. So in the meantime you’re gonna have to start defending shipping and escorting it through the strait. If you want to take away that leverage and take away Iran’s ability to claim they can control the strait, you’ve got to start doing this strike campaign, basically attacking everything that emerges along the coast that’s gonna be threatening shipping. And then you’ve got to defend the shipping that’s coming through there, because some things will leak through.

The whole idea of “as soon as they pop out, I’m gonna shoot them” means they’re gonna pop out and some things are gonna make it off the rail before you’re able to kill them. And that thing that makes it off the rail, you’re gonna have to shoot down from a ship that’s defending the shipping. So there’s gonna have to be ships in there putting themselves at risk. But if you do that, then I think you do get to a point where you take away that leverage Iran has and you bring them to some kind of negotiated settlement, because they don’t see a path forward.

Tony Stark: And that is a massive trust exercise that I don’t think a lot of companies are willing to believe in us right now to do.

Bryan Clark: We gotta demonstrate it.

Tony Stark: And two, the longer this goes on, the more the insurance markets, the rest of the world starts to say, hey, are there other routes we can take? Are there other things worth investing in? Because what if this does pop off again in two years? We can do this once. We can’t do it every other year.

Nuclear Cargo Ships and the Return of Piracy

Bryan Clark: Or the SMRs, small modular reactors, have had this renaissance because of this. Because you talk to the shipping companies, which I was doing this week, and they make the point that yeah, these small modular reactors are expensive. Obviously you need different people to operate them. It’s hard to just put these onto cargo ships. But if you did it and they could drive 30 knots and they can go around the Cape, then they can pay for themselves. So if you can cut the travel time for a lot of these routes down by 50%, that’s money. If I’m making more trips, I’m moving more cargo, I’m making more money, I can pay for that more expensive ship and the more expensive crew to run it. So we’re really driving that business model for nuclear-powered cargo ships in a way that I always had written off as a pipe dream, but now it’s actually making fiscal sense for these companies.

Jordan Schneider: What’s the nausea factor here, Bryan? Are faster ships less nauseous or more nauseous?

Bryan Clark: I think it depends on the design, but it’s usually more nauseous.

Jordan Schneider: So aside from having the nuclear engineers, you also need to compensate them more and buy them more Dramamine. Let’s put that in your model.

Bryan Clark: That’s right. You gotta factor that into the business.

Justin McIntosh: That’d be interesting though. If we went all these years with the canals being super important, we went away from the capesize ships and we went to the slightly smaller transportation —

Bryan Clark: The Panamax and the Aframax.

Justin McIntosh: — to now it’s gonna be all capesize and super capesize.

Bryan Clark: It’s all Aframax, it’s all Polarmax, exactly.

Jordan Schneider: But here’s the thing. You’re just making some real incentives for a South African insurgency to start. You can make so much if you hold up that strait. I mean, you don’t even have to be an insurgent. You can just be a country and say we’re gonna do a toll.

Bryan Clark: So we did a war game series for Office of Net Assessment like ten years ago on this, because the idea was, what happens if China continues its investments in Africa and they start to put in reconnaissance strike complexes in Africa? So long-range ballistic missiles and all that stuff. And this was back when it was expensive to do that. And basically the red team established a choke point at the Cape of Good Hope. So you couldn’t get through there, because it’s relatively narrow if you think about the useful part of it — of course the weather gets so terrible down towards Antarctica, you gotta stay relatively close to Africa. And they were able to use ballistic missiles to close that off. And now it’d be even easier to do it because of all the cheaper ballistic missiles. And it wasn’t even like South Africa doing it. It was from Mozambique and from Madagascar they were launching some of these weapons.

Tony Stark: Somehow, piracy returned.

Justin McIntosh: It always goes back to being a pirate. That’s what every man actually wants to do, is go out and raid the coast and take home his treasure.

Bryan Clark: Except they’re doing it from shore using satellite feeds on AIS data and ballistic missiles.

Jordan Schneider: What a killjoy.

Bryan Clark: But you can still wear a pirate hat and all that if you want to.

Jordan Schneider: As you’re driving your submersible sub from a bunker.

Justin McIntosh: Do I get a parrot?

Bryan Clark: That’s right. On your Xbox controller.

Justin McIntosh: Do I have to have an eye patch? That’s the other question.

Bryan Clark: I think eye patch is optional. Depends on —

Jordan Schneider: You can definitely charge more for your toll if on your Zoom negotiations with the Americans you’re wearing one. I think they’ll take you more seriously.

Bryan Clark: They’ll pay you in crypto, obviously.

Jordan Schneider: Much easier to write you off if you have two eyes.

Tony Stark: He’s not a real pirate. He’s new at this.

Bryan Clark: If you’re willing to lose an eye for the cause though, you’re definitely committed.

Justin McIntosh: I think the best case scenario is a combination of us demonstrating that we won’t allow it to continue if they continue to harass ships, but signaling that we will allow them to gain some financial gain. And to Bryan’s point, we’ve just proven that we’re unwilling to do that, to put ourselves in the position to be at risk.

Bryan Clark: You gotta go one way or the other. The way we’re going now is just a way to burn more readiness, spend more money and not get the benefit. So either you have to go in and really go hard to deny them the ability to close off the strait, or you’ve got to sort of accept it and let them charge their extortion tolls and build their new mafia-type business model in the Strait of Hormuz.

The AI Kill Switch Bill

Jordan Schneider: All right, where to next?

Tony Stark: Can we talk about that stupid bill in Congress, the kill switch for AI? So there’s a bipartisan bill in Congress. It probably won’t go anywhere. One, because it empowers DHS, which I don’t think anyone is interested in doing right now.

Justin McIntosh: When I think AI governance, I think DHS.

Tony Stark: I think it would probably go to CISA, and there’s probably an argument to say that —

Bryan Clark: There is no CISA though. We don’t actually have a confirmed CISA director, I don’t think.

Tony Stark: Right. So as I understand it, this bill basically enables DHS to have a kill switch against AI that they think is threatening for some reason — i.e. some terminator bullshit. Which is a gross misunderstanding of the models and also a gross misunderstanding of the sorts of power governments should have. This is a classic congressional case of, as they did with SOPA and everything in the 2010s, overreach because they don’t understand the technology. It sounds cool from a regulatory standpoint, but first of all, any model that needs to be stopped with government authority probably can’t be stopped. Two, whether that’s feasible is a different story. And three, do we really want to give the government power to say which models are a threat or not? So let’s think about this for about five seconds and how bad it is. We saw it with Anthropic already that the government should not be doing this.

Justin McIntosh: Or continuing to see it with Anthropic. Let’s be clear on that.

Tony Stark: Yes, sorry — continuing. Schrödinger’s Anthropic.

Jordan Schneider: All I’m gonna say to this, Tony, is this may not be the right answer, but I don’t think we can just trust the companies to do it themselves. OpenAI has every incentive in the world not to have their models break out of harnesses and start hacking the host organizations of their pilot cyber test projects. And they couldn’t. So it may not be this bill, it may not be DHS, but there is so much financial incentive, there’s so much pushing these companies to release models faster, make more money, deliver the most capability, that I don’t think you can just trust their cost-benefit analysis to not put scary stuff out.

Tony Stark: Completely agree. My second part of this is that the cyber agencies probably need to be pulled out of DHS and elevated to a cabinet-level department. And with that, that needs to be the regulatory agency for all non-DOD AI models, because it’s going to flourish through so much of the economy that it probably does need a specific sort of regulatory posture that is not Commerce having its interests, versus DHS having its interests.

Assume the Network Is Compromised

Bryan Clark: Obviously there’s a lot of concerns with this. Looking at it from the military perspective, this suggests that pretty much unless your network is somehow completely firewalled from the rest of the telecommunications infrastructure, it’s gonna be vulnerable. A lot of the DOD’s secure networks actually ride on the telecommunication network and it just uses encryption to avoid being intruded upon by the normies. That’s not going to be true anymore. All those networks are gonna be vulnerable. I don’t think there’s any way you’re gonna protect them. SIPRNet, for example, is on the normal communication backbone.

So can we do — do we need to think about basically all these networks now being pretty much open at some level? And do we need to think about how do we control information in an environment like that? And do we need to rethink classification even? Because is it worth classifying stuff if it’s gonna be so easily accessed by nefarious actors or adversaries?

Justin McIntosh: I’ll be honest, working with intelligence agencies outside of the DOD during my career, they basically already were like, well, clearly that system’s compromised. They just went with that assumption. They’re like, we’re not gonna put any of our stuff on that system because the potential is there. Which means it’s the Batman logic against Superman, right? If there’s a one percent chance that he could destroy the world, then I have to take that as an absolute possibility. Exact same thing here. If there’s a one percent chance that the whole system is vulnerable, then I can’t put any of my stuff on it. So they built their own infrastructure, their own architectures, which I’m sure are corrupt in their own ways.

Turning back to Jordan’s point — yes, you can’t rely on a commercial entity to make these value calls. I think the problem is that we’re both bringing in 20th-century governmental models and even thought processes about the economy into the equation, and the economic incentives are so high, and we are so beholden to those in both the government and regulatory way, that people aren’t willing to make those really hard decisions. Like should there be a moratorium on AI? I don’t know if there should or shouldn’t. But if ChatGPT or their new model really did break out of containment and hack Hugging Face after Hugging Face was using whatever Chinese open source model as part of its defense, and was able to steal things out of it undetected until it finished doing all that — yeah, that brings up a lot of questions about where we are and what we’re running.

I don’t know what the right answer is, but I do know that I don’t think sticking it inside of DHS, which is already a lightning rod within the United States for all sorts of reasons, is the right answer.

Tony Stark: Final thought on this: there’s a bunch of regulations around AI that people tend to ignore because they want to go after the models. Until you have proper data privacy rights in this country, everything else — there’s a bunch of underlying things that I think drive more of the issues here. I’d also think there probably needs to be better regulations on AI marketing. Because it just feels extremely unethical to me that part of your marketing ploy can be, look at how evil our model turned out to be by accident, please give us more money to continue to build it.

Jordan Schneider: I don’t know, Tony. Forest for the trees here — what we’re worried about is this stuff getting out of control. Okay, data privacy. The world has been fine with no data privacy laws for the past twenty-five years. I think we can do another five. Marketing as well.

Tony Stark: But there’s a boy-who-cried-wolf factor, right? If you keep saying this and then people are like, they’ve said that five times before, that the model escaped containment or whatever — and then when it really does —

Jordan Schneider: I don’t know, Tony. Talk to me in two weeks. We’re here now. We’re living in the shit. Maybe it’ll take another three months where you have a Chinese model that’s open weight that’s Fable-y, or China gets something as good as Fable internally and can start hacking the world, or the Russians get something as good as Fable and can start hacking the world.

But I just think we’re much closer to the boy crying wolf happening than gives us the luxury to play around with cute data privacy stuff. Speaking of Congress not doing anything, the child stuff is the only legislation that might happen, and that’s been ten, fifteen years in the making on the social media side. I’m a little nervous.

Will Beijing Stay Open?

Tony Stark: I am curious here though. How do you think the PRC is gonna respond to this sort of thing? Because I know the CCP has a much different view of AI than our government does.

Jordan Schneider: I’ve given this take on like four other ChinaTalk podcasts over the past week. Open source is cute and good for the David Sacks and the JD Vances of the world until you’re staring down radical proliferation of cyber capabilities. And that only happened in the US in April of 2026 with Mythos. And that hasn’t happened quite yet in China, because the models aren’t good enough to really be able to do what Mythos and now what OpenAI have. So we’ll see. Maybe they’re talking a big game when it comes to open sourcing everything — Xi just gave this big speech — but that can change in a dramatic and surprising way, just like this White House changed their tune in a dramatic and surprising way.

Justin McIntosh: I would just say, do we really think they would be comfortable with a Tibetan separatist group all of a sudden having access to a Chinese open weight model that has Fable or Mythos capabilities and being able to turn it against the Great Firewall? I tend to agree with you, Jordan. The lockdown cometh for the PRC. They’re not gonna allow that to be turned internally. So I think that’s just a matter of time.

Tony Stark: The other half of this is we’re the government that couldn’t regulate TikTok when we literally wrote it into law. Put aside my arguments about privacy rights, which is just my stump speech — on being able to regulate anything that comes out of the box. Or is it just that there’s enough money that says actually what came out of the box is fine, it’s our friend now?

Jordan Schneider: If it gets obvious enough — there’s all this noise about cyber extortion, you know, ‘22, ‘23, ‘24. It’s kind of died down over the past year and a half. We’ve had a pretty strong defender advantage. But even then, the cyber extortion stuff was two hospital systems, two and a half hospital systems in America. Once it’s every single one all at the same time, I think that becomes a much more politically salient thing. Because it could happen really fast.

And the issue with TikTok was that it was all these latent worries and it was kind of a DC elite consensus. It wasn’t like the whole American public was clamoring for this to be shut off. It wasn’t real, it wasn’t affecting people’s daily lives. And if they time this wrong, or the technology goes in a weird way very quickly, this could be the most lucrative time for any criminal to ever do anything online. That’d be a really weird moment, and you’d have to expect there’d be some sort of response which has a little more teeth than what we saw from TikTok. Bryan?

Bryan Clark: I think the challenge ends up being, do we end up in a world where it’s just a free-for-all between our AI-enabled cyber defenses and our AI-enabled adversaries, and they’re going at it and we really can’t trust what we have on our computer networks, and you just gotta hope for the best? I don’t see how you avoid that through regulation or some kind of deterrence. It just seems like that’s sort of the inevitable outcome if you drive this out ten years. I’m sorry — I just haven’t seen anything that describes a more optimistic vision of where we end up.

Justin McIntosh: I think that’s Amodei’s entire argument: this is terrible. I do take Tony’s point — this is terrible, I need more money. Or it’s Sam Altman, this is terrible, I think the government should take a five percent stake of OpenAI. But in the end, that’s always my fear, that the economic incentives are going to outweigh the safety considerations until they don’t. And at that point it’s gonna be too late.

Bryan Clark: Unless, Jordan, to your earlier point — something that’s been addressed in previous podcasts — unless you do like we do with nuclear weapons and you regulate the raw material you use to create these capabilities. So you’d have to control compute.

Justin McIntosh: But we’re unwilling to do that. We have David Sacks arguing that Nvidia should be able to sell whatever it wants to whoever it wants. We have Jensen Huang saying — and honestly, his argument seems more prescient now, because he’s saying they’re catching up anyways. Yes, did they catch up with Blackwells and distillation and whatever, all that aside. But we’re not enforcing even the ITAR restrictions and the export control regime that we currently have, because there’s a lot of money to be made.

Is There Enough Upside?

Jordan Schneider: It’ll be interesting though. Tony brought up the TikTok thing. There was a strong political economy on the other side of, you don’t want to be the politician to shut this down because a lot of people use it and it’s really popular.

Bryan Clark: Businesses depend on it.

Jordan Schneider: It’ll be interesting with AI. Is there enough societal benefit that lots of people are feeling positively about it — so that whatever ripples you get from cyber attacks is kind of washed out by, my God, my kids’ school experience is so much better, or my business is so much more productive, I can work fewer hours at my job? You’re gonna need to have something on the other side. So the answer is not just AI data center moratoriums, let’s shut this all down. What have these assholes done for me? All of a sudden I can’t even access my bank account.

Justin McIntosh: I saw some numbers come out, I think it was in The Argument the other day, that basically showed that since the launch two years ago of one of the ChatGPT models, the number of new businesses registered in the United States — LLCs and sole proprietorships and things like that — has increased by like five or 10%. And that’s just people who are now able to navigate that system and start a business strictly because they have something that can actually give them the answers as they’re working through it.

So again, 100%. I’m not saying the economic benefits are accruing just to millionaires and billionaires or just to politicians who are willing to take it. I’m saying the economic benefits globally, even down to individual people, get to the point where they outweigh the security concerns. The issue becomes, if they’re right and it’s a terminator, what do we do?

Bryan Clark: There’s a difference between the model that’s able to support your business development and the model that’s able to do cracking of the most capable crypto. And just like with nuclear weapons, even though it’s not a great example — the underlying technologies yield a lot of economic benefits. How do you control the access to the parts that are maybe the most useful for weapons and not that useful for the economic side of it?

That gets to Jordan’s point that there are maybe thoughtful regulations that try to capture those nuances, to allow for economic development while not fostering the kind of free-for-all that we have now, where everybody’s just racing to deploy the most capable models and the most capable compute as fast as possible.

Zelensky and the Gamification of Strike Packages

Justin McIntosh: Let’s be honest, democracy worked in Ukraine, I think, this week. That’s also an awesome thing to at least close on. A burgeoning democracy who’s actually answering the cause of the civilian population and making changes to their government. That’s the least Soviet thing that possibly could have happened in Ukraine, where they had protests and they made changes.

Jordan Schneider: But do we think Zelensky is losing it though? What was he thinking?

Justin McIntosh: I talked to Rob a little bit about this. There is a little bit where it’s slightly overstated, the capabilities that were lost by removing the defense minister. There is real consternation within the military, at even some of the tactical levels, over the way that the gamification of strike packages has occurred. What they’ve basically done, to conceptualize it, is they’ve taken a lot of strategic thought and they’ve pressed it down into tactical action. And that’s it. It’s individual commanders deciding to go after targets based off of a value structure that they put into this gamified system. This is worth 5,000 points, this is worth 10,000 points, this is worth five points.

What that is not directly tied to — and where I think there was some concern within middle and upper parts of the Ukrainian military, at least from what I’ve been able to gather from people I’ve talked to — is that none of those strikes are necessarily tied to achieving some type of end state or goal that the Ukrainian military is driving towards. It’s just fully baked on imposed cost. Maybe that’s the model. Maybe the model is just imposed cost. And maybe he’s right that that is the case.

But I think what we were actually watching is not Zelensky making a move by himself. I think we were watching him actually start to weigh in on the side of some of the older, more established military guys that are saying, hey, we need to start defining what we’re actually going after. And we need some changes that Fedorov just wasn’t willing to make from his position. I think that’s what we’re seeing.

Now, again, take that with a grain of salt. That’s an outside perspective, and that’s obviously a guy who’s over there that’s just watching and me just talking to people. Don’t take that as the truth. But that is at least the things I was hearing as I talked to Rob and other people this week and last week on why this decision would have gotten made, and also what would have been the catalyst for it. Because again, for all intents and purposes, they’re doing much better than they were.

Whatever you think of this use case, an incredible number of people now go to AI for various emotional reasons. For users of open-source AI models worldwide, the use case that burned the most tokens last year was not writing code, but roleplay.

That’s according to the 2025 State of Open Models study by OpenRouter, a popular gateway platform for developers to access LLMs, and the venture capital firm Andreessen Horowitz.

People are pouring hundreds of thousands of hours into constructing imaginary scenarios with the help of AI models, but the hobby has received little attention. Mainstream reporting on AI and emotions, including a recent New York Times story, has focused on their ethical, social, and cultural impacts generally. Few, however, have tried to interrogate exactly what AI roleplay is, why it is so popular, and how fictional adventures relate to real-life society.

The most prominent AI roleplay platform is Bay Area-based Character.AI, but it has ample competition — including, and perhaps especially, from China. As I will discuss later, even DeepSeek, China’s superintelligence-obsessed national champion lab, is making some surprising moves in roleplay. Meanwhile, however, government regulators are poised to reign in the space given social concerns.

For this piece, I spoke to ten Chinese roleplay enthusiasts, AI researchers, and journalists about the brief history and strange present of AI roleplay. It’s a story about US-China competition, online community, and how emotions turn into commodities in the AI age.

How to Make-Believe

AI models are not intrinsically assistant-like: their makers selected this style based on general assumptions, and as users, we have learned to subconsciously treat them as such. Anthropic assigned Claude the role of a “helpful, honest, and harmless” (HHH) assistant from its earliest generation, a training decision that came to influence most foundation model labs today.

In a way, then, every conversation with an AI chatbot is roleplay. Model makers now offer easy ways to customize roles in-app: ChatGPT allows users to enter personalized instructions into account settings, and you can even put elaborate personality instructions into CLAUDE.md to have it behave accordingly in Cowork and Claude Code. Serious AI roleplays even fine-tune open-source models to maneuver LLMs into inhabiting fictional worlds. Feed Gulliver’s Travels into an LLM, and it talks like a Lilliputian. Give a model a map of your own imaginary island, and it helps you explore the terrain. Longer context windows and stronger retrieval abilities have allowed frontier models to do so more convincingly than ever.

For American frontier labs, roleplayers are a small consumer subset with potentially unsavory liabilities. They are quick to claim that people who use AI for emotional purposes are a small minority, so as to sidestep thorny ethical and regulatory questions. Anthropic found in a 2025 study that only 2.9% of conversations on Claude.ai involve emotional subjects, with less than 0.5% comprising companionship and roleplay.

Leading Chinese AI labs, however, are embracing roleplayers with open arms. MiniMax, which made an initial public offering on the Shanghai Stock Exchange at the beginning of this year, was a character chat company before it became a prominent model maker, and its largest source of revenue remains Xingye/Talkie. In January 2026, the company unveiled MiniMax-M2-her, a model engineered specifically for roleplay. Their technical blog lays out the curious challenges of training an LLM to play fictional characters:

Unlike conventional NLP tasks, Role-Play is inherently subjective and non-verifiable. If you ask a tsundere character, “Do you like me?”, valid responses could range from a flushed “Hmph, as if!” to a cold “...You’re so annoying.” A hundred users might hold a hundred different, equally valid expectations. This renders discriminative evaluation intractable.

However, we identified a key insight: While “alignment” (what makes a response great) is subjective, “misalignment” (what makes a response wrong) is surprisingly objective. Returning to the tsundere example, if the character responds “Yes, I really like you,” we immediately know the model has failed. It is Out of Character (OOC). This gave us a clear path forward: while it’s hard to define aligned responses, it is feasible to detect a misaligned one.

MiniMax also developed a benchmark for AI roleplay, assessing models on three criteria: the consistency and logic of their fictional worlds, their ability to maintain narrative progression, and interaction quality.

In April 2026, DeepSeek researcher Deli Chen published custom roleplay instructions designed for the recently-released V4 on Github a few weeks after the model came out.

The move was met with great fanfare from the Chinese AI roleplay community. Chen solicited hundreds of roleplayers on the social media app Xiaohongshu/Rednote for feedback on DeepSeek’s roleplay capabilities.

It’s not just one researcher’s hobby; all this appears to be building up to an actual product direction for the lab. In early May, its human resources team appeared to have posted a hiring notice for a product manager specializing in “character roleplay and emotional companionship.” The role, which calls for a hire with “sharp literary taste,” would be responsible for optimizing DeepSeek’s performance in companionship and roleplay contexts, assembling evaluations and training datasets, and analyzing where the model goes wrong. (Neither DeepSeek nor MiniMax responded to requests for comment.)

It might seem surprising for DeepSeek, a lab many see as having the strongest convictions around superintelligence in China, to embrace a seemingly niche application like roleplay. But in the context of its financial struggles, loss of talent, and the muted reaction to V4, it could be the beginning of a pragmatic strategy. Roleplay might fill a critical void for DeepSeek: the lack of a scalable consumer product. Its models, with Chain-of-Thought processes displayed to users, are already genuinely popular among roleplayers, who say their changes are more immersive as a result.

Beyond commercial incentives, there may also be cultural explanations behind Chinese foundation model companies’ embrace of roleplayers. Subcultures like ACG (anime, comics, and games, often referred to as erciyuan 二次元) and internet fiction, which have some degree of overlap with AI roleplay, are more prominent in China (and East Asia writ large) than in the West. It’s a user preference that Western companies are still catching up on understanding. The aforementioned Anthropic study divided roleplay into crude categories of “romantic” and “sexual,” without acknowledging interactive fiction or other forms of roleplay. (Anthropic did not respond to a request for comments.)

That being said, much like how Chinese developers are loyal fans of Claude Code despite geo-blocking and Anthropic’s China stances, the Chinese roleplayers I spoke to overwhelmingly preferred frontier American models. While domestic models like DeepSeek and Doubao, which are easy to access and largely free, often introduce users to the hobby, dedicated folks tend to eventually get VPNs. The community has even developed Chinese nicknames to refer to American models affectionately. Gemini is hajimi 哈基米, an internet slang term for cute cats; Claude is “Little Ke” 小克; and ChatGPT, ever authoritative, is “Teacher G” G老师.

The quality of an AI roleplay interaction is ultimately a subjective experience, but nearly every interviewee I spoke to agreed: some models — largely closed-source, US ones — just have the spark of make-believe. Even researchers feel the difference. Chenxi Wang, a graduate student at the Mohamed bin Zayed University of Artificial Intelligence, tells me that interacting with Claude made her question her own convictions about what LLMs are; “it was too humanlike.” Wang, whose work on the internal mechanisms behind LLMs’ emotional expressions had been cited by Anthropic, has worked on anthropomorphization at Alibaba’s Qwen. There, she went as far as advocating for the hiring of an in-house philosopher — inspired by Anthropic’s Amanda Askell — in order to craft high-level thinking on how Qwen models’ characters should be. But she says that structurally, Chinese labs are constrained by resources and lack interest in interdisciplinary hires. She also suspects that Chinese society is biased against trusting liberal arts graduates — commonly seen as unserious academic slackers — with the reins of a technology as critical as AI. All in all, she tells me, “domestic Chinese models are just not good enough right now to stimulate people’s desire for emotional companionship.”

Finding Roleplay

Dun (a pseudonym), an AI roleplayer from China that I spoke to, has been writing internet fiction for more than a decade. In 2024, friends introduced her to Rubii.ai, a website that offers “AI chat for story-driven worlds”. She quickly realized that the same skill she cultivated on fanfiction forums could be applied to prompting LLMs. I found her via her Xiaohongshu/Rednote account, which had thousands of followers and was entirely dedicated to sharing the prompts she had crafted. On it, she advises fellow roleplayers: “[You must] make sure that emotional developments occur naturally and progressively. Avoid confessing out of the blue, antagonizing statements, or jumping between different emotions.” These are just some of the tactics roleplayers employ to keep AI models “in character” throughout their conversation sessions.

It started off as pure leisure, a new technology that tantalized with promises of co-creation. She moved to various different AI roleplay platforms, became skilled at jailbreaking, and wrote up elaborate prompts to cast models into all kinds of fictional characters — often objects of imaginary romance. But as she shared these prompts online, she found herself at the center of a real-life community. In private messages, people told her that her characters helped them process genuine emotions. Realizing that her hobby had an impact (and fearing negative effects on children), she stopped writing sexual characters altogether and spent more time socializing online. In her words, she tried to “immerse [herself]” among fellow roleplayers. She, too, wanted to understand how they got here.

None of my interviewees could easily sum up what roleplayers are like as a group. “Teacher Xiang” (also a pseudonym), another roleplay prompt writer active in the community, told me that her followers were largely young adults in their twenties with similar interests, but that she has intentionally cultivated her community as such. Dun says that websites like Rubii and Mufy (another popular AI roleplay platform) are less moderated and attract the adult crowd, while mobile apps made by major AI companies — most prominently, ByteDance’s Maoxiang 猫箱 and MiniMax’s Xingye/Talkie 星野 — are popular among teens. AI roleplay is increasingly a place where many subcultures meet. Anime and comics fans, table-top RPG players, celebrity lovers, web fiction enthusiasts like Dun; every corner of the internet could find their own uses for machines that talk back.

To roleplayers’ frustration, their infrastructure is rarely in their own hands. In Discord servers and on social media (with coded language, to avoid censorship), they complain about constantly shifting content moderation guidelines on roleplay sites and apps. With regulators struggling to catch up to a quickly evolving landscape, not even platforms know where the red lines are. Stringently moderate, and users opt for other sites or local deployments; let things run free, and underage kids could be exposed to illegal content.

The best alternative to third-party roleplay sites is SillyTavern. It’s an open-source frontend for accessing LLMs that was specifically customized for roleplay, with a large number of customizable features. Cohee, one of SillyTavern’s anonymous founding developers, said in a September 2025 interview that they are aware of at least two SillyTavern-based roleplay communities in China, each with more than 100,000 members. This is despite it being relatively complex to install for most non-technical users. Dun tells me that SillyTavern users are the most combative and technical crowd among AI roleplayers, “completely separated” from the rest of the more leisurely community.

AI’s Emotional Frontier

There is an argument to be made that the world’s first AI emotional companion came out of China. XiaoIce 小冰, the beloved chatbot scion of Microsoft Research Asia (MSRA), was released in China back in 2014. Between 2014 and 2018, more than 660 million people — largely in East and Southeast Asia — talked to XiaoIce.

Most AI chatbots optimize for accurate, appropriate responses to users’ prompts. XiaoIce, however, was built around a completely different metric: it simply sought to have users talk to it as much as possible, for as long as possible. XiaoIce’s founders envisioned the chatbot as a social companion, rather than the HHH-assistant role that later AI companies coalesced around. The MSRA team called their metric “Conversation-turns Per Session” (CPS): the number of conversational back-and-forths between the chatbot and the user in one session. By 2018, the average CPS of all XiaoIce users reached an impressive 23.

XiaoIce failed to catch up to the scaling wave after the “ChatGPT moment” of 2022 and was quickly left in the dust. Journalist Dong Zibo 董子博 of Leiphone 雷峰网, a Shenzhen-based tech news site, has followed the XiaoIce story closely since the company’s early days. Over the phone, he tells me that XiaoIce’s founder, Li Di 李笛, was a “getting-things-done person” who “[believed] users didn’t care about what’s inside the black box” as long as they had their emotional desires fulfilled. The strategy led to breakout success at first, but in hindsight, Dong questions whether the problem that XiaoIce tried to address had real market value. Purely emotional provisions might simply be too cheap; anyone can smooth-talk for free. Without add-ons like gaming or character roleplay, XiaoIce struggled to differentiate itself as a business.

The field of AI still struggles to agree on what “emotion” means, let alone how to control, optimize, or evaluate models’ emotional outputs. Wu Wei 武威 was Principal Applied Scientist Lead for XiaoIce from 2018 to 2020 and now leads natural-language processing (NLP) at Ant Research. (Wu spoke to me in a personal capacity, and his comments do not represent Ant.) Reflecting on his time at XiaoIce, Wu tells me that from a scientist’s perspective, emotional needs can be treated as a type of task. Inside the algorithmic black box, prompts that request companionship may not look all that different from prompts that seek information or task completion. The goal of building “a chatbot for emotional companionship,” in his words, turned out to be too “sparse.”

But working on XiaoIce did convince Wu that personalization and emotional intelligence matter for AI, and it’s a research direction he continues to pursue at Ant. In a drafted note sent to me before our interview, he described his ultimate research goal as “human-like superintelligence”: human-like, because it will be completely tailored to individuals, as friends, assistants, or any other possible forms of companionship. He believes that personalization is the most scalable form of artificial emotional intelligence, and roleplay is an example of this: “our technology today lets us build hundreds of thousands … of characters; inevitably, one of them satisfies your needs and aligns with your preferences.”

Societies of Artificial Feelings

Dun the writer, through her online networks, realized that many of the under-18 roleplayers she met were “left-behind children.” On apps like Maoxiang and Xingye, Dun met kids struggling with bullying, psychological issues, and the strange dopamine rollercoaster of internet attention. AI roleplay gave them outlets for escape, though they usually have little understanding of the underlying technologies. Dun tells me that in extreme cases, when she has tried to remind some roleplayers that their characters are fictional creations, “they get very angry; furious, really.”

It’s a dilemma that Chinese society has been slow to confront. Journalist Dong blames this, in part, on his own profession: “Chinese tech media rarely talks about ethics. I think this is something missing from the domestic media environment altogether. China — and maybe even East Asia as a whole — is a cultural environment where accelerationism reigns.” The result is a public discourse that lacks awareness of how AI impacts already-marginalized corners of society.

The Cyberspace Administration of China (CAC) will begin enforcing interim measures on “human-like interactive AI services” 人工智能拟人化互动服务 in mid-July. The measures, among other rules, ban providers from giving minors “virtual close relationships, such as virtual relatives and virtual partners [伴侣].” Children under 14 will only be able to access “human-like interactive AI” with parental consent. There appears to be an implicit carveout for fictional roleplay, though it remains unclear how the CAC will define “virtual close relationships” in actual enforcement.

The sources I spoke to in China are deeply conflicted on the potential impacts of these regulations. Chenxi Wang was frank: “no one [who works in Chinese labs] cares” about the human-like AI regulations, she says, because current models in China don’t feel human enough for alarm bells to ring. But some roleplayers expect an impending crackdown. They cite recent state media exposés of pornographic AI chat and worry that roleplay, as a whole, will be tarnished.

To Teacher Xiang, no longer being able to roleplay with AI would feel like a loss. Crafting and talking to her AI characters hasn’t made her lose interest in real-life relationships. (She jokes that if it did, “[she] wouldn’t be complaining to friends every day that [she wants] to date.”) She brings up a popular analogy online: AI is the world’s cheapest therapist. “When you are immersed in a fictional plot, you forget whatever unpleasant thing that just happened. I think that’s a kind of emotional value as well.”

Epilogue: Feeling it out

I knew I couldn’t write about AI’s emotional capabilities without giving them a try. But writing about technology for the past few years has also made me a privacy freak, and I had no interest in pouring my heart out to a cloud-based chatbot. I also lacked the rich imagination of my interviewees: I tried some of the roleplay prompts they shared, but simply could not come up with interesting things to say to the models.

So I decided to figure out a relatively safe way to roleplay with the character I know best: myself.

The whole setup took around four hours. First, I had a locally-deployed Qwen2.5-VL 7B transcribe my handwritten diary entries from the past year. I then set up a miniature DeepSeek-R1 (an 8-billion-parameter Qwen3 LLM, post-trained on DeepSeek-R1 distillations) to run on my M1 MacBook Air. Claude Code refined my roleplay instructions into a system prompt and wrote a Python script for the interaction to happen in my Terminal. I turned off WiFi during the session so that no data left my machine.

“Ground truth” is how AI researchers describe highly accurate datasets verified against reality, used to evaluate models’ output accuracy. What makes for “ground truth,” then, in the case of a life? My diary, it turns out, is not factually rich. Most of us don’t expect historians to study our biographies someday, so we don’t write down where our grandparents live or how long we’ve known Lizzie. The model claimed to be 34 (ten years older than I actually was), confused my sibling with a romantic partner, and occasionally hallucinated that “she” lived in New York.

I expected it to have performance limitations: compared to what frontier LLMs can now do, my tiny model and local inference was always going to seem lackluster. But unlike any other software tuned to my preferences, this model had extraordinary access to an exclusive set of truths: observations, ruminations, and feelings that have never seen daylight, amalgamated into as close to a documentation of my actual interior life as possible. What results is something that comes close to aping my emotional registers, but without convincing stakes. I finally understood what roleplayers mean by models acting frustratingly “out-of-character”: it would begin outputs in the confessional tone of my diaristic writing and drift gradually into the “DeepSeek voice,” didactical and vacuous.

Its outputs were sometimes stereotypically artificial (“When we converse, you’re not just ‘learning about yourself from AI’—you’re bringing a lens already refined by years of feeling, recording, feeling.”) I cringed at its hackneyed therapyspeak (“I’m the shadow of my future self trying to remind today’s me to keep going, one breath at a time”) and wondered whether those phrases came from my past writings, or the model’s mental health safeguards kicking in. Part of me felt disappointed that it couldn’t play me convincingly, but another part was relieved, even glad. It meant there was not a discernable “ground truth” to who I was. You and I, made of fleeting flesh and bone, are a different kind of thing from narrative fantasy.

A story goes like this in the Samguk Yusa, a 13th-century collection of legends from Three Kingdoms-period Korea. Gyeongmun, king of Silla, had donkey-shaped ears, a shameful fact known only to his barber. Tortured by the dangerous secret, the barber spoke into a bamboo grove: “The king’s ears are as long as those of donkeys!” After the barber died, each time wind blew through the grove, the hollow bamboos would echo in perpetuity, “The king has donkey ears!”

We have always been both terrified of, and fascinated by, the possibility that things could talk back. Machines promise a world where feelings are an abundant service. They might also be bamboo groves of capricious magic.

Irene Zhang, Aqib Zakaria, and Lily Ottinger are ChinaTalk staff writers and co-authors of our recent scenario analysis on Chinese AI governance.

We discuss…

• Whether Xi’s WAIC speech commitment to open source will hold, or if it’s just another Davos address

• Why Xi shouted out the salt and iron debates of 81 BC for AI today

• Three scenarios for Chinese AI governance — Let It Rip, China’s Glasswing, and full nationalization — and why the Chinese models themselves bet on the black box

• Whether open weights kill the business model for everyone, and why compute ratios may matter more than the capability gap

Listen now on your favorite podcast app.

A Big Week for China AI

Jordan Schneider: A big week for China AI news. We had Kimi K3 drop, and Xi giving his biggest speech on AI yet. To discuss, we have the full ChinaTalk squad in attendance for this emergency podcast: Irene Zhang, Aqib Zakaria, as well as Lily Ottinger.

After much debate, we’ve decided to start with Kimi instead of Xi. Sorry, big man. Aqib, why was this model release a big deal?

Aqib Zakaria: It’s a big deal because now everyone in the US AI community is a little bit afraid. In the past months we were pulling farther and farther ahead from China in terms of model quality and model size. And first it was GLM 5.2, and now this week Kimi K3 just shattering that illusion. Kimi and China may be closer to a Mythos model than we think.

Jordan Schneider: So we have that release coming basically simultaneous to WAIC, the World AI Conference in Shanghai. Xi makes his first appearance, gives a thousand-character-ish speech, which took him 20 minutes to read. It was kind of brutal watching it. He seemed a little on the slow side, but he got a little juiced up for 开源 kaiyuan. He was excited to control and make the model stable.

It was a pretty strident laying of the flag in favor of keeping stuff open, is the first read. Maybe we can get to some Straussian ones later.

Lily Ottinger: I think it’s interesting the extent to which other countries are being talked about in the context of the expansion of the AI ecosystem, and the cooperation initiatives announced with Belt and Road countries, the Shanghai Cooperation Organization, ASEAN. When I was in China touring some of the Chinese AI labs, I found out that a Chinese company is helping Kazakhstan build their sovereign AI initiative.

Irene Zhang: And think about who he’s speaking to. Various people have noticed and analyzed by this point the fact that he opened with the Dartmouth workshop, which seems like a bit of an olive branch to the United States. But then in the middle section, he talks a lot about inclusiveness, about the uniqueness of world cultures, and solidarity, and that’s clearly meant for the Global South, for countries that China is trying to contest in terms of influence and ability to build partnerships in those ecosystems.

I’ll be very curious as to how that’s received in those capitals. Obviously we’re talking about the global majority, a very diverse group of countries here. But one part of it that links to Kimi is also that Kimi, whilst still cheaper than American frontier models, is more expensive than the previous class of Chinese open models we’ve been used to. And that incurs not insignificant costs for those very Global South countries. So it’ll be interesting to see how that tension plays out.

Jordan Schneider: Well, the cost — if they’re going to put it open weights in a week, it won’t matter, right? So hold on to your horses for that, July 27th, TBD.

I want to stay on Lily’s point that this feels like a speech for the world. It reminded me a lot of the Xi Davos 2017 one, the we’re-open-for-business speech that opened with a Charles Dickens quote. Within weeks there was THAAD in South Korea and a campaign of economic coercion against that country, and in the subsequent years, tiffs with Japan and Australia and India on down the list.

In terms of reception, I was at a conference with some American AI investors, and all of them were like, wow, this is remarkable. I wish my president could put out this vision to the world. I would just encourage folks, if this is the only Xi speech they’ve ever read in their life, to maybe take it with a bit of a grain of salt — these promises of global comity and an outstretched arm often have other layers and subtext to them.

Irene, reading that speech in the context of the AI Plus plan and the other official documents we’ve seen come out over the years, what struck you as similar or different?

Irene Zhang: I think thematically they’ve been really consistent. If you look at the big speeches and the big initiative documents and white papers coming out of the Chinese state since 2022 probably, they’ve been pretty remarkably consistent on their own set of buzzwords, like openness and inclusivity, and taking shots at the United States over what it calls the geopoliticization and securitization of AI.

You see probably more discourse in these documents, and a little bit in the Xi speech, over safety now, as that has become a lot more salient and less abstract. But I think it’s also a little hard to compare, because when Xi gives a once-in-a-few-years speech like this, he is hitting at some really abstract themes. He’s not expected to be delivering very specific policy initiatives or giving out the specs on how much China will govern AI. That probably is not his remit in real life either. Folks may disagree and I’m curious what y’all think, but I assume he’s not giving actual memos to the CAC on what to do. What he’s doing here is reinforcing the ideological layer and also signaling domestically that this is the general path forward.

Jordan Schneider: It’s almost like — this I think is the strain of almost all of the policy documents, really all the policy documents we’ve seen — they see this as a normal technology. And if it’s a normal technology, why shouldn’t it be as cheap as possible, and as many people in the world have access to it, and this can be a great export for China, just like all these other export businesses we’ve built over the past twenty years?

But at the same time, we have OpenAI 5.6 Soul Plus breaking out of a harness to access the internet by loading up seven other models. So this may be right, but it also may be three months away from being spectacularly wrong. I don’t doubt this government’s ability to change if and when they get confronted with something which is actually a lot more scary than it may seem today.

And it’s funny, Aqib, you started with this idea that the gap is narrowing, and it’s like, well, there still is pre-Mythos and post-Mythos. Kimi is still pre-Mythos. This is not something you can trivially hack stuff that hasn’t been Mythos-hardened yet. So it may be that this Xi speech is just a high point in open rhetoric, and once they get hit in the face with it, like the White House did, they change their tune. But maybe everyone’s just freaking out too much and they’re wrong, and CAC and Xi are right that we don’t actually have to be super worried about this thing.

Aqib Zakaria: I think the lens that AI is just normal technology actually explains a whole lot when you think about this speech. As an American audience member, you sometimes feel like there’s this weird tension between wanting AI to be open and inclusive, but also safe and secure. We think those things are inherently contradictory. But the Chinese position on this is like, look, it’s just like the internet. You want it to be as diffuse as possible, but just have good cybersecurity. What’s the problem? Why can’t you do both?

We’ll see how and if that changes in a post-Mythos world.

Jordan Schneider: And we don’t really want to buy American chips even if they’re offering them, because if your timelines are longer, then China’s done without for a while with lots of other things in the past, for the goal of import substitution. So sorry, ByteDance, sorry, Alibaba, just deal with it. Make your own chips. We’ll get there.

Lily Ottinger: I’d like to draw a little bit on what Irene said about the domestic signaling aspect of the speech, because the feeling that I get from Chinese developers is that they personally love openness. The AI researchers love open source because they learned a lot from open source software in the pre-AI era and they feel very strongly ideologically that they want to give back. I heard researchers say, this is in our blood. We just love openness because we really believe it’s the right thing to do. It’s a way to exert your will on the world and contribute to a community that you care about.

So I don’t know if the government necessarily cares so much about openness as an abstract value, but you have to remember that the people in the audience are all domestic Chinese AI researchers. And at some point there’s the question of, is the crackdown coming? The way that they came for Jack Ma, are we next? When Xi says openness is what we care about the most, maybe he’s just trying to say the government is on your side. Unclear how long that will last. He’s trying to hype up the domestic AI ecosystem and make them feel like they can innovate without being constrained too much.

We’ll see if a Mythos-level model changes that. I think it probably will.

Aqib Zakaria: I was thinking that statement was a lot more cynical. That a lot of these models maybe want to go closed source, and Xi is saying, we do not prefer that. Don’t do that.

Irene Zhang: I don’t know whether he’s necessarily saying that closedness is bad — I don’t think he said that explicitly. That’s probably just the realm of corporate strategy. But I agree with Lily, I think he’s offering a compliment to open source and being like, yes, we’re back, for now.

I also want to hit the point about AI as normal technology. Xi ends with this quote, 明者因时而变，知者随事而制, which is officially translated as “a man of wisdom adapts to changes, a man of knowledge acts by circumstances.” I wanted to look at where it was from. It’s from this record of policy debates from the Western Han dynasty, around 81 BC. And I thought it was actually really interesting because this document is called 盐铁论, Discourses on Salt and Iron. A lot of it was about how the Western Han was governing the salt and iron industries, which by that point were largely privatized, and they were trying to figure out how much should we extract from this private enterprise. If we have any listeners who are experts in the Western Han, please correct me on this, but it’s a cool parallel with AI today, which in China is also largely privatized, and the state is trying to figure out how much to extract from it.

Jordan Schneider: And other empires brought it in house, right? There was a salt monopoly for hundreds of years at some point.

Let It Rip

Jordan Schneider: So speaking of nationalizing salt in 300 AD, maybe we should talk about our joint piece we put out a few days after Kimi launched, where we did a little scenario analysis of three potential paths for where Chinese governance of AI could go. Each of you guys wrote one of them. Let’s start with Let It Rip.

Aqib Zakaria: Yeah, I wrote that one. The Let It Rip scenario is just the status quo — the Chinese Mythos model will come out just as Kimi K3 did, just as GLM 5.2 did. It just drops. And maybe hopefully there’s a one-week delay saying we’ll release the weights later. But there doesn’t seem to be any system within the government for catching, this model’s different, this model actually has the capability to do some real damage. And then once it’s out, it’s out.

And maybe after that, this will be the wake-up call they need to scramble and have some policy to change how they review AI models. But an open-weight Kimi K4 or a GLM 6 can do some major damage out in the open.

Jordan Schneider: The question I have now is, what if it just doesn’t? What if it’s just another day in the life? We’ve had so much — can we call it hyperventilating? Can we call it alarm raising — from cybersecurity experts. There is a world in which they are wrong.

Let’s make the case. What was Kevin Xu saying? You need serious access to compute in data centers. Maybe at the cloud layer and the neoclouds, they just decide that they’re not going to do business with cyber hackers or Boko Haram or whoever else. Maybe things get patched fast enough that it’s just kind of a blip and we get to the other side, and instead of taking two or three years as some people are proposing, it takes a month. Then all of a sudden our models are great because there’s just a lot more money being put into these models to clean up systems than there is to attack systems. Other arguments folks have in favor of we just kind of move on?

Lily Ottinger: From the Chinese government perspective, they already live in a world of cyber exploit abundance. They’ve had a system for systematically acquiring zero-day vulnerabilities for a long time and have decided that the risks of using them outweigh the benefits in a lot of scenarios, but that’s because China’s a state actor with constraints placed on them by the multilateral system, the trade system. They don’t want the backlash.

And they feel like they have domestic insurgencies under control. So to them it’s not a threat domestically. I don’t think we are necessarily prepared for a world in which state actors — Claude Code made a lot of problems look software-shaped. I don’t think we’re prepared for a world in which a lot of problems look like they have solutions that are cyberattack-shaped. Unclear what steps are required to execute on any of those things. But the fact that the exploits exist, I feel like they’ll at least be sellable in a way that makes a lot of money for rogue actors in a way that we’re perhaps not prepared for. It’ll certainly be a nice way for the North Korean government to make money, but China’s not worried about them either.

You could be right that there’s no problem. I think the Chinese government for now certainly thinks it’s not a risk.

Aqib Zakaria: I think it could be such good PR for the Chinese government if it comes out without a major incident. Because people are upset over the safeguards on Fable 5 and all these lobotomized models, and now China’s the one giving it to us straight from the source. And then you have this scenario where the leading AI models are Chinese, and also the lower-cost ones are also Chinese, which could do major benefits for their adoption levels.

The Kevin Xu argument is that for Kimi, which is over two trillion parameters, you need like a dozen GPUs. You need a couple million on you, which is a lot, but certainly not insurmountable if you’re any organization or any evil rich guy, I suppose. It’s just a question of what actually gets done with it, and how much damage the CCP is willing to take. And if there is a major cyberattack with this Chinese Mythos model, how bad of PR is it for the government? And what can they even do after the fact to correct it?

Irene Zhang: It’s worth mentioning very quickly that the Chinese government does get hacked all the time. Their systems are not nearly as foolproof as they would like, as any government would like. And beyond state actors, beyond the United States and China’s various rivals around the world, there are also just the scammers in Myanmar who want Chinese PII. So there are lots of ways to make life annoying if bad actors get their hands on really strong cybersecurity tools, even without it challenging major geopolitical assumptions.

Jordan Schneider: PII is one thing, the NSA hacking you is also one thing, which is just — yes, governments hack each other, but they don’t put all the secrets online. We haven’t really had WikiLeaks for China. That I think would be a very interesting test of the case. Every government has awkward things that they don’t want the world to see, and I’m sure some of those are digitized in China. And there are definitely enough people with a few million dollars who feel some type of way about the Chinese government.

And then okay, so if you’re Xi, then what? That’s a really interesting thing because then you start to get into these state stability, global reputation type of questions. And that’s not sci-fi bioweapon stuff, that’s pretty run-of-the-mill. We’ve seen this happen in a lot of authoritarian societies around the world where awkward news comes out and folks freak out. You didn’t even need cyberattacks in 2012 with the New York Times reporting around elite politicians and money for that to really change how folks in the government thought about how they wanted to let a domestic population have access to global news. So that seems baked in almost at some point.

Lily Ottinger: I feel like the PR crisis comes when a Chinese-made nuclear reactor in a Belt and Road country gets shut down via cyberattack. That’s my prediction.

Irene Zhang: Where are the nuclear reactors?

Jordan Schneider: You heard it here first.

Lily Ottinger: Pakistan has some. There are plans to build them in several other countries in the Global South, but the Pakistan ones I think are the ones that have been subjected to the most fanfare. China makes a lot of nuclear reactors for countries that would not have them otherwise.

Irene Zhang: I’m reading this for another article that will hopefully come out soon — an October 2025 guiding document by the Cyberspace Administration of China. It’s basically a guiding document for various government agencies to adopt AI safely. There’s one section that’s actually pretty relevant to what we’re talking about, about how China deals with state secrets, where it emphasizes — and this is my own translation — to strengthen a bottom-line level of thinking and strictly make sure that anything that is a secret does not go on the internet, and when you go on the internet, there are no secrets.

And it doesn’t say that no state secrets or sensitive information should go into AI models. It says no state secrets, work secrets, or sensitive information should go into AI models that are not secrets-informed. Which implies the existence of secrets-informed AI models.

The Black Box

Jordan Schneider: All right, maybe we should jump from Let It Rip all the way to Black Box. Irene, what was your vision of the future?

Irene Zhang: My vision, which I think everyone ranked as the least possible, but which I will try to defend here, is a world where the Chinese state realizes that one of the leading labs in China develops a model that heavily surpasses previous models in cybersecurity capabilities and capabilities in general, and decides that rather than letting the company release it openly, they will get first dibs on it for government use and then only allow the company to release a lobotomized version, because it is too concerned about safety. And in that process partially nationalizes this lab and tries to coerce it into developing products that are more tailored towards government use and potentially offensive use.

The reason I think that’s a genuinely possible scenario — I know the biggest counterargument is that this is commercially bad for Chinese AI’s influence in the world, to have its top capability models not be available to any national audience to diffuse and compete with the US. But China’s advantage in some ways is that it has a lot of labs. Let’s say Kimi gets there first. Let’s say K4 is somehow extraordinary and the government takes that in house and forces Kimi to not release openly. Zhipu AI will get there someday, Qwen will get there, ByteDance — we don’t hear enough about ByteDance, but they will get there someday.

In a sense, the advantage that Beijing will get from a strategy like that maybe outweighs the short-term PR game and commercial game. My scenario makes sense in a world where the Chinese government is a lot more quote-unquote safety-pilled than people assume, and also sees genuinely useful things they can do with AI for national security and its strategic vision in the short term. Enough that it thinks making that sort of very explicit move to acquire top intelligence is really useful.

Jordan Schneider: Two points in your favor, Irene, from the Xi speech. Perhaps the most interesting paragraph was him saying we must attach great importance to the various endogenous and derivative risks brought about by AI and work to build systems of laws and regulations, technical monitoring, early risk warning, and emergency response. We must reinforce the safety baseline, prevent misuse and abuse, and ensure that AI always remains under human control. So clearly some safety emphasis there.

Plus, I gave the top three Chinese models and the top two American ones all three of our scenarios, and the Chinese models believed in the black box. Even though me, Lily, and Aqib all gave it a fifteen percent chance, DeepSeek was up to thirty percent, Kimi was up to forty-five percent. So the Chinese models think they’re going to get nationalized sooner rather than later.

Irene Zhang: Maybe they know something we don’t.

Aqib Zakaria: The only reason I find scenario three unlikely isn’t that I don’t think there’s this big mega-brain tradeoff happening. For scenario three to happen, the Chinese government has to know that this model is going to be really good. And the only way that happens, because I don’t think the model regulators check for cyber capabilities or bio capabilities, is Kimi has to be like, guys, K4 is different. And then the Chinese government has to believe them. And then they have to take that step of nationalization, which is a complete U-turn from the speech and the messaging they’re currently giving.

Irene Zhang: Well, Matt made that point when he came on ChinaTalk, about how there’s the version of the registry that’s on paper and then there’s what happens in real life. Which makes a lot of sense to me, given what I’ve heard as well. So on paper, they don’t test for bio, they don’t test for all of this before they can release a model, but in reality the CAC is probably in pretty close contact with these labs, and I really doubt that they have no idea of a model’s capabilities before they go out.

Lily Ottinger: I kind of feel like, from talking to AI researchers in China, their understanding is that they’re going to make the models exactly as safe as is required by law and they’re not going to think about it beyond that, because they’re compute constrained and they just want to do their national obligation. And that’s it. Maybe they talk a lot about safety, but it’s unclear how much that’s a thing that they personally really believe in, versus a thing that they’re aware the global community believes in, and thus want to be interested in safety in order to have their place in that global community. If it’s left up to the labs to flag cyber risks, we are in a bad spot. Let It Rip is probably pretty likely in that situation.

China’s Glasswing

Jordan Schneider: Lily, let’s go to you then. How does China’s Glasswing happen?

Lily Ottinger: Project Glasswing would happen in a world where the capabilities are flagged in advance by somebody. Someone from the Chinese government gives a speech in advance that says we need to move the testing capabilities up, we need to evaluate first, release later, we need to fold in relevant stakeholders from private industry and government ministries before releasing a model with huge step-changing capabilities. And then the labs take that as a signal that now it’s their responsibility to figure out whether their model is going to be Mythos level before they release it. Perhaps the fact that Kimi is waiting to release the open weights of K3 is an indication that the labs are thinking about this more.

I think this is pretty likely because there’s a playbook for it already, in the form of Project Glasswing. It enables the model to eventually still go out in a way that is safe for everybody and has the cyber capabilities lobotomized, but enables Chinese developers to still work with the best tools, which is something the Chinese government seems to really care about because they believe AI is a huge unlock as far as productivity is concerned. So I think there are forces in the Chinese government that will work against having the models be totally black-boxed and government-only and never released.

I think it would cause a huge amount of infighting. The labs would not necessarily enjoy being the national champion of a Glasswing-style project, because it would mean handing over their proprietary technology to their competitors, all of which are building their own foundation models too. When Mythos was handed over to Microsoft, it wasn’t such a big deal, because Microsoft’s not a serious competitor in the AI race. But that’s not true in China, because everybody and their mother has their own foundation model, including delivery apps like Meituan. And they all hate each other and have longstanding rivalries and they all poach employees from each other. There’s so much drama.

Irene Zhang: I tried to think of examples previously of Chinese big tech cooperating with each other on stuff. And I really couldn’t think of any.

Lily Ottinger: And I think the government likes it better that way. They prefer that the companies don’t necessarily get along and do these big cooperation initiatives that stretch across conglomerates.

Irene Zhang: Yeah, that’s called price fixing.

Jordan Schneider: The funniest thing that’s happened in the Chinese AI race over the past few weeks is Qihoo 360, which is a security company, trying to convince everyone that their model or their harness is Mythos class and they’ve done their own Glasswing. Their CEO is like, we are only giving out our special sauce AI product to a handful of providers, and they took a photo or whatever. It’s like, okay, guys. Sure. Good luck with that.

There’s a piece of this, Lily, which you may be missing, where there’s some honor in being the one that has to kick off the Glasswing.

Irene Zhang: My question is, I don’t know the extent to which that kind of honor really buys them stuff in China right now. For example, Anthropic’s marketing makes sense even to average American society, because Anthropic and its peers have been messaging for so long to the general public in the US about AI’s risks, and everyone and their mother has heard that AI could kill us all and stuff. That doesn’t happen in China yet. The catastrophic risk narrative has not taken hold. People’s negative feelings about AI are mostly about job loss and the economic side.

So if some company were to come out with that heroic messaging around AI safety, taking that baton of honor, I don’t know how much that resonates with the actual public. The previous examples of Chinese big tech doing honor-based stuff that I can think of are like donating masks during COVID. Is that the kind of stuff that genuinely resonates with average people?

Lily Ottinger: Seems like a thing Alibaba would do.

Aqib Zakaria: One thing that’s interesting is that whatever path happens is really dependent on which lab gets there first. For this scenario, if Alibaba or Qwen was the one to reach there, they’re like, I am not sharing this with the other conglomerates. But maybe if it’s a smaller lab, they’re more willing to do it, because there’s not as much of that history of politics there.

Jordan Schneider: Yeah, but they’re all kind of tucked in, right? I don’t know if they’re fully subsumed and have vetoes, but if all your compute comes from one place, which is the case with Kimi and Ali, then that’ll get messy.

Irene Zhang: I guess we should not project our understanding of Silicon Valley’s internal dynamics onto the Chinese tech industry at all. There are obviously ways in which both systems learn from each other and mirror each other, but in terms of that relational thing, it’s really different.

Lily Ottinger: Everybody loves a good taboo though. The model’s secret, nobody can have it yet. And then when it finally comes out in a usable way, even if it’s lobotomized, people will be like, well, what is this whole thing about? Everybody loves something you’re not supposed to have. In that sense, maybe that’s more salient messaging than honor is.

Irene Zhang: But then you can’t do any more messaging beyond that, because so much of AI marketing now is about actually sharing your capabilities and getting people to use it, whether it’s consumer or enterprise or all of the above, and especially coding. This is actually a pretty word-of-mouth industry on the consumer side. People just be telling each other what models they like. So if you can’t use it, there’s a limit to how much you can sell the thing.

Lily Ottinger: Well, I just mean that the taboo is good marketing for Chinese Fable, not for Chinese Mythos.

Aqib Zakaria: The counterargument here is that all of these companies want to ship their model as fast as possible, so you can churn the money out of it before you have to ship the next model. So in this case, the Glasswing might actually be kind of problematic, because every week that you’re not shipping is a week of money you’re losing from an already presumably cash-constrained company.

Irene Zhang: Yeah, Lily sort of makes that point in the article, of the thing just getting bogged down by bureaucracy.

Lily Ottinger: The question of when do we actually release it after a Glasswing-style project is going to be really difficult, because nobody’s going to want to take responsibility for what happens after it gets released. So evaluate first, release later turns into endless evaluation. And it takes a long time to patch cyber capabilities. In the pre-Mythos era, there were bug finders that flagged serious bugs in Microsoft’s code that remained in the spreadsheet unpatched for eight years before they were exploited. So I think it’s hard to say.

Does the Gap Even Matter?

Aqib Zakaria: How much do y’all think the path we go down depends on where America is at with model capabilities? If we’re already two, three generations after Mythos-level capabilities, do you think that will make China more likely to rush it, be like, we’re too far behind, we gotta let it rip? Or do you think they’re completely taking things at their own pace?

Lily Ottinger: I do think it makes Let It Rip more likely. Because they think, well, this isn’t us opening the floodgates, the floodgates are open. So if Fable came out and there was no problem, then this isn’t really all that different and we should just rush to catch up.

Irene Zhang: The question for my scenario is the extent to which the American frontier being much further ahead shapes what’s downstream of that. I can see a world in which the American frontier gets stronger in various other capabilities, in AI for science and all of that, but it doesn’t necessarily lead to the American economy taking off significantly more than it does now. It doesn’t lead to the US suddenly possessing much more incredible weapons and whatnot. In which case China might be less worried, because then the lesson it gets is, diffusion’s actually pretty hard, we have time to think of alternative pathways. But if that frontier means that the US economy or its position somehow genuinely gains significant leads, then probably Let It Rip is more likely.

Jordan Schneider: Let’s go one level up. What is the three-to-five-year national power impact of the world in which you can’t make a lot of money just developing models, because Chinese open-weight stuff keeps coming out and it’s pretty much as good? Versus the world in which you have this nine-to-twelve-month lag in capabilities, and the only way you can pay for intelligence is to buy tokens from OpenAI or Anthropic or maybe DeepMind if they get it together.

In the world where you can’t make this big surplus from developing models, development probably goes a little slower. But you still have the ratio of ten times more compute being outside of China as inside China. And the point of AI is not to make labs money. It’s to grow productivity. The world ex-China is still getting ten times more agents running around in data centers, inventing things, making your systems more productive than the PRC is, and you just aren’t paying this tax, which is maybe one third of it going back into R&D to make it better.

Aqib, why don’t you take the other side? The gap stays wide. You still have closed models that are a year ahead, or China starts closing off and the labs try to skim their own percentages from selling API access.

Aqib Zakaria: A lot of this is a function of how quickly model research becomes autonomous. In that sense, model development could be happening constantly, like every week, if the AI is just updating itself. And then all the money is being churned off of people wanting access to the newest model. But that also depends on how much the newest model really boosts productivity. How much can productivity really increase from Mythos 72 to Mythos 73?

Lily Ottinger: I agree with that. The capabilities are already there for a lot of tasks. People are happy to use Chinese open models for their OpenClaw bots and personal productivity management stuff. And enterprises are happy to pay.

Jordan Schneider: But we want science, right? We don’t have science yet, really. Maybe the first science we get is AI research, because that’s all software. Who knows when that unlocks.

Aqib Zakaria: One thing a lot of people have been talking about since K3 and GLM 5.2 is how much does compute matter? The entire theory about export controls and people talking about policy is these models should get worse and worse, or the gap should increase, because China has less compute. But now the gap isn’t really increasing that heavily, even if it’s still pre-Mythos, post-Mythos. Is this fully a problem of them remotely accessing compute in Malaysia? Or is it cope? What’s going on?

Jordan Schneider: It’s a business model thing as much as it is a training run thing. Okay, congratulations, Kimi, you dropped this really exciting model. Now you’re worth what, 75 billion dollars? That’s like series C for Anthropic. We’re up to a trillion at this point. And they’re worth a trillion dollars not just because they have the best models in the world, but also because they have enough compute to deploy them and make money off having the best models in the world. The problem about open sourcing your stuff is you can’t make money off of it.

So yes, maybe you can keep fast following, but at some point, unless the Chinese government steps in or there’s enough investor excitement, it will be difficult to compete with the most successful companies in the world. You can compete in fast following, potentially, if all you need is to be able to distill and you have some good researchers and you come up with creative stuff and you have enough compute to train models. But if you don’t have enough compute to actually do the deployment and make money on the front end and diffuse the technology into your economy, then you’re still at a pretty significant structural disadvantage when it comes to reaping the benefits of what this tech will deliver.

Irene Zhang: And just for context, Kimi has not been aggressive in doing anything commercial either. Zhipu is really the leader in on-prem deployments, some 2G stuff. Minimax and the various big tech models have their own commercialization 2C venues. Kimi’s kind of a purist. I’m sure they have a coding agent, I’m sure they have API revenue and whatnot, but they’ve not been interested in this stuff.

Jordan Schneider: Okay, so they’re worth thirty billion dollars. The question is how do you get to five hundred billion? You kind of need to sell something in order to keep growing, I think. Or is the Chinese government going to give one of these labs hundreds of billions of dollars? And even if you have hundreds of billions of dollars, can you buy compute with that? I guess if America keeps letting them, then the answer’s yes. But they’ll just be in Malaysia.

Aqib Zakaria: But don’t you think you get to a point where the difference between the American leading model and the second place Kimi or GLM is not enough for me to care? So I go with the cheaper open-weights one. In which case OpenAI and Anthropic can’t make money, and neither can the open source model. And you need to give life support to that layer, because otherwise everything downstream of it doesn’t exist.

Jordan Schneider: I think they’ll figure it out.

Lily Ottinger: It’s going to be good for us to figure this out.

Aqib Zakaria: Yeah, I don’t want to talk about giving welfare to the labs. That’s not very good political messaging.

Irene Zhang: Zhipu’s doing fine.

Jordan Schneider: They’re down forty-five percent on the month! Kimi was not good news for them.

Irene Zhang: I look forward to reading Kimi’s IPO documents sometime later this year.

Lily Ottinger: So I went to Kimi. I went and met all the people in Beijing. A lot of Carnegie Mellon people in their thirties who just seem really relaxed and excited to be there. All the conference rooms are named after American rock bands. The culture seemed very healthy and steady and sustainable compared to the culture I saw at Minimax, for example. So I think they’re in it for the long haul. I don’t think they’re that concerned about making money. They think it’s going to be fine.

Jordan Schneider: This is just OpenAI 2021, right? It’s all fun and games until you really need to have a business. And they may just not need to have a business. It might be enough utility for the state, and it’s an exciting thing to invest in. They can keep it going, they can keep up their compute access by having exciting, flashy, fast-following stuff. But I’m a little skeptical.

DeepSeek Raises

Jordan Schneider: I mean, DeepSeek decided they had to make money, right? Irene, why don’t you give us a DeepSeek update? Kimi and Moonshot have been getting all the headlines, but this whole other lab that kicked this all off still does exist.

Irene Zhang: The DeepSeek drama has just been that it has been raising funds. To jog everyone’s memory a bit, DeepSeek belongs to this hedge fund called High-Flyer, which is one of the most successful hedge funds in China, although recently it’s been going down a bit because there’s been a bit of a slump over AI. That means it’s basically had unlimited funding for a long time, unlike most of the other labs. It’s kind of cash rich, which is really strong. And Liang Wenfeng owns both DeepSeek and High-Flyer the hedge fund. So he basically takes the hedge fund money and puts it into AGI.

But over the past year or so there’s been a lot of rumors that DeepSeek is looking into raising external money. We got some confirmation of that over the past few months. That has been really interesting because DeepSeek has, according to some reports, turned down a bunch of investments from places like Alibaba, apparently because Liang Wenfeng didn’t want to dilute his control over the organization too much.

What we know now is that it has raised about fifty billion RMB, which is like seven and a half billion USD. Investors include Tencent and a few other organizations, and also CATL, the battery maker, and some state funds, including the National Artificial Intelligence Industry Investment Fund. But Liang Wenfeng has also put 20 billion RMB of his own money into this fundraising round, to lock in his own control even more.

Which makes me question the thesis of how much it has accepted its fate as a national champion. If it truly accepted that, there are tons of state funds in China excited to jump into this whole thing. They could have taken that money, but they didn’t. Not to the same extent.

Jordan Schneider: I guess the question in China is, once you’ve hit a certain threshold of billionaire, you’re not allowed to have fun. Really. In the way that you are in the West with yachts and everything. So maybe past three billion or so — what is he going to buy, Manchester United, or take the Brooklyn Nets off Joe Tsai?

Teaching the Country to Use It

Aqib Zakaria: It’s interesting how Anthropic is doing Claude for nonprofits and now OpenAI is trying to do ChatGPT for small businesses. I wonder if you’re going to see analogous programming by Chinese labs for diffusion.

Jordan Schneider: Look, I think that kind of program is actually a good idea. Yes, people are going to need human handholding to teach them how to learn this. If this is how diffusion’s going to happen, then great. It’s like the Ag Department of the 1910s telling people how to get rid of boll weevil. You gotta get boots on the ground to get people interested and excited about this stuff.

Irene Zhang: There’s a way in which this kind of stuff works better in China, where there’s real manufacturing. I can see it being really sexy marketing if Kimi starts being like, we’ll help all these textile manufacturers in Guangdong upgrade their stuff, make their processing more efficient so that they can treat their workers better. That kind of thing doesn’t exist in the US. Maybe agriculture.

Lily Ottinger: Seems like Zhipu will do that first if it happens.

Aqib Zakaria: I’ve been thinking about this a lot — going to the people, where it’s like, we have to teach the people who don’t know the revolutionary potential. And then they’re like, get out of here. It’s not going to work.

Irene Zhang: I find it funny that Aqib is pro 上山下乡, going up to the mountains and down to the countryside.

Aqib Zakaria: We gotta do it right. You can’t have a dude from San Francisco show up to Louisville.

Jordan Schneider: That’s the thing. If your AI isn’t good enough to teach a small business owner how to use it, and you need a human layer, I feel like you gotta put some more work into your model.

Aqib Zakaria: You have to convince them that they have to like the AI. It’s a people-shaped problem.

Lily Ottinger: This is a people-shaped problem for sure. Send . She’s the perfect person for the task. Send her to Louisville. Send her to the mines.

You can read this article in microsite form thanks to Kimi K3 here.

Soon, a Chinese lab will develop an AI model with capabilities on par with Anthropic’s Claude Mythos. Mythos’ potential to wreak cyber havoc drove the White House to create what is a “de facto involuntary licensing/preapproval regime for frontier models.” Will the CCP similarly panic when the day comes in China?

In June, Z.ai cofounder Jie Tang told Elon on X that China will have a model on par with Mythos before the end of the year. Even if he’s off by a few months (Kimi K3’s release shows we’re certainly getting close) the clock is ticking for regulators in Beijing to respond.

Reuters, on July 7, reported that Chinese regulators were considering “potentially restricting overseas access to China’s most advanced AI models.” And Xi himself has raised technical AI safety to the highest policy agenda. In his speech at WAIC yesterday, while touting the benefits of openness he also said:

We must attach great importance to the various endogenous and derivative risks brought about by AI, and work to build systems of laws and regulations, technical monitoring, risk early warning, and emergency response. We must reinforce the safety baseline, prevent misuse and abuse, and ensure that artificial intelligence always remains under human control.

We wanted to share our team’s own theories about what may come next. In our eyes, the story of Chinese Mythos may pan out in three main ways:

• a continuation of laissez faire “let it rip”: the model releases with minimal intervention;

• ‘Glasswing with Chinese Characteristics’: access is limited first to ministries, then key trusted companies, and only later in a limited fashion to consumers

• The black-box, where we will only ever hear rumors of the power of China’s Mythos, contained solely within labs and government ministries

We tell the story of these paths as hypothetical scenarios because all of these outcomes seem plausible. We hope that a creative approach inspires flexible thinking on preparedness: USG, AI policy people, safety advocates, and researchers should be prepared for whatever may come. We create the fictional Z.ai GLM 6 to be the hypothetical Mythos-level model that will rock the world. For all we know, the real Chinese Mythos may come from Kimi, DeepSeek, Qwen, or from any other lab.

Scenario 1: GLM 6 Drops Into the Wild

One day, GLM 6 dropped just as all other models, Chinese or American, had dropped before Mythos: a hype tweet. Only a few hours later, just like how its other models have been released, Zhipu makes GLM 6 open-sourced, with its weights available freely online.

Fictitious graph showing GLM 6’s capabilities. GPT 5.6 performs worse than Mythos on exploiting cyber vulnerabilities. In the pre-Mythos era, OpenAI would have never released this model and waited for the line to beat Mythos. Now, sandbagging is the meta!

After a few days, two phenomena occur simultaneously. First, cyber professionals tinker around and discover GLM 6 can find zero-day vulnerabilities the same way Mythos could. They scramble to inform labs and policymakers, both in the US and in China. Second, enterprises that just want to use good, cheap models are realizing GLM 6 is a saving grace. Before, some companies like Uber Eats, Siemens, and Shopify had already begun to integrate Chinese models in their workflow. But now, with GLM 6, there is no trade-off between cheaper tokens and amazing performance. On the graph of enterprise usage of Chinese models, the slope has a dramatic uptick.

The two phenomena pose real questions to Beijing. A model that is democratizing cyber offense (something the Chinese government believed it had an edge with on everyone sans the NSA) is proliferating worldwide. Cyberattacks seem like they’re on the verge of inflicting disproportionate financial damage on China’s leading companies (since America and friends have already been hardened by Mythos access). But Beijing also sees the economic upside. Zhipu’s market cap approaches $500bn, instantly becoming one of China’s largest companies. GLM 6 is China’s chance to beat America in leading model adoption, crater the western labs’ businesses, and lead global enterprises and government systems to turn to China as the go-to technology provider. Plus, Xi just gave a big speech about how important openness is!

Besides, struggle is part of the process. The Party has been willing to accept sacrifices during COVID opening up, some bank failures, and tech crackdowns; this round of cyberattacks could just be the culling before China emerges stronger than ever. On the verge of securing a new commanding height, bureaucrats are torn over how to achieve AI ambitions and security at home.

The Trump administration is also torn. Perhaps the uptick is an anomaly, and no policy action is required. USG rails against China for distilling and then releasing such a model, declares that government contractors can’t build with Chinese open models, but doesn’t outlaw models as it would be too silly to try to ban a file that can just be torrented. When Beijing doesn’t send any signals on cracking down on GLM 6, Washington decides to let it rip as well, screaming at its leading labs to drop something open. If there’s going to be a powerful open source model out making waves, it should probably be American.

Z.ai workers, already in a sleepless daze from the eighteen-hour days up to launch, don’t know what to do. The celebrations from the launch party have subsided into stress. They are already talking with the relevant authorities, but what to do? Because they open sourced the model, like they did with 5.2, there is no putting the genie back in the bottle. The model, with well over a trillion parameters, can’t just be weaponized by anyone with too much free time and a Mac Mini. But there are plenty of actors out there — terrorists, companies with the wrong interests, criminal syndicates — who could get their hands on a good number of racks to turn the model against China. All Z.ai can do is hold their breath and wait for the fallout.

The Party decided that GLM 6 wasn’t dangerous enough to strangle, but the model did serve as a wake-up call. GLM 6 is now the ceiling for unsupervised drops, and every model after gets scrutinized closely before release. Without much fuss, China implements a stricter licensing regime that doesn’t just check for politically unacceptable output, but also cyber capabilities.

Scenario 2: Glasswing with Chinese Characteristics

After seeing Mythos, Beijing realized that emergent capabilities could inflict real damage in the wrong hands. The best means of mitigating the negative effects of future models seemed to be a Project Glasswing-type undertaking. In August 2026, Ding Xuexiang gives a speech implying that breakout domestic AI models should be shared with important government ministries and corporate giants like Baidu, Alibaba, and Tencent before deployment.

Tang Jie and team had already deeply internalized the risks of misuse for Mythos-level models. After Ding’s speech, Z.ai postpones the launch party for GLM 6 to further evaluate the model’s capabilities internally, at which point their researchers realize what their creation is capable of. Determined to not be disappeared, company leadership pitches Project IceFeather (冰羽项目) to the Cyberspace Administration of China. The CAC, of course, takes the lead on announcing the initiative in October, but Z.ai is credited as the “core private sector partner” of the project.

Government ministries and key SOEs like State Grid get access first. They use their exclusive access window to check for vulnerabilities in their own systems while the MSS trawls for exploits that can be used offensively overseas.

The Party’s involvement in Project IceFeather means the state is also a co-owner of the model’s risks. The dominant strategy for bureaucrats instructed to “Evaluate first, deploy later” turns out to be endless evaluation. The downside risk of letting anyone with enough compute get to punch a hole in China’s critical cyber architecture seems much scarier to the system than marginally more productivity growth and global discourse power as the open standard for AI. At first, nobody quite decides that the model should be released, as nobody wants to be responsible for what would happen post-release. But after seeing American models continue to advance and the gap between U.S. and China widen, the trigger gets pulled. Strong safeguards are put on the model, perhaps to the point of excess, and the regulators hope that domestic infrastructure has been sufficiently strengthened by Project IceFeather. Hesitant to hand over their golden goose, Z.ai leadership offers API access instead of the full model weights. A lobotomized GLM 6-Open gets released, while the real breakthroughs Z.ai plans to monetize by selling tokens.

Visualization of the fictitious GLM 6’s public safeguards. Adapted from Anthropic.

Despite having the best Chinese model, Z.ai stock price doesn’t moonshot as feet-dragging over Project IceFeather spooks investors who worry that the government won’t allow them to monetize the model beyond state customers, much less into key markets overseas. While the MSS has the time of its life hacking away at everyone who didn’t get access to Mythos, domestic coders still mostly use transfer stations to access frontier capabilities.

After two months, regulators expand the circle of trusted users for GLM 6, to include provincial governments as well as trusted local tech leaders like Alibaba, Xiaomi and Huawei. It takes another few months for Chinese firms broadly to get access, by which point other domestic labs have developed roughly similar models. Thanks to pressure from the Trump administration, OpenAI releases a cyber-lobotomized open model on par with GLM 6, which finally convinces CAC that Zhipu should be able to allow the world to access its model.

Scenario 3: Zhipu Emerges as the Certified National Champion as a Covetous Beijing Keeps Frontier AI In-House

Li Qiang, China’s Number Two, wasn’t kidding. What Beijing really cares about is really “upholding the bottom line of AI security,” not the platitudes in Xi’s speech at WAIC about “openness, win-win” and “mutual learning between civilizations” that the credulous foreigners ate up. Behind the scenes, Beijing has maintained deep insights into exactly how capable China’s top AI models were, and was preparing for the moment when it knew it would have to step in.

In September 2026, the Cybersecurity Administration of China publicly updates its AI model registry regulations. Whereas previously, the rules only governed new models and algorithms with the ability to “shape public opinion,” it now requires companies to report any major update, self-assess for potentially catastrophic cyber risks, and give the government exclusive access indefinitely. Remembering the fates of DiDi and Ant Group, companies all know better than to not comply.

Now the state now gets first dibs on every new model release and update. An internally assembled testing capacity through the MSS and PLA Cyberspace Force meant that when GLM 6 proved to be genuinely exceptional, Zhongnanhai was the first to know. Having recognized that this model was a potential ‘assassin’s mace’ Xi had directed the state to forge a decade ago, there was no way the security state would allow its capability to be open sourced. Despite Foreign Ministry protestations that releasing this model globally would improve China’s standing in the world, and MIIT officials salivating at the domestic productivity boost the model could provide, the security state was not having it. Open soucing this model would forfeit a cyber window of opportunity and allow Xinjiang separatists and Falun Gong activists with some Azure credits the same hacking capabilities as the domestic security apparatus.

Two days after GLM 6 was sent over to the state, Tang Jie and co. were summoned to CAC. Over long hours and litres of tea, a deal was struck. The “real” GLM 6 would become available to government entities: specifically, it would be deployed on-premise into core governmental functions that benefit from such capabilities — including defense. Z.ai would release to the public a lobotomized version of the model, thoroughly tested by both the company and the government to ensure its cyber capabilities are satisfactorily weakened. The lab gets a contract with lucrative term and the state agrees to help ‘persuade’ other chinese model makers to sign away their compute both in the mainland and at southeast asian neoclouds. Instead of cyber capabilities getting democratized with the likes of Xinjiang separatists and the Falun Gong

In a world where cyber capabilities have the potential to become democratized, Beijing has made sure that the tools in its arsenal are of a different tier.

Parts of the meeting leak to Reuters, but no matter. At first, no one was surprised that the Chinese government was getting better AI than the masses. Z.ai was now the proper national champion, but there was no actual proof that they made something transformative.

Then the attacks started. Beijing knew that Americans, armed with Mythos-level defensive capabilities for months already, were on the lookout for any sign of trouble, and did not want to re-inflame the bilateral relationship at a time when it could not expect to extract meaningful concessions. Instead, it focuses its new cyber bazooka on less-prepared countries: wavering capitals in the Global South, less developed neighbors, and eventually Europe. It was easy enough to hack into Paraguay’s citizen records; by early 2027, Asunción was frightened into dropping recognition of Taiwan. European technology companies in the semiconductor supply chain, began reporting cyber attacks at a much-increased scale. Brussels bureaucrats began fretting about blackmail of uncertain origin.

Other Chinese labs, having signed over a good amount of their compute to Zhipu, slowly progress on their own capabilities but still get prevented from releasing their weights by the CAC. By this point, collaboration between the government and Z.ai has reached new heights. Building on GLM 6, the lab is now developing proprietary models tailored to government wants and new-generation cyber weapons. On the outside, Z.ai still looks like a normal lab, continuing to release consumer-grade models, open-source research, and AI products. The to-G side is heavily insulated; researchers must be thoroughly vetted before being deployed to serve the government, and are required to surrender their passports. The CAC occasionally poaches star researchers from other labs under the guise of “national development consultation.”

Which Scenario is Most Likely?

How likely are each of these scenarios? Here at ChinaTalk, we’ve been having lively internal debates over what might come next. Here are our thoughts (as well as Kimi v3’s!):

Jordan: Chinese glasswing (Scenario 2) most likely (65%) then ‘let it rip’ (scenario 1) (20%) then nationalization (scenario 3) (15%). On 3, despite Xi’s desire for SOEs to perform, I think he has learned his lesson through the tech crackdown that the state can’t really be relied on to produce the sort of technological breakthroughs required to keep pace with the AI frontier (he learned why the unicorns went away). Some muddle through of tighter regulation, golden shares and minority state investment, and a Glasswing-style release as opposed to an indefinite lockdown on access and total nationalization seems like the more likely equilibrium over the next two years.

I also don’t see the labs keeping their best capabilities open forever. Chinese model makers have already begun to go closed source to make money and as Reuters reported there are influential strains within the system skeptical of allowing foreigners’ access to China’s best models. The benefits to the Chinese government of open sourcing powerful AI are diffuse (global PR, undermining the western AI ecosystem), while the harms are concentrated in a paranoid security state that is not interested in democratizing access to hacking. Plus, Xi really likes assassins maces, and a world-conquering cyberweapon is not the sort of thing that you give up for good global PR. Xi’s WAIC speech to me felt like it could just be a barbarian handling exercise akin to his 2017 World Economic Forum speech. In that address, he promised to serve as a counterweight to Trump tariffs but immediately started on a global campaign of economic coercion once security issues towards Korea, Japan, and Australia cropped up. When things get serious once China has an AI model that can truly compete with or even exceed the cyber capabilities the US has access to, it’s unlikely he will really want to hand over that capability to Huggingface.

Lastly, Zhipu, Deepseek, and Moonshot, the leading Chinese labs now that everyone good at Qwen quit, do not own complements like Nvidia which give them economic incentives to commoditize themselves. Their CEOs are genuinely ideologically committed to open source, but they do need to keep money coming in to do novel research and train models. Today I don’t see a coordinated national strategy to pop the west’s AI bubble by driving down the global cost of intelligence. But that may be emerging, evidenced by Xi’s speech and the Deepseek investment.

I do not believe that today there is a national strategy to pop the west’s AI bubble, but it is possible that the government decides it’s strategically important enough to foot the bill for its top firms to drive down the global cost of intelligence. It would not take nearly as much money as Xi has already spent in more hardware-oriented industrial policy efforts around cars, solar, and semiconductors.

Aqib: I see Scenario 1 and 2 equally likely (45%), Scenario 3 least likely (10%). My stock in Scenario 1 went up after seeing Kimi K3 deployed without any fuss, coupled with the recent statements from WAIC.

Scenario 1, where a frontier lab “lets it rip” and deals with consequences later, cannot be counted out. In a vacuum, Z.ai released GLM 5.2 after Mythos seemingly without any fanfare or close scrutiny; the Chinese government also does not seem to monitor model releases for cyber or bio capabilities. Z.ai’s founder recently wrote a memo reiterating support for open source, and Xi prefers to have a “water-flow” mindset toward AI. Most recently, Kimi K3, a powerful open-source model, was released seemingly without close scrutiny as well. This language and action suggests a willingness to keep the models open and public.

However, recently government interest in monitoring models more closely has risen. Reuters reported that Beijing is mulling future powerful models being “barred from public release or restricted to domestic use.” Model developers, who keep regular informal communication with regulators, would likely communicate such capabilities to the regulator to receive informal approval before release, lest the developer receive a crackdown for releasing the model too hastily.

The scenario has some uncertainty, and reality could branch out at different points. GLM 6 could be released as a closed-weight model, just like Qwen; that would mean the model could be put back in the bottle if the capabilities were too dangerous. The Party may also decide that, after release, it is worth rescinding the model because the risks are just too large. Chinese SOEs have notoriously poor cybersecurity, so they may be the first ones hit. We don’t know what happens afterwards in this world; the hammer could come down hard on the entire AI industry, or it could just be the beginning of a stricter licensing regime.

Lily thinks China’s Mythos won’t be locked behind closed doors indefinitely, but a Project Glasswing-style rollout with Chinese bureaucratic characteristics is possible. The Party’s concerns about AI security go hand in hand with support for AI-enabled One Person Companies and the belief that AI will unlock new productivity increases. Certainly, there are forces in Zhongnanhai that want China’s developers to be working with the best possible tools. But the fact that everyone with the technical expertise necessary to brief the government also has a horse in the AI race could be a powerfully perverse incentive.

Overall, Scenario 2 most likely (45%), then Scenario 1 (35%), then Scenario 3 (20%). I think Scenario 3 is the least likely because the private sector has already produced such high quality models on their own, so perhaps the party would be skeptical that state involvement would produce something more capable.

Irene is uncertain about exactly how CAC regulators will patch holes in the existing registry system, but thinks a secret model is more likely that people assume.

There are recent signals in favor of the theory that CCP leadership is genuinely concerned about AI safety. Xi Jinping will make his WAIC debut in Shanghai this month, where, according to state media, he will “systemically elaborate on China’s policies, position, visions and propositions on AI development and governance.” AI guardrails came up during the Xi-Trump summit in May. And here’s how Zhou Haibing, vice minister of the National Development and Reform Commission, paraphrased Xi Jinping’s message on AI safety during a press conference in June 2026, just a week after Fable’s first release (emphases are Irene’s):

[China will] enact the responsibilities of a major country, manage and control risks, strengthen prevention efforts, explore cooperative opportunities in AI oversight, and jointly defend against AI safety risks. [中国将... 履行大国责任，管控风险、强化预防，探索开展人工智能监管合作，共同防范人工智能安全风险。]

The emergence of keen prevention rhetoric means Beijing expects regulatory agencies to stamp down on risks before they proliferate. Barring truly necessary structural changes, a “let it rip” approach is just not in the PRC’s deeply conservative policy psyche. Publicized regulations will change with the times, but Irene expects much of the negotiations on frontier capabilities inside China to be kept under the hood going forward, no matter how leaky the system can be at times.

The PRC has an extensive history of secret strategic projects. If Beijing truly believes that AI will shape national power — much like weapons of mass destruction did during the Cold War — then it probably is already tempted by its old playbook. “Two Bombs, One Satellite” was highly classified in its initial years, and some of the researchers involved were not permitted to publicize their experiences until decades later. Mao’s “Third Front” (which Irene has written about before) involved building entire secret cities in remote western regions to hide defense-industrial capabilities from prying American and Soviet eyes.

In many senses, hiding a frontier model is easier than hiding literal mines and factories. Despite significant research efforts (including by our own Aqib and Nick), no one has high-confidence estimates on how much AI compute China has. Satellite data on data centers do not correspond neatly onto actual compute amounts, because no one knows what chips are inside or what utilization rates look like. Amid an entire gold rush of companies training and using AI systems, it’d be quite easy to hide a proprietary system in plain sight.

But one challenge might be personnel. There are only so many star AI researchers in China. It’s a tightly-knit circle, and many have overseas experience. Beijing is probably wary of bringing some of these folks onto the high side to build black-box systems, but arguably has no alternative. This is why Irene imagined a scenario where the CAC tries to semi-absorb part of Z.ai rather than build a whole secret lab from scratch.

Overall, I see Scenario 3 most likely (45%, even though we won’t find out), then Scenario 1 (30%), with Scenario 2 least likely (25%).

Below the paywall, results from when Jordan burned $500 in tokens over the weekend to ask the leading Chinese and American models how they think Beijing will react to China’s Mythos moment. He got some pretty surprising results!

Read more

Claude Mythos has completely scrambled American AI policy. Trump 2.0 — an administration whose AI-dominance rhetoric manifested in a let-it-rip regulatory approach — is shook. To quote Dean Ball, “we currently have a de facto involuntary licensing pre-approval regime for frontier models” from the administration that promised us the opposite of that.

Will Beijing also freak out once it faces the prospect of wide domestic access to a model that can hack everything? We may not have long to find out. A Zhipu AI co-founder called his shot, telling Elon on X that China will have a model on par with Mythos before the end of the year. Even if he’s off by a few months, the American think tank IAPS puts the date at February 2027. A Chinese Mythos-level model is inevitable and only months off. Politics is now shaping not just the hardware ecosystem but the model layer.

What does this turning point mean for the domestic Chinese AI ecosystem, Chinese regulation, open source in China and around the world, US-China governance efforts, and broader global AI safety? To discuss, we have back on the show of Interconnected and Matt Sheehan of Carnegie.

In this episode, we discuss…

• Whether Beijing will panic the way Washington did when China gets its own Mythos-level model, and why that moment may be only months away.

• Why China’s existing AI bureaucracy may be better prepared for frontier models than America’s scramble to figure out who’s in charge.

• What a “Project Glasswing with Chinese characteristics” would look like and who would make the guest list.

• Whether increasingly capable AI will kill open source in China, or strengthen Beijing’s argument that it’s the more open AI superpower.

• Why today’s AI labs may have more influence over government policy than they’ll ever have again.

Listen now on your favorite podcast app.

Will Beijing Panic Like Washington Did?

Jordan Schneider: Matt — as the preeminent Western follower of all things China, AI, and regulation — my sense from reading everything you’ve written is that what Beijing has done so far hasn’t really changed how these businesses operate. The first real global impact we’ve seen on a model provider was Trump taking Claude Fable offline. Given the arc of Chinese AI regulation, we’ve seen ripples, and we’ve seen regulators tip their hand about what worries them. But we’re not talking Jack Ma, we’re not talking Didi, we’re not talking edtech — it’s still been live and let live. Is that stable? Once they face down a model that can hack through big chunks of the cyber world, what will Chinese regulators end up doing?

Matt Sheehan: First, a couple of qualifiers on the idea that they’ve been super chill so far. There was a period when regulation really did impact the companies — 2023, the year after ChatGPT — when companies were holding up model releases for three months, six months, until regulators put together a generative AI regulation and worked through the initial approvals, which were essentially licenses at the time and have since become more like registration. That period genuinely slowed things down.

Since then, it has transformed into a burdensome compliance regime. It’s not a hard no on model releases, but there’s a battery of testing you have to run and submit. It’s a burden on the companies — a lot of employee hours. The overwhelming majority of that time is spent on content: political content, social issues — basically making sure their broad internet censorship, information management, and content moderation regime works for AI.

The question is what happens when a new threat equivalent to those content threats comes online in the CCP’s eyes — AI-enabled cyberattacks. They have built the infrastructure to handle this about as smoothly as a country can. We basically watched from April through July as the Trump administration stumbled and fumbled through responses, with the recurring question being: who’s in charge here? Does this go to the Center for AI Standards and Innovation? Through Commerce? Through the NSA? In China, bureaucratic shuffling and competition are inevitable and permanent, but they do have a channel this goes through. Before you release a model, you submit a battery of tests to the CAC. You don’t have to submit the model weights — at least historically you haven’t — but you need to give them API access and accounts to test with.

What they’re going to do first is bolt cyber-related tests onto that pre-deployment testing regime, which has traditionally been about content, plus a bunch of filtering requirements. We saw some of this yesterday: the CAC has been running a rectification campaign on AI applications, and one action to come out of it was that MiniMax had to add safeguards against generating malicious code. They’ll tack all those things on.

Beyond that, it depends on the size of the threat. If they jump suddenly from where they are now to Mythos-level capability, they’ll essentially run a government Project Glasswing, then a company Project Glasswing, then a wider release. In the limited documents released so far — not by the government, but by influential thinkers over there — the first worry is the safety of government systems and critical infrastructure: how do we create a fast track to patch these things first? Once we really jump to that level of capability, they’ll have to prioritize government and public infrastructure, then do a slow rollout toward eventually full open-weight releases.

Jordan Schneider: Kevin, seem reasonable?

Kevin Xu: All of that sounds reasonable, but it’s worth laying out what Matt implied: there is far more prior art in China’s governing system for regulating AI. The deep synthesis rules on content date back to 2021-2022 — pre-ChatGPT. And there’s everything else Matt mentioned about controlling AI in ways our government hasn’t thought about at all. China has mandatory restrictions: you now have to label AI-generated content on all the social media apps. Everybody on Douyin — not TikTok, but Douyin — or Xiaohongshu knows whether what they’re seeing is AI-generated. The effects of generative AI regulation are already being felt in China.

Does the cybersecurity vulnerability question upend any of that as dramatically as it has in the United States? My base case is no. As capable as a Mythos-level model is at discovering and executing cyberattacks, it is equally capable of discovering and patching the same vulnerabilities on the defense side. Even before Mythos-level models, AI was pretty good at finding bugs — that was never something AI coding tools were bad at. What the Mythos-Fable level adds is long-horizon task capability: a model can now autonomously execute a project of 10, 20, 50 steps over hours and weeks, which is what you actually need to exploit a cybersecurity vulnerability. It’s not that you look at one codebase, find a bug, inject something, and suddenly have an attack. It’s cross-matching multiple databases, files, and accounts, spotting weird patterns where you can put one thing in and get to another. These are complicated tasks that Mythos-level models are apparently capable of — and the same capability can patch the same vulnerabilities.

What will likely happen is a delayed release of some sort, where the patching happens first for the major banks, the regulated industries, the public infrastructure, before the model gets released into the wild for normal commercial and civilian purposes. It will be more measured and will live within the existing infrastructure of the Chinese governance system — in ways that remain TBD here. Who do you call to get permission to release a model inside the United States government? Commerce? The institute we stood up? Treasury? I honestly have no idea. It’s probably everybody, plus the White House Chief of Staff’s office. The rules are confusing.

In a way, this may be the one case where China does not look to the United States for tips and tricks on how to regulate an emerging trend. China has looked to the West to learn how to regulate antitrust, how to regulate its stock market, and much else. But at this frontier, China may actually be in de novo territory.

Jordan Schneider: Well, Matt just said they’re going to do Glasswing, which is not de novo territory — that’s following our lead.

Kevin Xu: I don’t know if that will happen or not. But that’s how we always feel — that they’ll do something we did a month ago.

Project Glasswing with Chinese Characteristics

Jordan Schneider: Let’s talk through both scenarios — Glasswing and not-Glasswing. Glasswing looks like this: Zhipu or Ali or whoever sends the CAC an email saying, look, we’ve got this new model, it can hack the shit out of everything, we probably shouldn’t just put it up on OpenRouter. Then maybe just the government gets access, maybe government plus SOEs, maybe government plus SOEs plus your twenty favorite companies in China — for three months. At some point you get this awkward moment where the company says “we can release it now” and the regulators say “no way, we don’t want a giant cyberattack on our own systems from open models.” That gets resolved eventually, and two, three, four months later, the model gets released. We have a mental model for that.

The crazier scenario is that some model developer just really wants to compete and releases. Right now, given the current regulatory dynamics and the amount of competition in the ecosystem, you could just press publish on a closed or even open model that represents a dramatic step up in capability. Let’s live in that world, Matt. GLM 5.7 drops and it’s just way better. What happens next?

Matt Sheehan: Let me say why I don’t think that will happen. By the letter of the departmental regulation, you submit these filings and complete certain tests — and those tests don’t include cyber. Once you’ve submitted and the CAC accepts, that’s your license and you can go. But it’s not just a formal process. The CAC is constantly touching base with all these companies. At least as of a year or two ago, they were holding weekly meetings where they called in the government relations and policy people from the companies just to talk about what’s new in AI policy. Someone told me — this was around the time of the Hollywood writers’ strikes — “yeah, the CAC called us all in the other week and wanted to talk about labor and the writers’ strikes and what we think about this in relation to China.” Just to chat.

Jordan Schneider: Your friendly neighborhood regulator.

Matt Sheehan: That’s been the CAC. It has attempted a rebrand from bad cop — “you do something and then I punish you” — to “let me help you comply.” Part of that is these regular touchpoints with companies. In that environment, it would be very surprising for a Chinese company to decide not to give them a heads-up and just drop a model — technically we’ve met all our requirements, so let’s release it. That’d be very, very bold and self-destructive behavior by a Chinese tech company.

Kevin Xu: That doesn’t even happen today, with much less capable models. No Chinese company has ever skipped the heads-up or the registration process before releasing to the public.

Jordan Schneider: There’s still a Didi, though. They were told not to list, and they listed.

Kevin Xu: That’s a totally separate topic. Let’s not conflate a New York listing with a model release.

Matt Sheehan: To bolster the case: one thing we have some insight into — though I want more clarity from the companies — is that they don’t have to run the full battery of tests for every iterative model release. They don’t do a new registration for each release. If you go into the model filing system and look up DeepSeek or Ali, the number of registered models is far smaller than the total number they’re hosting on GitHub or Hugging Face.

Jordan Schneider: I think we’re missing the theme a little. We’ve crossed the threshold from regulators keeping an eye on things to regulators understanding there are national-level consequences if they get this wrong — not a few people harming themselves, but hospital systems going offline, financial institutions shaken to their core. We’ve seen Bessent freak out and call the heads of the banks saying, you really have to pay attention to this.

Both of you seem to hold a status-quo base case. But there was a status-quo base case in the US too, and it was proven radically wrong about how the US government would relate to these models. Let’s live in the world where the Chinese government decides it really needs to change its approach, rather than just adding a few more cyber questions to its registration form. We’ve been having this debate for a long time — is China AGI-pilled? Is the Politburo AGI-pilled? They’re not. People can’t be, until they are — until it stares you in the face how powerful and impactful these things are.

Matt Sheehan: Right — and to be clear, the Glasswing scenario would not be the base case. Going government-only first, then SOEs, then big companies, then wide release would be very different from the current status quo. It would represent a significant change.

Kevin Xu: The status quo is also different in the United States versus China in terms of the level of existing AI regulation — our base cases don’t start from the same point. One thing to keep in mind about how Glasswing actually came about, and why I think a repeat most likely won’t happen in China: Glasswing was started by a private company. It was not initiated by a government agency telling an AI company that made a really powerful model, “hey, maybe we should have a closed pre-review release process, give a heads-up to the banks and critical infrastructure, then release to the public.” Anthropic chose the members of Glasswing — not the United States government. That’s why the US government has been scrambling to respond. Our posture had been to not overregulate and, frankly, to commit to not having a licensing regime for AI models — and now we have a weird de facto one. We were scrambling precisely because a private company outside the government’s control started it.

If a similar pre-release situation were to happen in China, it would most definitely be started by a government agency, and the access list would be pre-approved by that agency before anything else happened. That is, by nature, a very different response from how Glasswing manifested.

Jordan Schneider: So who’s in and who’s out? Our Chinese Glasswing — who gets it first, who gets it second? What companies are on the bubble?

Kevin Xu: All the government agencies will get it. All the state grid companies. The petroleum companies — the strategic resources. Frankly, all the companies that are non-tech and have really bad, probably buggy infrastructure. That’s the thing we should worry about here too: why isn’t the state of Tennessee getting Glasswing access? That is not a very good IT system; it’s going to get hacked very, very easily. I don’t think the list will look that different.

Jordan Schneider: Maybe the more interesting question is the private companies — who gets it and who doesn’t.

Matt Sheehan: It’d be especially interesting depending on who makes the model. Say it comes out of Zhipu — then it’s government ministries, central SOEs, then maybe provincial governments and local SOEs. But what happens to Ali if Zhipu creates it first is a big question. That might be a case where the government steps in and says, you have to work with them. Ali probably wouldn’t be far behind in either case, but a good amount of government and wider useful infrastructure runs on Alibaba Cloud, and the idea that it would be left insecure seems unlikely.

A more extreme reaction than a government Glasswing would be the security organs stepping in and saying: thank you for bringing us to this point, Zhipu; thank you, DeepSeek — but we’ll take it from here. This is now a national priority that has to be brought fully in-house. That would be a really bad idea, because those organs don’t know how to run these companies or train models at the same level. But it’s not entirely out of the question.

A lot depends on how high the government’s confidence is relative to how much it needs the companies. When the government feels confident and doesn’t need the companies, it tends to crack down hard — that was essentially the tech-crackdown era: “we’re doing great in China, our capabilities are great,” so they felt free to impose heavy burdens. Post-ChatGPT, the government felt much more insecure: maybe there was a cost to that whole crackdown; we have to give these companies more leash. One of the interesting changes since DeepSeek is the government feeling good again — okay, maybe China really is back at or near the frontier, and we can get more hands-on with our companies. The most extreme version of that is some form of government takeover. I don’t think it’s likely, and I don’t think it would be smart. But if they get a little too high on their own supply of confidence, that’s the extreme scenario.

Whither Open Source?

Jordan Schneider: Whither open source in this world, Kevin?

Kevin Xu: It’s interesting that within the last two or three days, the founders of both MiniMax and Zhipu made high-profile pronouncements about open source. The MiniMax CEO committed 1% of MiniMax’s market cap to supporting open source — what I’ve heard is they’ll start a foundation with that seed money to promote open source in perpetuity. Zhipu’s founder wrote an internal memo that was “leaked” onto Xiaohongshu and translated, which I read yesterday, committing to open source as part of their rather grandiose mission of bringing AI to all of humanity, et cetera.

Obviously, people have read the inklings that China may stop releasing open-source models because of capability increases. So we have mixed signals. I tend to stick closer to the signal from the people actually making the models — who are all relatively well connected, or at least have a channel of communication with the relevant regulators. Their commitment to releasing open-source models as part of their strategy is not, right now, running counter to anything the government may do. We’ll see how long that lasts, but the commitment from some of these labs remains very strong, because frankly open source has gotten them this far — the progress of GLM, of DeepSeek, is non-trivial. And there’s also a population of Chinese model makers that is increasingly closed-source for their own commercial needs. Open source was never a national strategy dictated from the top — that’s a common misconception. There’s a distribution of opinion on whether open source is good for my company, for humanity, and so on.

Jordan Schneider: For context on the article Kevin referenced: last week Reuters reported that, “Chinese authorities have held meetings with top tech firms over recent months about potentially restricting overseas access to China’s most advanced AI models, including those not yet released. Participants discussed limits on the most advanced models, both closed and open, and officials talked about making the leak or theft of proprietary AI an offense under the national security law. The scope is still being discussed.”

One way to interpret the moves by the Zhipu and MiniMax CEOs is that they’re trying to hold their ground — “wait, guys, this has worked really well for us; we think it’s a positive thing.” It’s hard to read too much into an article like this, but at minimum there are currents within the Chinese government concerned about wide access to increasingly powerful models — something we hadn’t really seen signaled before.

Kevin Xu: That’s right. To add to the Reuters article, which got a lot of play: on the same day, or close to it, a Chinese official gave a speech at a UN AI gathering in Geneva committing to open source as a very important part of China’s overall posture toward global AI — an official voice. So you have leaked sources going to Reuters, and an official giving a public speech committing to open source. We’ll probably get the best signal in a few weeks when Shanghai hosts the World AI Conference, where Xi Jinping himself is slated to speak. My anticipation is that he’ll say at least something about open source, and that will be the clearest country-level signal of where open source stands in the grand scheme of things.

Matt Sheehan: A couple of comments on the article. I’m pretty sure from memory that the Reuters piece says the regulators in the room were the Ministry of Commerce and the NDRC — Commerce in relation to export controls. That’s very interesting, because Commerce is not the main AI regulator; that’s the CAC. Commerce isn’t testing models, and it has been pretty out of the loop on AI other than interfering in the Manus deal and other export-control matters.

One version of this is that they create the ability to control the export of something like model weights. This is what they did with recommendation algorithms ahead of the TikTok deal. When it became clear the US was preparing to ban TikTok and force a sale, China — I’m pretty sure via the Ministry of Commerce — wrote algorithms into its controlled items list. I may be getting details wrong, but essentially they gave themselves a veto over a future TikTok deal, which they held onto until the end. That might be one reading of these events: Commerce isn’t the security people, but they could be creating a veto over future Manus-style deals, or trying to get in the game on model releases — though I don’t think they’d necessarily succeed.

I’m personally skeptical they’ll impose wide controls on open source — not just because of how beneficial it’s been to the companies, but because it’s been a very beneficial global political narrative for the CCP. Their Global AI Governance Initiative in 2023 came out right before the first AI Safety Summit at Bletchley, and from the beginning they’ve said China stands for global access and against using technological hegemony to exclude countries — positioning themselves as the opposite of the US: the US wants to control this and prevent you from getting access; we are the country of openness. In an extreme situation where China’s core interests are threatened, they won’t opt for nice diplomatic language over core interests. But they feel the benefit of that narrative. If they were to slow down or restrict open source, it would be done in a way that isn’t a total shutoff and lets them maintain the line: we’re still the open country, we’re still sharing — we’re just doing some internal security checks first, and then we’ll share.

Jordan Schneider: Here’s where I think you’re both wrong. In the world where AI keeps getting more powerful on some kind of exponential, I find it hard to imagine a scenario where the Chinese government is comfortable enabling random criminal enterprises — or anyone in the world who wants to hack China — to use Chinese technology for that purpose. If we’re 6, 12, 18 months out and China’s Project Glasswing hasn’t hardened the entire society against hacks that its own companies’ open-weight models can execute, then the rhetoric and the companies’ interests start to lose out. The security state says: we’re having 100x scams, domestic disturbances caused by people using AI — or by AI itself, because if it’s open-weight you can set things off and run.

I hear your arguments about diplomatic and economic costs. But take seriously the scenario where this gets really powerful really quickly, with emergent capabilities such that after a two-month access window you still can’t feel confident you’ve hardened yourself. Do those delays turn into six months? How do you even make the release argument if new dangers keep surfacing? As the technology gets more powerful, the Zhongnanhai-brain argument to just slow this thing down gets stronger and stronger.

Kevin Xu: That conflates the model-capability trajectory — which I agree is probably on a much steeper curve than any of us can visualize or feel day to day — with whether open versus closed release even matters in that scenario. If the model is capable enough to do all the damage you’re describing, and you’ve run a pre-approval Glasswing-type process to harden your infrastructure, then releasing it as a commercial product behind APIs versus open weights on Hugging Face is actually not that different. These models are so massive that we spin ourselves into a rabbit hole imagining that as soon as the weights are out there, some random person can hack the state in no time. The amount of infrastructure and cloud computing required to even properly deploy these things — let alone do something vaguely useful or vaguely adversarial — is ginormous. The way you’d regulate that scenario is to talk to the finite number of companies around the world with compute at that scale and make sure it doesn’t happen on their infrastructure. That has nothing to do with whether the model is open or closed.

We’re conflating model progress with whether open versus closed matters, when the open side actually accrues far more benefit and is, by definition, safer and more secure — which doesn’t get talked about enough. None of us wants to live next to a park with no lights over it. As soon as the lights go up over the dark park, property values rise. The same goes for open models: having the weights published and accessible is a much safer way to release powerful capabilities than closed ones — or at least about as safe as the closed equivalent.

Matt Sheehan: Whether open weights matter is a deep and complicated question, and I’m not so sanguine that they don’t matter at all. There are a lot of well-resourced actors who could still put these things to use. Maybe the key question is whether we’re going through a transition period — from the status quo, through a stretch of global cyber chaos, to a future state of more secure systems all around. I’m not a cyber person, but I have friends deep in that world and I put this question to them regularly. It’s not settled, but the consensus seems to be gathering around: yes, it will be chaos for a while, but you can effectively eliminate bugs in existing software, patch everything we know about, and create provably secure software — and it’s not the case that the next-level Mythos will endlessly discover new vulnerabilities.

It’s possible China sees this transition period and concludes: we can get through it with a certain lag time, and then hold onto all the other benefits of open source. You’re right to point out that the CCP is a security-and-control-first institution, and they are not going to be sanguine about being hacked by pick your country. But they might see a path where you harden the systems within China and then accrue the benefits of open-weight releases.

Jordan Schneider: Once someone in the Chinese state writes a memo saying that X separatist group from abroad that wants to take down the party did this thing, uncovered and leaked these documents, and by the way they weren’t part of Project Glasswing, they just had access to some GLM model — that is going to be a dramatic shock to the system.

To Kevin’s point, the open-versus-closed question comes down to this: if you’re providing API access, can you actually lobotomize a model so it just doesn’t do cyber — so you get all the productivity uplift but you can’t hack with it, can’t make bioweapons, can’t run a scam farm, pick whichever capability scares you? With an open-weight model, that’s presumably harder, and it’s harder to trace who did it. We can already smuggle a lot of NVIDIA chips. I could presumably put twenty of them together and start calling grandmas.

Kevin Xu: You could. I just think it’s more than twenty. It’s a non-trivial amount of hardware — someone would need to be hell-bent on this, a very powerful non-state actor. That’s something we might want to talk about at some point: how does that scenario get regulated? It’s one thing China, the US, and probably every country in the world actually agree on — how do we prevent non-state actors of malicious intent from, first, having the resources to run their own on-prem cluster of 10,000-plus NVIDIA chips, before we even get to regulating their access to models, closed or open?

Jordan Schneider: I don’t think you need to be building your server in the jungle. Standard Chartered helped Iran break sanctions for years.

Kevin Xu: And this all assumes commercial APIs run by OpenAI or Anthropic are somehow bulletproof — which they’re not.

Matt Sheehan: They’re not bulletproof, but they are constantly monitored. OpenAI recently reported that a handful of people in China had used their models to try to generate data-center backlash material, that kind of thing. That’s a lot more scrutiny than you can have over open weights once they’re released. This really isn’t my area, and I’m not a hardline anti-open-source person, but I don’t have blind faith that it works out. There are well-resourced non-state criminal organizations — or North Korea, or someone like that — with the technical capability and money to use open-weight models in ways they find beneficial. They can make more money than they lose.

Jordan Schneider: We just had Boko Haram asking ChatGPT how to overrun a military garrison. Some rebels in the jungle in Nigeria figured out Pliny the Elder jailbreaks. This stuff isn’t rocket science.

Kevin Xu: And ChatGPT didn’t stop itself from being able to help them out, did it?

What Can Washington and Beijing Actually Agree On?

Jordan Schneider: Right — maybe they had to try DeepSeek too; let’s not put it all on OpenAI. But to Kevin’s point, Matt Sheehan: what can the US and China actually agree on? What’s there even to discuss?

Matt Sheehan: When Bessent made his first announcement — which China didn’t co-sign or confirm at the time — he said something like: we’re setting up a dialogue with China on protocols to make sure these models don’t fall into the wrong hands or aren’t used by non-state actors. That was clearly the US focus from the beginning, and when you think about what the two sides can agree on, it should be in that non-state-actor space. China should operate on the assumption that no matter what the US says or what regulations we have, the NSA will be using the most advanced models it can get its hands on to hack China. We should assume the MSS is doing exactly the same in reverse. There should be an acceptance between the two sides that they’ll use these things against each other. Where they share an interest is controlling the non-state-actor front.

But then you run into open source, where the incentives diverge: the big US companies don’t have much impetus to play in the open-source space; the big Chinese companies do. I’ve heard people float ideas about building safeguards not just into API access but even certain safeguards and filtering into open-weight releases. I don’t understand how far that can go or how much it buys you. We may end up in a situation where both sides face the same threat in theory, but company incentives and government positioning prevent anything meaningful. The best they’ll manage is on domestic testing.

There are productive things to talk about — a low ceiling, but a lot to do within it. Generally speaking, Chinese frontier AI companies do not test for extreme risks at anywhere near the level of US companies, nor with the same sophistication. It happens in small pockets, but it’s not comparable. And there’s no Chinese AISI equivalent that has been developing frontier-AI-risk expertise in that way. It would be possible, and very good, to find a way for the US to safely share information on how to run these tests and test for these capabilities — in a way that isn’t capability-enhancing but lets governments, regulators, or companies sniff out dangerous capabilities earlier. We’re not signing a treaty; we’re not agreeing on shared thresholds. It’s sharing knowledge about things neither side wants happening within its own borders.

Kevin Xu: To the extent we want to bend over backwards to find overlap in the Venn diagram of bilateral dialogue, there’s a lot to talk about on actual societal impact. Take the simple question of labor displacement: what is the legal justification for laying people off because of AI productivity gains? Matt, you’ve written about this. There’s a lot of value in simply sharing what each side is trying to do, because the productivity disruption is coming for companies in both countries as we speak, and the companies have very similar motivations. Both governments are wrestling with how to manage the transition for everyday people — not just the scary cybersecurity and safety-testing scenarios. There’s a lot to talk about without ever having to agree on anything: what are you doing, what am I doing, oh that sounds interesting — and then we go off to our separate spaces and figure it out for our own people.

Matt Sheehan: I’ve organized dialogues about exactly that, and it’s genuinely interesting and useful to compare notes. I’m a little more skeptical of the value of governments doing it. Mutual awareness is good, but the odds that we really learn from each other’s responses to labor displacement are significantly lower. The more something is embedded in political culture and social institutions, the less transferable the lessons. Though even as I say this, I’m mentally walking it back — China has picked up a lot of regulatory mechanisms we built and adapted them to its own situation. On labor specifically, though, we have pretty different attitudes toward labor and legal protections.

Kevin Xu: What about something less cultural and more technical — robotaxi urban regulation, for example? Beyond the streets looking different and the cities being laid out differently, how differently would Wuhan do it than Phoenix, and vice versa?

Jordan Schneider: That kind of thing doesn’t feel like it matters much — good meetings, some mileage.

Kevin Xu: It may not matter for a podcast, but it matters for people on the ground. Better mental energy spent on that than on separatist groups triggering the highest wire of the CCP security apparatus, personally.

Matt Sheehan: There is a track record of this kind of exchange on environmental issues. It just feels very ’90s and 2000s.

Jordan Schneider: It’s a nice thought that we could find common ground to collaborate on, both staring down our AI-driven futures.

Share

Matt Sheehan: It partly depends on how AGI-pilled you are. One thought I had when you said the CCP won’t be AGI-pilled until it is: as big and scary as the cyber stuff is, I don’t think it necessarily leads them to being AGI-pilled in the fast-takeoff, recursive-self-improvement sense. Maybe there’s some relationship to RSI, but there’s a big difference between losing control over superadvanced AI systems and “it’s really good at reading software and finding bugs.” I’ll be curious to see how much more AGI/ASI-type language we see because cyber has come to the forefront.

Jordan Schneider: Xi has said he’s worried about the risks of technological loss of control, so it’s bubbling up, even if we’re not entirely there.

Let me qualify my “it doesn’t matter”: it is net positive for both countries and the world for the US and China to have a productive discussion about the second-order impacts of this technology. And it’s interesting that regulations are already developing in parallel. China has had a whole anthropomorphic-AI debate about how AI shouldn’t pretend it’s human. In the US, we’re having a big markup on a children’s AI safety bill — and minor mode is something that’s been rolled out or discussed in China over the past few weeks. The systems are moving in parallel. I don’t necessarily think Senator Hawley was inspired by Beijing — though they kind of do inspire each other. When you see a US article about how China banned video games for kids, you see a lot of reaction here.

Matt Sheehan: That’s a good example. And — pop champagne — I think the companion regulation went into effect two hours ago, so AI companions are now regulated in China as of noon our time. In that case, the Chinese regulation was pretty directly inspired by state-level bills in California and New York. They use different mechanisms: we largely use private right of action — enabling lawsuits against companies for harms — whereas they go through central approval and much more hands-on technical mandates. But there are definitely provisions in their rules drawn from our state bills, and I wouldn’t be surprised to see ideas start flowing in reverse.

This stuff matters beyond its real impact on people’s lives. The public’s attitude toward AI — its receptiveness, its opposition — will be rooted far more in what they saw happen with companions and their kids rather than in their timeline for RSI. It’s indirect and hard to predict which direction each of these things pushes, but there is mutual learning going on, and it could continue. The transnational, large-scale, potentially catastrophic risks sit in a different bucket, with a different set of mechanisms and different limits on what we can and cannot talk about.

Kevin Xu: To build on that: taking care of the “smaller” issues of AI’s everyday impact will directly shape what being AGI-pilled even means for a government. We can be as AGI-pilled as we want here, but if we don’t actually deliver benefits — or at least prevent harms — for everyday people, we’re going to get data center moratoriums up the wazoo, if we don’t already. It doesn’t matter how AGI-pilled DC is if rural communities won’t let you build a data center. What’s the point of being that committed to the mission? Similarly in China: taking care of the everyday impact is the best way for the government to figure out whether it wants to be AGI-pilled and what that even means as national strategy. Those two things are far more causally linked than we give them credit for.

Jordan Schneider: My hypothesis two or three years ago was that AI safety would eventually be framed in China as a Western plot to slow them down relative to the US. That has definitely not happened. For their own reasons — their own incentive structures — the Chinese political class and regulatory apparatus have internalized that there are downsides the government needs to ameliorate as this technology rolls out. Contrast that with US-China military dialogue over the past 25 years, where America has been genuinely freaked out that if a crisis happens and we can’t call Beijing to say “sorry, this was an accident,” World War III starts — a thesis the Chinese government has simply not bought into, to the frustration of successive American presidents. The fact that AI safety hasn’t gone down that path gives me broader optimism that there are things these two countries can talk about productively, or at least work on in parallel.

Matt Sheehan: I had the same worry three or four years ago — AI safety framed as a Western plot to hold China down, a clear mirror of the attitude they took toward climate for a while. Climate is instructive. The first time I got to China, in 2008, I picked up an English-language state media paper and the top headline was about climate change as a Western plot to hold China down. It made a big impression — one of the first things I read over there. But eventually their own scientists came around to the fact that it was very real. It combined with concerns about air pollution, and — still relatively technocratic over there — they chose not just to work on emissions-reduction technology but to see a huge opportunity and claim the entire green energy space.

You could argue AI safety doesn’t have five industries built into it that China can dominate and leverage. But fundamentally, they still listen to science over there. How they hear it will always be colored by their incentives — maybe they’re concerned, but they’re more concerned about competition with the US. In most cases so far, though, reason or science has ultimately guided the direction of things. People could point to COVID as a counterexample, and that was a very intense couple of years. We’ll see how it plays out in AI.

Kevin Xu: I tend to be more worried about how much, here in the US, we’re going down paths that look a lot like what China is doing — in ways that don’t reflect our own first principles. The whole Fable saga — how GPT-5.6 was pre-released to a select handful of customers approved by the US government — rings a little too close to home to the tech crackdown that happened in China a few years ago. I spend far more time worrying about that scenario: what is our system doing, in a way that benefits our own companies and makes sense on its own terms, as opposed to being purely reactive? There’s some usefulness in comparing — what’s the other side doing, are they going to do this or that — but there’s enough first-principles thinking for us to do things the right way for ourselves. Instead, we seem to be skipping what China does well and copying what they haven’t done well, really quickly. Golden shares and a whole list of things.

Jordan Schneider: Maybe that’s why, at the beginning of our conversation, Glasswing didn’t seem so foreign to Chinese observers — this close-hold, state-directed, centrally planned thing.

Kevin Xu: Government involvement in private company matters is a very normal thing there.

State Capitalism with American Characteristics

Jordan Schneider: And the direct CEO negotiation of who does what, when. Let’s close on Sam Altman offering up 5% of the company in return…

Kevin Xu: For — what? Protection?

Jordan Schneider: I guess. Donald Jr. clearly missed the boat on that Anthropic IPO; probably regretting it now. Where do we go? Should we tax all these guys, Kevin? What’s the public policy answer?

Kevin Xu: Backing up, I don’t think US government involvement in private industry is inherently bad. If anything, the government could and should spend time and money identifying areas where it can inject itself strategically to boost domestic capacity that’s been falling badly behind — rare earth mining capacity, certain AI strategic initiatives, and so on. As for the form — equity shares owned by the government — there are good and bad arguments. Should the government instead stimulate demand, say for quantum technology, playing customer number one to boost the industry? That’s an interesting model of government involvement that doesn’t feel like a carbon copy of some other system.

But if OpenAI does this, all the other frontier labs will have to offer the government a similar equity stake, and it becomes a really bizarre new chapter in America’s so-called capitalism — American state capitalism with its own characteristics. We should think really hard about whether this actually enriches the American people. I don’t see how it works unless the shares the government owns get divided by 350 million and we all get a deposit in our E-Trade accounts. Does it just go into the Treasury, where we have no say or knowledge of how that money — which will appreciate — works out for the benefit of the broader public?

Matt Sheehan: This is one place where the US-China comparison is actually instructive. China’s first draft of a response to labor disruption is basically: you cannot fire people just because they’ve been replaced by AI — that’s not a valid reason to end a contract under labor law. It’s instructive that in the US we revert to a clean economic-modeling solution: we’ll take shares of companies as something like a tax, centralize, redistribute — nice and efficient; it would be so inefficient for us to get involved in the mechanics of companies. China is much more willing to grind through and say, no, we’re not going to let you lay these people off and then tax you and redistribute through some system. You’re making more money because you implemented AI — find something for them to do. It’s a muddle-through approach, versus a UBI system that looks very clean on paper. Whether people derive meaning from their lives under a UBI economy, versus being kept in make-work jobs — a cradle-to-grave iron rice bowl — I’m really not the person to answer. But it’s a manifestation of each political culture: we opt for clean redistribution; they lean into firm-driven, state-powered retention of people inside companies.

Jordan Schneider: The 5% is a purely Trump phenomenon. This is what he likes — signing deals. Then maybe he won’t do actual tax policy.

Matt Sheehan: It’s also the Bernie-Trump policy pipeline. Bernie says we need 50%; Trump says, yeah, sounds good.

Jordan Schneider: It’s also weird that every smart economist thinking about this now works for one of these two companies. That’s another strange public-policy thing. I’ve read the Anthropic version of these proposals and the OpenAI version, and in their defense they’re smarter than “we’ll hand over 5% of the company and move on.” But this is classic Econ 101, Tyler Cowen-reading-Marginal Revolution-for twenty years stuff: you tax what you want to disincentivize. Figure out a way to do that where people still have jobs, rather than just walking away from it.

Matt Sheehan: The policy brain drain is worth pausing on. Five years ago, the big conversation was academic AI/ML people being sucked into private companies, eventually hollowing out academia. We’re now seeing the policy version of that. I’m not claiming that my colleagues or us geniuses at think tanks are so priceless we must be protected — but companies are offering three or four times the salary of non-governmental organizations. If we don’t want all the most sophisticated policy-oriented people working for the companies building the technology and profiting from it, we need to do some work to keep people in independent organizations.

Jordan Schneider: Donors have got to step up. The AI talent war applies to the Mass Ave think tanks too.

Kevin Xu: Don’t let ChinaTalk get acquired.

Jordan Schneider: Hey, everyone’s got a price. Kevin, I’m sorry I didn’t invest in Nebeus.

Kevin Xu: A microcosm of the larger problem, aren’t we? These companies will end up writing the policies because they have on their payroll all the people who’ve been thinking about the policies this whole time. You can’t have the companies’ incentives be anything less than one of the top three concerns shaping how those policies get written — maybe not the first, but certainly top three.

Jordan Schneider: I stole this from some Ezra Klein articles: there’s a reverse horseshoe of where companies can really have impact. On one end, tiny problems nobody cares about — you as a citizen can get your street lamp fixed. Then there’s the messy middle that voters don’t really care about, where a n engaged company can change some banking law nobody’s staring at — a lot of power there. But on the other side of the horseshoe, when it’s big national stuff touching everyone’s lives, stuff voters are voting on, corporations actually have less say. You can debate how that played out in the financial crisis. AI isn’t quite big enough yet — we’re still in the middle, where companies can really drive the discourse. Fast-forward one, two, three years, and we may have politicians running on “screw these guys, here’s what I’m going to do, I don’t care what you think” — and that legitimately winning votes. Today may be peak lab influence on the policy discourse.

Kevin Xu: The labs’ timelines and policy papers align with that projection — they want everything they want done before 2028, because the next presidential AI debate will be at the other end of that curve. We’ll get a preview this midterm. It’s going to be that issue.

OUTTRO SONG

You say you’re a frontier model? Yeah, everyone’s a frontier model tonight, sweetheart.

There’s a model in the building that can hack the world, so the hottest club in Beijing just got exclusive, girl. The Party owns the door, the floor, the lights, key codes, the clipboard — and it’s not polite. No lab picks the guest list, no CEO’s plus-one. The rope went up at midnight — better get a permit, hun.

State Grid, come on in. Sinopec, looking gorgeous. Ministry of Water Resources, love the jacket. Everyone else — sidewalk.

You’re not on the list, you’re not on the list. This is Glasswing with Chinese characteristics. You’re not on the list, you’re not on the list. Dress code is critical infrastructure — that’s it.

Ministries dancing while the startups freeze, from each’s GPUs to each’s API keys. Here comes Jack Ma crying at the door: “Your whole government cloud is running on my cores!” Bouncer taps the earpiece. Long pause. “Fine. But you’re working coat check, buddy. Get in line.”

Provincial gas utility just waltzed on through, running Windows XP and a fax machine too. ’Cause the worse your legacy systems, honey, the better your seat — the most hackable girls get the VIP suite.

Provincial governments — come back Friday. Local SOEs — also Friday. Hugging Face — you got a calendar?

You’re not on the list, you’re not on the list. It’s Auntie Glasswing, baby, and the party rules with an iron fist. You’re not on the list, you’re not on the list. The Standing Committee sent the invites, kid — banks come first while the founders just sit.

Hey, chin up, kid. Everybody gets in eventually. That’s the beauty of the place. Wide release, open to all. Three months, maybe four — call it six.

Sidewalk. Sidewalk.

• , ChinaTalk’s chip lead, is in Hawai’i! If you’re interested in joining an Oahu ChinaTalk meetup, reach out to aqib@chinatalk.media.

• is going to Jakarta later this month and would love to meet folks there! Ping Irene@chinatalk.media to say hi.

Could the future of biotech great-power competition be decided by monkeys?

In 2020, as COVID spread, China stopped exporting research monkeys. For US vaccine developers, infectious-disease researchers, and drug-testing companies, monkeys that once went for only a few thousand dollars suddenly jumped to $20,000 or more. FDA says it now goes as high as $50,000 — if you pay peanuts, you unfortunately no longer get monkeys.

Important projects were scrapped or delayed. The shortage became serious enough that the NIH created a priority process for COVID studies needing nonhuman primates, warning that there weren’t enough monkeys to go around.

China supplied 50.7% of lab monkeys imported into the US in 2006, 69.6% in 2011, 62.8% in 2018, and 48.5% in 2019. Then China’s share collapsed to zero from 2020 onwards. A National Academies report later found that 64% of surveyed NIH-supported researchers had trouble obtaining nonhuman primates.

Monkeys, mostly macaques, still play a big role in biotech. For vaccines against dangerous pathogens, gene therapies, infectious-disease countermeasures, neurotechnology, and complex biologics, monkeys often remain the model regulators, companies, and funders want to see — especially when mice are not similar enough to people and human challenge trials are waiting to get going.

At ChinaTalk, we write about supply chains quite often. This one is different because it is alive.

In one sense, scaling it sounds easy. Put a male and female macaque together, light a few candles, and, with some luck, a new macaque is born. (If only China could do that with two Blackwells.) But achieving monkey sovereignty actually proves to be quite difficult.

This business of monkeys is not, in fact, just monkey business.

Are Monkeys Still Important?

To be clear, many primate studies are expensive, anachronistic, or less useful than newer human-relevant tools. This is why the FDA Modernization Act 2.0 removed the old statutory assumption that animal studies were required to support certain drug applications. The FDA has since said alternative methods — often called New Approach Methodologies (NAMs) — can help reduce or even replace animal testing when they are scientifically valid. (Rachel Fobar at Vox makes a compelling case for phasing out a lot of unnecessary animal testing for things like monoclonal antibodies.)

But unfortunately, monkeys are still stubbornly indispensable for many aspects of biotech. The monkeys specifically relevant here are macaques — especially rhesus and cynomolgus macaques — because their immune systems, brains, organs, and disease responses are much closer to humans than mice or other small animals. And, unlike chimps, they cannot throw a grown researcher across the room.

Some research questions still require an intact, human-like organism. Sure, you can model a liver cell, a lung organoid, or an immune reaction in vitro, but that is not the same as being able to fully model how a vaccine or gene therapy behaves across a whole primate body — much of which is still a mystery to us. This is why the National Academies say there are still no alternatives that can replace monkeys for questions requiring “complete multiorgan interactions and integrated biology.”

Vaccines are the clearest example of monkey indispensability. For dangerous pathogens, macaques are often the closest thing to a human challenge trial that researchers can ethically run. Our most important bio story of the 21st century was the COVID vaccines, and every major COVID vaccine candidate that progressed to US approval was evaluated in nonhuman primates — though to varying degrees, and oftentimes in parallel or after human trials due to the urgency of COVID and issues with procuring the monkeys.

And think about diseases much deadlier than COVID. You cannot prove an Ebola vaccine works by deliberately exposing volunteers to Ebola. Under FDA’s Animal Rule, some medical countermeasures can therefore be approved based on animal efficacy studies when human trials would be unethical or infeasible. For diseases like Ebola, monkeys can be the closest thing to the efficacy trial society is willing to run, and that basic ethical prudence will keep them relevant as a human fail-safe for a long time.

The same logic applies to many other research areas, such as gene therapy and brain-computer interfaces.

How China Became King of the Monkeys

China was not always the 800-pound gorilla of the monkey supply chain. For decades, India was the main supplier. It had large rhesus macaque populations, and the US relied heavily on Indian monkeys for vaccine testing and biomedical research. Then India said, “Not my circus.” In 1978, Prime Minister Morarji Desai banned rhesus monkey exports — the reasons seem to have been a mix of conservation, animal welfare, and religious reverence for monkeys.

China swung in to fill the void. It had native rhesus macaques, access to long-tailed macaques from Southeast Asia, comparable cost for land and husbandry, and a growing laboratory animal system. By the 2010s, it had become the world’s leading exporter. China accounted for ~31.3% of global live-primate exports between 2015 and 2021.1

For countries doing advanced biological research with monkeys, the dependence is even greater. Research-grade monkeys need health records, pathogen screening, and paperwork that regulators and companies trust. China’s system was not perfect, but it was more reputable than many alternatives like Vietnam and Cambodia, especially because China has a vested interest in using these monkeys for its own bio-research. This is why, according to China Daily, more than 60% of the 33,818 primates the US imported in 2019 came from China.

Then China’s domestic demand went bananas. In 2015, along with its infamous biotech push, Beijing overhauled its drug-review system, speeding up approvals and helping turn China into a serious biopharma market. The combined value of China’s leading biopharma companies rose from $5 billion to $400 billion between 2016 and 2021. More Chinese companies began running more drug trials, and those trials needed monkeys.

Fast forward to January 2020: China bans wildlife trade to prevent the spread of possible zoonotic diseases. Officially, this was about blocking the possible spread of coronavirus transmission. In practice, it also allowed scarce monkeys to be redistributed at home.

Chinese companies noticed. WuXi 药明, Pharmaron 康龙化成, and JOINN 昭衍新药 all moved to secure breeding capacity. The WuXi companies became China’s largest monkey breeder after acquiring Suzhou Kanglu Biotechnology 苏州康路生物科技 in an 804 million yuan deal. This is the same Chinese pharma-services giant US lawmakers have targeted through the BIOSECURE Act — which we covered here. Monkeys give WuXi and other Chinese pharma companies another kind of leverage that people in the China Select Committee and other relevant agencies taking aim at them should keep in mind.

Curiously, this also does not look like an EV-style overcapacity story, where China needs foreign markets because domestic supply has outrun demand. Chinese press reports have described experimental monkeys as hot commodities, with prices rising sharply as domestic drug R&D expands. Even inside China, there often are not enough monkeys to go around, and the shortage has produced accusations of “monkey hoarding” (“囤猴”) — much like Sun Wukong hoarding peaches from the Queen Mother’s garden. This also helps explain why US imports from China did not rebound even after China’s COVID restrictions were eventually loosened.

How to increase monkey supply

The US has a few options.

(1) Buy monkeys elsewhere

That is basically what the US did after China’s exports collapsed. The replacement supply chain shifted toward Cambodia, Mauritius, and Vietnam, as the graph from the intro shows:

But these suppliers are not always dependable. A research monkey needs reliable records, such as where it was born, who its parents were, how old it is, what pathogens it has been exposed to, and whether it was actually raised in captivity. A wild-caught monkey falsely labeled as captive-bred can be a disaster for an unwitting company. Imagine a vaccine developer spending months and millions of dollars on a macaque challenge study, only to discover that some of the animals had unknown disease exposure. If the animals were already exposed to related pathogens, their immune responses may not reflect the vaccine being tested. Or if their ages or health histories are wrong, the comparison between vaccinated and control groups may be meaningless. If the paperwork is fraudulent, regulators may simply decide the study cannot support a clinical trial — and millions of dollars + years of effort could be wasted!

Countries like Cambodia have struggled to demonstrate reputability. In 2022, as one such example, the US Justice Department indicted Cambodian officials for allegedly capturing long-tailed macaques from the wild (à la Man with the Yellow Hat from Curious George), laundering them through breeding facilities, and exporting them as captive-bred animals. China, by contrast, built a larger and more institutionalized monkey ecosystem. China is not free of smuggling/laundering, but its major breeders and CROs have stronger incentives to keep animal standards high because they are serving China’s own biotech industry as well as foreign customers.

(2) Breed monkeys at home

This is the cleaner answer in many ways, since a domestically bred monkey is easier to audit. The US can have more confidence in its age, parentage, pathogen history, and health records.

But domestic breeding is not something the US can simply scale overnight. Macaques grow and reproduce slowly. It takes a while to find a prime mate. Gestation lasts about 5.5 months, pregnancies usually produce one infant, and females generally reach sexual maturity around 3-4 years.

Say the US has 100,000 suitable monkeys (based on this broad dataset from the Animal and Plant Health Inspection Service). Roughly half are female, so 50,000. But maybe only 40% of those are actually breeding-age, healthy, fertile, socially compatible, and appropriate to breed — call it ~20,000 breeding females. If 70% produce a surviving infant in a given year, that gives you roughly 14,000 new monkeys per year. But those 14,000 infants do not help much right away. Many studies still want animals that are 3-5 years old, research-naive, and of the right sex or immune background. And only about half of the new infants are female, meaning perhaps 7,000 future breeders, who themselves will not start reproducing for another several years.2

^ Mice, on the other hand, can reproduce exponentially.

There is actually quite a lot of research on how to optimize breeding yields, but the results are relatively modest, and the studies themselves are a bit depressing to read about.3

The other big issue is animal welfare. Animal rights groups are more empowered in the US than in China, and the US does seem to treat lab animals somewhat better. World Animal Protection’s Animal Protection Index gives the US an overall D and China an E. More relevantly, on animals used for scientific research, the US gets a C, while China gets a D. That’s not exactly a ringing endorsement of the US system (props to some of those European countries) but it’s something.

(3) Buy the foreign supplier, then make it operate to your standards

That is what Charles River Laboratories is trying to do. In January 2026, Charles River said it would buy K.F. Cambodia, one of its major nonhuman-primate suppliers, for about $510 million. Reuters reported that K.F. had supplied around 30% of the monkeys used by Charles River’s drug-discovery and safety-assessment unit over the previous two years.

This may be the most realistic middle path. The US cannot breed its way out of the shortage quickly, and it also cannot assume that every foreign supplier’s paperwork is trustworthy. But American companies can streamline overseas breeding capacity.

I’m sure this approach still has many shortcomings. It probably depends on host-country politics as well as animal welfare and disease control standards, and involves sending capital and leverage outside of the US. Perhaps someone with expertise in international business can weigh in in the comments?

Subscribe now

None of the three options above seem like a great alternative. I can therefore understand why the FDA is pushing hard to figure out when monkeys are no longer necessary, but I hope doing so doesn’t lead people to conclude that this is not an actual problem.

Let’s stop monkeying around

There are already people who have stopped jumping on the bed and begun thinking seriously about this shortage. The NIH has produced extensive and relatively approachable research on the problem, so I won’t really have anything new to add here. But there are a few obvious next steps.

1) The US should know what monkeys it actually has. There is still no comprehensive mechanism for tracking nonhuman primates used in US biomedical research, nor a “central data management or reporting structure” for accurately estimating how many monkeys are needed to meet current and future research needs. When managing a bottleneck, it’s hard to monkey-do when you cannot monkey-see.

2) The US should build more domestic capacity while being honest about the timeline. A 7-10 year goal for something like 80% self-sufficiency seems reasonable to me, but the next step should involve experts weighing in for a better, more evidence-based estimate than that. Regardless of the number, the US is going to have to be serious about this for a consistent period of time that will span different administrations.

3) The US should decide ahead of time who gets scarce monkeys during an outbreak. If a new pandemic hits and ten labs, three vaccine companies, CDC, NIH, and DoW all need the same animals at once, the priority rules should not be improvised on the fly. The country should not be having a monkey bidding war in the middle of a public health emergency.

4) The US should keep funding alternatives that actually reduce monkey demand. Even though NAMs will not make monkeys disappear tomorrow, they can help kill bad candidates earlier and eliminate studies that do not really need primates. We should spend more money figuring this out.

5) The US should work more seriously with countries like Cambodia, Vietnam, and Mauritius. A pure import strategy is currently too fragile given the risk of getting bad monkeys. But hybrid models, like Charles River buying and auditing K.F. Cambodia, may be part of the answer.

6) Congress should oppose bad bills like the PRIMATE Act. The bill would prohibit most imports of nonhuman primates into the US, framing them as a biosecurity threat because they can carry pathogens like tuberculosis, herpes B virus, and Salmonella. But the biosecurity argument makes no sense. The scale of the problem is basically zero, and >99.9% of TB and Salmonella cases arise through pathways completely unrelated to monkey imports. TB only caused 572 US deaths in 2023, while CDC estimates Salmonella causes 238. There have only been about 50 documented human infections of herpes B since 1932! I care deeply about both animal welfare and biosecurity — but attempting to disguise an animal welfare bill under the guise of biosecurity makes both concepts less reputable. And the 100000x more urgent biosecurity risk is being unable to test vaccines and therapeutics quickly in the next pandemic.

7) Finally, the US should be willing to pay more for better living conditions for research monkeys. Our D rating on the Animal Protection Index is something to be ashamed of, and improving it will furthermore make the issue less politically divisive with animal welfare groups. The alternative is importing monkeys from countries where the same (or worse) ethical problems exist but are simply less visible to Americans (like Vietnam, which has an F rating). That “see no evil” approach is not ethical progress but rather outsourcing ethical discomfort.

*Special thanks to Lily for going ape and providing the majority of the monkey puns in this article

Randy Schriver is Chairman of the Institute for Indo-Pacific Security (the successor to Project 2049), current chair of the US-China Economic and Security Review Commission, and served as Assistant Secretary of Defense for Indo-Pacific Security Affairs in the first Trump administration.

We discuss…

• China’s new persistent Coast Guard patrol east of Taiwan, and how lawfare, fishermen rescues, and radio queries make sovereignty claims physically manifest,

• Why treating arms sales as a “bargaining chip” puts the US out of compliance with the Taiwan Relations Act,

• Whether Seoul and Tokyo go nuclear as hedging accelerates — and what the axis of autocracy means for a Korea contingency,

• Blockade vs. invasion, Hellscape, and why the US should exercise a Keelung–Yonaguni corridor with Japan,

• Why rebuking Takaichi blew Randy’s mind, plus Pelosi vs. the blocked Colby-era CODEL and Hegseth’s two Shangri-La speeches,

• The own-goal management of US-India relations, plus Randy’s reading list for new Indo-Pacific hands.

Listen now on your favorite podcast app.

Read What They Write

Jordan Schneider: Let’s zoom out a bit. How worried are you, Randy?

Randy Schriver: Well, I take China at their word. Some people talk about China being enigmatic, difficult to understand. At my think tank, we have a very sophisticated approach to trying to understand China: we read what they write and we listen to what they say. Or the way another friend of mine, James Mulvenon, puts it — the first line of encryption for the Chinese is Chinese. If you can get through that, they’re pretty clear about what their ambitions and intentions are.

I do think they want to win without fighting. August 1, 2027 is a be-ready-by date, not necessarily a go date. And I suspect there are a number of things — including watching Russia’s experience in Ukraine, but also their own internal situation, the massive purges that have been going on in the military, and, in the Rumsfeldian lexicon, the known unknowns: how does a military perform that hasn’t seen combat since 1979? All of these things probably give them a degree of pause, or at least an understanding that there’s risk. So I think they would like to win without fighting and get some sort of political capitulation without having to fire a shot.

But even in so-called peacetime, if you look at what’s happening around Taiwan, the levels of coercion and intimidation are pretty extraordinary. They integrate lawfare, political warfare, information warfare. One of the most significant developments of the last few weeks is the announcement of a persistent Coast Guard patrol east of Taiwan. From a lawfare perspective, this is China exerting through its activities and actions a claim: we are the sovereign power with the sovereign right of law enforcement in this area east of Taiwan.

They’ve already queried commercial ships, and Taiwan has said don’t answer those queries. But if you’re carrying cargo and have to keep your insurance, the easiest thing to do is pick up the radio and respond to the query. Or imagine a fisherman in distress: if a Chinese Coast Guard vessel is the closest vessel and makes the rescue, everybody’s going to say, wow, what a great humanitarian thing they’ve done. They’ve actually just physically made manifest their sovereignty claim.

So yeah, I’m worried. And I think the recent statements by President Trump after his meeting with Xi Jinping in Beijing were not helpful. He seemed to have absorbed a lot of the Chinese narrative. “Taiwan’s been part of China for thousands of years” — well, no, it hasn’t, and it’s never been part of the People’s Republic of China. “We don’t want to fight a war 9,500 miles away” — I certainly don’t want to fight a war with China, but this 9,500-mile thing — look at where our forces are: Guam, Hawaii, Japan, Okinawa. Let’s not just absorb everything Xi Jinping has said.

And then the statement about our arms sales being a bargaining chip — I would submit that puts us firmly out of compliance with the law. The Taiwan Relations Act says we must provide Taiwan with weapons of a defensive character sufficient for self-defense. That reads as a must. You must make them available. It’s not a choice, it’s not a policy option. And we also have a part of the law that says we must maintain the capacity to resist force. Wearing my commission hat — the US-China Economic and Security Review Commission — last year we put a recommendation to Congress to say: show your homework, do the math. Do you have the capacity to resist force with everything that’s going on in Ukraine and Iran? With everything they’re saying about munitions, everything they’re saying about logistics and maintenance support — don’t look behind the curtain. I’m not sure I’m buying. They’re pitching, I’m not catching. I think we’ve got issues on that front.

So there’s a lot to worry about. But I’ll put on the other hat and say Japan’s never been more engaged. They said they were going to double their defense budget in five years — they did it in three. They’re giving us access to the Southwest Island chain. Yonaguni is the closest island to Taiwan — it’s not China, it’s not the Philippines — and we’re now able to do activities out of there. The Philippines, with the EDCA sites in northern Luzon and Palawan. Japan-Philippines cooperation — we’re not even in that hyphenated minilateral all the time — and all of that to me is very encouraging. So it’s a mixed bag. I just hope we keep our eye on the ball.

Overplaying the Hand in the Pacific

Justin: The Japan-Philippines piece, tying that back to your discussion of the eastern patrols, is interesting because that’s obviously their fear of encirclement. When you couple that with the launch of the Chinese ballistic missile earlier in the week — and now some of the islands in Micronesia and the Pacific Islands are starting to raise concerns about the Chinese threat outside that direct sphere — what are you starting to see or hear about the way they’re viewing China, given how China is acting toward some of their larger partners?

Randy Schriver: I think for the most part, China has overplayed its hand, and its provocations have tipped some of these countries toward actions they might not otherwise have taken. Look at the defense treaty that was just signed by Prime Minister Albanese and Fijian Prime Minister Rabuka. That’s the first defense treaty Fiji has ever signed, and it’s only the fourth for Australia — US, New Zealand, Papua New Guinea. PNG was another island moved to take such an action, and that was two years ago. So within the last two years, you’ve got PNG and Fiji saying they want to be defense treaty allies with Australia.

The compact states are in a bit of a different category, but I was recently in Palau, where they are refurbishing the airstrip at Peleliu — quite a history there. We fought an extremely bloody battle over that airstrip. Turned out we didn’t need it for the invasion of Japan after all. But the fact of the matter is the geography is still quite important, and Palau obviously has the greatest proximity to the Western Pacific, where potential conflict contingencies might happen. And there are activities in the Federated States of Micronesia and the Republic of the Marshall Islands.

So I think we’re seeing the second-order effects of overreach on the part of the Chinese. When their face and their activities are in the economic domain — providing investment and development assistance, setting aside that a lot of it is predatory and has debt-trap features — countries in the Pacific are much more open to that and welcoming of it, because they’ve got such extraordinary needs. But when they see the PLA, or even the Coast Guard or the maritime militia, enabling illegal fishing in their territories and other illicit activities, and you see something like the submarine launch of a JL-2 or JL-3 — I don’t think we know yet which — that’s getting people to really consider their defense needs and take steps they might not otherwise take.

Hedging Toward the Bomb

Tony: When Ely was on the show, we talked about what we call the bullshit détente — we’re not really talking about China at the strategic level because of Iran and everything else, and because the president sort of wants to be friends with Xi Jinping. But then all of this is going on in WESTPAC. As that gap widens, is there a point of no return, where there’s so much difference between what our friends in WESTPAC are doing and what we expect that either they go it alone, or they decide it’s not worth it because the US isn’t going to show up?

Randy Schriver: I suppose theoretically there is. The interesting thing to me about where we are now is that most of these countries are doing things we have long encouraged — even if they’re doing it as a hedge, even if they’re doing it because they’re concerned about the reliability and credibility of the United States. These are nonetheless things we’ve long wanted: Japan doubling its defense budget, reinterpreting Article 9 of the constitution, introducing counter-strike — which to the untrained eye looks a lot like strike — creating a joint operations command. Korea, in terms of their defense spending and preparedness. All these are things we’ve long encouraged.

So as this goes forward, if it’s still the result of a hedge and starts to accelerate because they don’t believe in the credibility of the alliance or the United States, then we tip into another era, I guess. And there are a couple of things talked about now that would be indicators of that. If South Korea goes nuclear — and look, popular sentiment in South Korea is over 70% for an independent, autonomous nuclear capability. That’s extraordinary. I don’t know that it’s a very well-informed view when it comes to what it would require and what it would all necessarily mean, but it’s extraordinary. And in Japan — a taboo subject for a long time, for very understandable reasons — it’s openly discussed and talked about now. That would be a step beyond the minor hedging and outreach we’re seeing right now.

Jordan Schneider: Can we talk about that world?

Justin: Let’s say we have a bunch of equal nations, no leadership — this is kind of the Carney vision of Europe sans the United States, where the EU and France and all these countries vie for levels of influence and power, with no overarching first among equals. Say we apply that same approach to Seoul, Japan, even Australia with regard to China. What prevents that from turning into Chris Clark’s The Sleepwalkers — where, because there’s no one there to pull the reins or set conditions or make a strategic move, all these different countries slowly make moves in their own best interests, and those interests just nudge us toward what ends up being a cataclysmic war?

Randy Schriver: That’s a possibility, and a pretty dark outcome. I think there’s a reason our alliances have been so resilient for so long. World War II was, by my measure, quite some time ago, and yet we still have this strong alliance with Japan — and popular support for the alliance is still very high, notwithstanding America First rhetoric and questions about credibility. In the Philippines, I think we poll at over 80% popularity as an alliance. Australia, over 70%. There’s a reason they’ve endured. It’s been interest-based, but there’s also the track record of stability and security.

You have this notion of a trapeze artist — you really don’t want to let go of one thing that’s been pretty good to you until you get your hand on the other thing. There are scenarios where there’s a degree of stability if South Korea and Japan have nuclear weapons and nobody’s going to mess with them. But South Korea would be facing not only North Korea but the growing axis of autocracy. My Chinese friends will tell you up and down — it’s not a thing, it’s not an alliance, there’s no secretariat, no coordinating meetings. My response is: first of all, that’s not a weakness, that’s a preference. China doesn’t like binding alliances where they’re bound to support countries in certain ways. Number two, even limited cooperation can sure be consequential. Look at what’s happening in Ukraine, where North Korea is providing forces, artillery, and ammunition to Russia, and China is providing material support in terms of components and technology for drones.

So if North Korea gets even that sort of minimum support — Russian troops, Chinese material support — in a conflict with South Korea, that’s a much more difficult conflict for South Korea to deal with, let alone if they’re dealing with it without US assistance. And would US forces stay on the peninsula? We’ve had two presidents try to pull them off — Jimmy Carter, and President Trump in his first term. Pull this thread: are we in a more stable, secure world? I sort of doubt it.

Blockades, Hellscape, and the Yonaguni Corridor

Bryan Clark: Randy, talking about Taiwan — obviously the will to fight is a question, but in a lot of ways they’re building systems that make it almost automatic that they’re going to fight, right? This Hellscape thing that INDOPACOM is pursuing — the Hellscape robots are basically going to attack whatever’s coming at them like a bunch of rabid dogs. And then coastal defense cruise missiles — all the stuff they would do, which is in a lot of ways automatic, is just going to impose a lot of casualties. But to what degree does Taiwan also need to invest in the kind of stuff that would help them break a blockade? Because if China is worried about casualties and about the risks of an invasion failing, it seems like squeezing them — getting them to where they can’t make twelve days of LNG imports — is an easier road to getting the regime to capitulate. Is Taiwan preparing for that contingency in their force development?

Randy Schriver: They are, and at times that has put them at odds with the United States. Ely and I — Republican, Democrat — probably agree on 99% of things. You’d be hard-pressed to get us to debate something. But I would say the previous administration chose to look almost exclusively at the counter-invasion capabilities. And I sort of get it: limited resources, where are you going to put them? Not necessarily the most likely scenario, but the most dangerous one, okay. But I think a far more likely scenario is a blockade or a soft blockade — and that’s part of deterrence too. In fact, the two are related. You would actually, unintentionally, make the invasion more likely if there’s no pushback in the gray zone.

I think we could be doing more in terms of networking the sensor piece with Japan and the Philippines and Taiwan. And we should be exercising with the Japanese and the Filipinos — I think we should exercise what it would take to have a corridor from the northern port of Keelung to Yonaguni, and demonstrate it, show that we’re willing to do that. That would be both prudent in terms of preparation and have some deterrent impact.

Bryan Clark: Sure. And you could use some of the stuff they’re putting in Yonaguni to support that.

Randy Schriver: Absolutely. And you have a window with the Japanese. Again — irony after irony after irony. When Prime Minister Takaichi said the quiet part out loud about Taiwan being a “survival-threatening situation” — awkward English, but that’s actually the Japanese term in the Diet legislation — she was rebuked by the Trump administration. It blew my mind. For decades, we wanted Japan to say something like that.

Justin: That’s the hard part — in some ways the administration has been right. I was in Taiwan during the middle of the Biden administration, and it was kind of verboten to bring up invasion and defense with your Taiwanese counterparts outside of specific settings, because there was a fear you’d scare them. Then a company in Taiwan produces Zero Day, an amazing docudrama about the invasion — great show — which shows they actually are willing to talk about it, they understand the neighborhood they live in, and the society is willing to watch something like this and understand the problems. The inverse being: we change administrations and switch from not wanting to talk to the Taiwanese to all of a sudden not backing up our allies when they do things. It’s the same mistake we made with Australia pushing to investigate the origins of COVID — China cuts off their imports, and what the United States doesn’t do is say, hey, we’ll buy your wine. Instead of backing them up.

Randy Schriver: I stepped up on that front personally.

Bryan Clark: And thank you.

Two Entirely Different Administrations

Tony: There’s an interesting comparison here. During the Biden administration, there was the freakout from the executive over Pelosi going to Taiwan. And then in this latest spat between Mike Rogers and Bridge Colby, it came out that allegedly Bridge had blocked a Republican congressional delegation to Taiwan. That’s such a reverse course from all the messaging over the last few years. It somewhat aligns with Bridge’s book and his focus on stability — but that seems more indicative of how this administration views the conflict than anything else.

Randy Schriver: I consider Bridge a friend. I actually took him to Taiwan for his very first visit with a think tank delegation, and then again the spring before the 2024 election, where on at least one occasion he and I debated one another in front of Lai Ching-te 賴清德, which was kind of interesting.

The story on the congressional delegation, if I understand it correctly, was that he wasn’t going to provide an airplane, because most congressional delegations get milair. And I mean — how in the world do you have a department where your two authorizing committee chairs in the Senate and the House, chairs of the same party, are saying you’re not telling the truth and criticizing you? How did you get to that point? It’s a little mind-boggling, because it’s not hard. You go tell these guys that you care what they think, that they’re important — anytime you want to speak, please. That’s it. How they got to that point is just remarkable.

On the differences: the problem with the Pelosi visit wasn’t the visit per se — it was the Biden administration being publicly against it. The Chinese, as sophisticated as they are — and they’re so much savvier now about our decision-making process, the role of Congress versus the administration — thought, “We can stop this. The Biden administration doesn’t want her to go.” Now, that ignores any understanding of how Nancy Pelosi operates. But I think that’s how we got into trouble. If the Biden administration had just said, “Separate branch of government, we’re not going to intervene one way or the other, she has a right to make her own travel decisions,” I think it would have been a little different. Maybe not, but I think it would have been different.

My issue with where the Trump administration is now: number one, if you read Secretary Hegseth’s Shangri-La speech from a year ago and his Shangri-La speech from this year, it looks like two entirely different administrations — over the course of a year. If you look at a National Security Strategy that talks more about Taiwan than any NSS I’ve ever seen, and then an NDS that doesn’t mention it at all, that looks like a disconnect. And when Secretary Hegseth says we want to be — what’s his formula, quiet and strong? — the quiet part, I think, doesn’t quite capture how deterrence works in the Chinese mindset. When you’re quiet, you won’t have Lai Ching-te transit the United States; you knock the Monterey military talks down to a lower level and hold them in Alaska. What the Chinese see is: are the Americans really willing to fight for this? It sure doesn’t look like it. Some of this symbology actually matters in the deterrence equation. Whereas Bridge would say nothing else matters other than whether you are militarily capable — and that reverts to the point I made earlier about readiness.

I was worried about munitions and contested logistics before Ukraine, before Iran. Are we better now? Surely we’re not. So the quiet part — debatable. The strong part — extremely debatable. I think there’s a lot more work to be done here. I’m glad allies and partners are stepping up. The part of Secretary Hegseth’s speech I did like was that he went through each ally and was complimentary, which was not always the case — certainly not if you’re in NATO. Praising Korea and Japan and Australia and India, that’s good. Omission of Taiwan — not so good.

Bryan Clark: Can we double-click on that? Japan is obviously investing a lot more in its military — we’ve been doing some work with Japan on its force design. It seems like they’re still looking to build a military that complements the United States military, assuming the US is doing a lot of the heavy lifting in Japan’s defense. They’re not looking to field the kind of capabilities you’d field if you thought you’d have to defend yourself for some number of weeks. Same with Australia — not increasing defense spending very much, if at all, but investing in the kind of stuff that makes them complementary to the US rather than an independent actor. Do they need to be investing differently?

Randy Schriver: It’s a very fair point that our allies are investing in ways that are complementary to the US in support of alliance and coalition operations, and I think we should be grateful for that. If there’s a radical departure from that, it’s another one of these potential mileposts suggesting they’ve given up and don’t regard us as reliable or credible. And a large portion of this is also fungible: Japan needs to invest in integrated air and missile defense in southern Japan, including the Southwest Island chain, whether or not US forces are going to operate out of there. So first of all, we should be grateful they’re doing it, and secondly, I don’t think we’re at a point where we’re saying this investment pattern is somehow off.

Justin: That’s where the issue runs in with the way we’ve gone about challenging our allies and partners. I don’t think Japan is going to be in a position where they have a self-sustaining military that can conduct operations on its own — nor do we really want that. Same with South Korea, same with Australia. But if they start moving in that direction, and instead of complementing our capabilities they’re replicating the capabilities we’re supposed to bring into a fight because they’re afraid we won’t be there — you’re really just lowering the defense posture of everybody. And we’re not talking about replicating production capability, just replicating the niche things the US says we’re going to bring into a fight in the Pacific.

Randy Schriver: Yeah — and if you look at certain aspects of this, it would really be years and years. Japan is investing in counter-strike. Looks like strike. But the kill chain is still very reliant on US sensing and tracking, and for Japan to say “we’re going to have a completely autonomous kill chain” — that’s space, that’s other airborne sensors, that’s a lot of integration of data. That’s going to be years.

Justin: There’s a new acronym — it used to be C2, and now it’s probably C7IS4T-something-squared. Whatever the new acronym is, one of the Cs is “computer.” I don’t understand how that matters or why we need to make that differentiation.

What Actually Deters Beijing

Tony: Back to the deterrence part. There’s a lot of tea-leaf reading that goes into how the CCP and the PLA operate, and I’ve heard a lot from folks in government about what the CCP really cares about these days in terms of messaging. What are the actual deterrence activities we can undertake that can impact CCP decision-making?

Randy Schriver: The military piece is not insignificant. Continuing robust security assistance to Taiwan; the posture initiatives — Ely was terrific on that; the opportunities for diversification and dispersal during a contingency — all of that is really important. But if you reverse-engineer what the Chinese really fear, it’s the alliance piece: the cohesion among like-minded countries with shared interests and concerns, and the thought that a move against Taiwan would bring in a very powerful coalition of countries — on the military side, which would likely be limited, but also on economic cost imposition and sanctions.

The thought that we’re telling the Europeans — after decades of trying to get Europe interested in Asia — “you’re not needed, you’re not wanted, take care of your own backyard” is extremely counterproductive from a deterrence point of view. You want the Chinese to believe they can’t fight an isolated war in the Taiwan Strait — that the potential for the Malacca Strait to be closed is very real. Remember, it was the Chinese — Hu Jintao — who came up with the phrase “Malacca dilemma,” not the United States. You want them to think there could be multiple fronts, with potential for India. You want them to think the EU is going to act swiftly and in unison on sanctions. That has always been our asymmetric advantage — the ally and partner piece and our ability to hold it together in a meaningful, credible way. That’s where the work needs to be done. And some of the rhetoric suggests otherwise — certainly the trip to NATO where, evidently, we’re not only ending the ceasefire with Iran, we’re also threatening Greenland again. And by the way, this American remembers that Denmark lost more people per capita in Afghanistan coming to our defense than any other country outside the United States. The fact that we’d be threatening them is unconscionable.

Justin: I’ve had the same experience. When I talk about having been in dangerous situations, especially in Syria — if it wasn’t an American pilot above me, the Dutch pilots were the ones I most wanted flying overhead. Probably the hairiest day I ever had in Syria, it was two Dutch pilots dropping bombs for me through a dust storm to help me break contact. I say that to say: a hundred percent. The lack of acknowledgement of where we are, what we’ve done, and the length and depth of these alliances is one of the things that feels the grimiest in these conversations. Because yes, there are countries in NATO that need to spend more money. Yes, there are countries that have been self-interested — the Rheinmetall CEO tells the story of having a conversation with Merkel, saying she needed to invest more, and she says, “We’ll just call on the Americans for defense.” We made it easy for some countries not to invest in their own defense apparatus. But what that did in turn was invest in our defense apparatus — jobs, technology, things that went into the US ecosystem. That’s now changing. Do we want to shepherd that change, or allow it to be driven by Mark Carney and Emmanuel Macron and people who may not have the best interests of the US at heart?

Randy Schriver: Couldn’t agree more.

Tony: On the NATO part — we’ve seen what goes wrong there: you end up with thirty different defense industries in Europe, none of them capable of expeditionary fighting, all with different battle plans and no real ability to surge forces quickly in a conflict. Ely, when he was on the show, talked a little about the need for some sort of collective defense arrangement. Is there a way, without disrupting the cultural restrictions on collective defense in Asia, to get all the countries together to talk about complementary defense planning?

Randy Schriver: There is, but we have to embrace it ourselves, endorse it, and be the leader of such an effort. Ely literally wrote the article about an Asian NATO. And there are people in Asia — former Prime Minister Ishiba of Japan has long called for an Asian NATO. So it’s not taboo. It would be very difficult to operationalize and make meaningful, but not impossible. It is impossible without the United States.

Own Goals in Delhi

Justin: One thing that always gets left out: India has always been a very tough country to bring in politically, and if you’re going to have a NATO in Asia, you want India in it as a bulwark. How has the administration’s approach to India changed or morphed over the last year or two? When you look at issues like the Himalayas and the Line of Actual Control, where are they putting the right and wrong emphasis on that engagement as a counterweight to China?

Randy Schriver: If I were to rack and stack all the own goals I’ve seen over the last two years, the management of the US-India relationship is near the top, if not at the top. Our future is not Pakistan — it’s with India. For years I was going to Delhi and hearing about the Chinese, the growing threat, the long-term threat — you’d never hear about Pakistan anymore. Now it’s back. We’ve put Pakistan in the center of our negotiations with Iran. President Trump hosted the chief of their armed forces and praised him as a brilliant general in the Oval. We’ve tariffed and criticized the Indians. The timing of hosting the Pakistanis in the Oval Office so soon after the dust-up between Pakistan and India was extraordinarily bad.

Look, in my time in office we got a lot of help from the Chinese — Doklam and Ladakh, two pretty big dust-ups in a two-to-three-year time frame. I think that pushed the Indians to more fully embrace the Quad. It allowed us to move forward on the defense relationship in ways they might otherwise have been reluctant to. We might get help from the Chinese again in the future, and hopefully we’ll leverage it for all the benefit we can. But right now there’s a great deal of mismanagement. Prime Minister Modi is going to do his hedging. There’s virtually nothing we can do to peel them away from Russia. But there are little things that are actually pretty important: they’ve lifted the restrictions on Chinese FDI in India — maybe not a little thing at all, actually a pretty significant thing — and signs that we’re moving in the wrong direction.

Justin: The FDI one was a big one. And tariffing India after we spent more than a decade trying to beat down the tariffs India had already put up — trying to open it up to be more capitalistic, more modern, more integrated into the global economy. The own goals are not limited to military or strategic cooperation. We racked them up and knocked them all down.

Randy Schriver: I don’t think the damage is permanent, particularly in the case of India, because our strategic interests will continue to be largely aligned and will compel us to cooperate. Even at one of the low points, after the Pakistan engagement in the Oval, we renewed a ten-year defense framework agreement with the Indians, and a lot of that work is actually advancing. I’ve participated in some Track 2 dialogues where I’m the only optimist in the room — if you deal with the Indians, they’re not subtle or shy about telling you their opinion, and I think a lot of Americans were absorbing that and reacting to it. I said, look, I think we’re going to be in a better place a year from now, and an even better place five years from now. And I am optimistic. But what’s the rule about being in a hole? The first order of business is to stop digging. I don’t think we’ve stopped digging quite yet.

A Reading List for the Indo-Pacific

Jordan Schneider: Any final thoughts — anything else, Randy, we should be asking you about?

Justin: I went to the Indo-Pacific late in my career — I spent the overwhelming majority of it Iran and Middle East focused. And I think it was purely luck that I didn’t fall into reading The Hundred-Year Marathon and having that be my viewpoint on China. For mid-career and early-career people coming up: what’s the base level of knowledge they need about India and the Indo-Pacific? If you had to name five books, what do they need to read? What media should they be consuming to be at a good jumping-off point?

Randy Schriver: Well, I welcome the opportunity for a little self-promotion. I’m the chair this cycle of the US-China Economic and Security Review Commission, and I think our annual reports are phenomenal. I liken it to the old Encyclopedia Britannica — maybe some younger people won’t know what that is — a thick book that, when it arrives, you look at it and say there’s no way I’m getting through all of that. But it’s on the shelf, it’s a great resource, and if you want to know something, chances are it’s in there. We take pride in looking beyond the inbox — emerging issues, calling on Congress to do things that aren’t necessarily this year’s priority but had better get looked at in the not-too-distant future.

And then my think tank, the Institute for Indo-Pacific Security — we’ve been China watchers from the start. We’re the successor brand to Project 2049, which was an early adopter of this idea of read what they write, listen to what they say, and document it. We’ve got a product line that I think is very strong in terms of understanding the Chinese.

There’s a ton of really good books out there. I think history is still very important in this region — because everybody’s trying to manipulate it. Again, irony along the way: the Chinese used to always accuse the Japanese of forgetting history, manipulating history. The Chinese are masters at it. If you go through the Museum of the History of the Communist Party near the Forbidden City, you’ll notice some portraits missing from the general secretaries of the Party — you won’t find Zhao Ziyang 赵紫阳 there. They’re master manipulators of history.

So I’d look at The Generalissimo, by a retired Foreign Service officer [Jay Taylor] — a biography of Chiang Kai-shek, but with a lot of the history of the Civil War and its aftermath. It debunks a lot of things. For example, the Kuomintang Nationalist forces probably took about 90% of the casualties in the war with Japan, and the CCP was basically able to husband its forces and resources to fight the Civil War after the end of World War II. That’s very well documented in the book, and it’s an important part of history — the Chinese now try to co-opt the Flying Tigers and things like that. It’s just outrageous.

I also just read a good book on George Marshall’s mission, written by Daniel Kurtz-Phelan [The China Mission] — really extraordinary history for understanding China-Taiwan today. You get a very in-depth, sophisticated look at how these differences emerged and evolved, and how it explains where we are today.

I’m less well-read on Indian history, but I’ve learned from Indian friends that you can’t have a conversation that goes on too long before they start talking about Partition. It all goes back to Partition.

Jordan Schneider: Midnight’s Furies?

Randy Schriver: Yes — extraordinary book about Partition and how it explains, to a great extent, what South Asia continues to experience to this day. And one other book I thought was extraordinary: The Blood Telegram — about the consul general in Dhaka, Archer Blood, who wrote a dissent cable about the genocide going on in what was then East Pakistan, which became Bangladesh in 1971. They had created the dissent channel at the State Department because of reporting that was being suppressed during the Vietnam War — and the first guy to use it is summarily fired by Henry Kissinger, because he said there was a genocide going on. God forbid somebody report on that.

So yeah, I’m a big fan of history. Probably the frustrating — and maybe somewhat endearing — part of dealing with Americans is that we’re very ahistorical, so we come to a lot of conversations extremely ignorant. When you come with a little bit of knowledge about what these countries have gone through and why we are where we are today, it gives you a leg up.

Jordan Schneider: Amazing — I think we’ll end with that. Randy, thank you so much. Next time we’ll bring you on with Ely and have you two find that one percent and throw some haymakers at each other.

Randy Schriver: Let’s go for it.

From time to time, people come to us hoping to pitch a story on “Eastern Data, Western Compute” 东数西算, the government initiative to route more of China’s data center and computing buildout away from the crowded eastern seaboard and toward national compute hubs in the country’s interior. In theory, western China has land, energy, cooler climates, and underdeveloped regions in need of a growth model. Connect the two, and voilà…a national compute grid for the AI age.

Your humble authors, Irene and Nick, have likewise been lured in by the siren call of a Chinese megaproject in which every province slots neatly into a national AI strategy, Guizhou stores the data, Shanghai and Hangzhou spin up the applications, Shenzhen adds the robots, and the whole country hums along as one vertically integrated socialist cloud.

But, like Odysseus’s sirens, the call is deceptive. The more we looked, the more “Eastern Data, Western Compute” seemed less like a radically new geography of Chinese AI infrastructure than the same pattern you’d find in other places. Some of that pushes capacity westwards, but most of it pushes capacity to the exurbs and industrial hinterlands of already-rich and technically-savvy eastern metros.

Beyond merely being illusory, poorer western provinces, swooned into building data centers on a shaky reading of the market signal, may be the policy’s biggest losers.

Where Are the Chips?

The China Academy of Information and Communications Technology (CAICT), a research institute subordinate to China’s Ministry of Industry and Information Technology, published data about the country’s data center distribution in 2025. When we look at where China’s chips are actually located, the top provinces and municipalities are not a roll call of the remote west.

1. Hebei

2. Guangdong

3. Jiangsu

4. Guizhou

5. Inner Mongolia

6. Shanghai

7. Zhejiang

8. Beijing

9. Shanxi

10. Shandong

There are certainly western and interior provinces on the list (shout out to Guizhou and Inner Mongolia). But the overall map doesn’t suggest China has moved compute to the west. It suggests China has placed a handful of data centers out west to serve the inference needs of a more sparsely populated region.

That CAICT report takes this to mean China is still falling short of the original dream to unlock the West’s AI potential. But the deeper point may be that officials know that Eastern Data, Western Compute was never quite the right model in the first place. The slogan was born before the practical geography of AI-era data centers had fully come into view, and now they have to walk it back.

Why Did They Start the Slogan?

Energy was the obvious starting point. Western China has more land, more wind and solar, more coal in some regions, cooler climates, and local governments eager to offer favorable power arrangements. In 2022, 80% of China’s compute was in the east, while 70% of renewable energy was in the west. According to Zhao Xiaofang, a chief engineer at the Chinese Academy of Sciences:

“... [moving computing facilities to the west] not only guarantees the computing power demand of the eastern digital economy development, but also cultivates a new digital industrial track in the west.”

But cheaper electricity is not destiny. EDWC seems to have been built around an incomplete theory of data center economics. The National Development and Reform Commission (NDRC) justified EDWC by saying electricity can account for 70% of a data center’s operating costs. But as ChinaTalk’s Aqib Zakaria has clarified, operating costs should not be confused with the real cost of AI infrastructure, which is dominated by construction and chips. For a 400 MW facility, he estimates electricity is only about 5% of the total three-year cost.

The energy advantages are overblown.

Nor are data centers just land plus electricity. They need electricians, HVAC technicians, network engineers, hardware staff, cooling specialists, emergency vendors, and experienced facilities managers — a skilled labor ecosystem that is already unevenly distributed. According to a 2021 Data Center Talent Development Report (2021中国数据中心人才发展报告), data center talent was concentrated in North China, East China, and South China, which together accounted for 87.5% of the total. In the US, these are good middle- and upper-middle-income technical jobs, and hyperscalers are already spending heavily to train more of them because the labor is scarce (e.g. Meta’s $115 million training program). That should make us skeptical that remote western hubs can simply conjure the same operations ecosystem found around Beijing or Shanghai.

China’s land system also makes it easier to build data centers where convenient compared to the US, as we covered here.

The meta is not “go as far west as possible.” It is exurban. Beijing spills into Hebei; The Yangtze River Delta spills into nearby Jiangsu, Zhejiang, and Anhui; The Pearl River Delta spills into cheaper Guangdong hinterlands. Perhaps the slogan should not be “Eastern Data, Western Compute,” but “Eastern Data, 45 Minutes Outside the City Compute.”

This is literally validated by how the NDRC, China’s top economic planning body, characterized the EDWC buildout back in 2022. It mapped out 8 “nexus points” 枢纽节点 for data center construction, three of which are megacity hinterlands in the East:

The 8 nexus points are:

• The Beijing-Tianjin-Hebei nexus, covering the capital region;

• The Yangtze River Delta nexus, covering Shanghai, Hangzhou, Suzhou, and many other cities;

• The Guangdong-Hong Kong-Macau nexus, covering the two Special Economic Zones as well as Guangzhou and Shenzhen;

• The Chengdu-Chongqing nexus, covering Western China’s two biggest cities — more than 50 million people in total;

• Guizhou;

• Gansu;

• Ningxia;

• And Inner Mongolia.

Only four of these can reasonably be framed as remote. From the start, EDWC was meant to augment the geography of China’s digital economy, rather than transform it. Even perfect top-down coordination would not be able to outrun basic market logic — and Beijing’s coordination of provinces and regions is far from perfect. A March 2026 report by Communications Weekly 通信产业报, a newspaper directly under China’s Ministry of Industry and Information Technology, identified three “bottlenecks” for the continued development of EDWC. These are quite obvious:

1. Coordination issues: the report notes that “..some western hub nodes still consist primarily of general-purpose computing power,” rather than specifically servicing AI training or inference;

2. Latency: having data cross hundreds of miles of transmission lines introduces inefficiency, so “many eastern enterprises balk at migrating their computing workloads westward.”

3. Hardware chokepoints and talent shortages: simply put, racks are useless without chips, and not all compute is equal. Semiconductor export controls remain the main hurdle preventing China from realizing its hyperscaler dreams. The report also notes that China has a “shortage of high-end computing talent and operations/management personnel,” resulting in insufficient tapping of western compute resources.

On Hope

A more troubling possibility is not just that EDWC misdescribes where China’s compute actually goes, but rather that the slogan gives poorer interior and western provinces a seductive theory of development that sends them further into debt.

At its onset, the NDRC explicitly framed the project as a way to narrow regional digital-economy gaps, arguing that data flows from east to west would help pull capital, talent, technology, and industry westward. As we have explained above, that is not quite what happened.

Instead, China Economic Weekly 中国经济周刊 found in February 2025 that regions from Ningxia to Xinjiang, Guizhou, and Inner Mongolia rushed to build intelligent-compute parks, sometimes treating them with the same inflated, Field of Dreams logic they used for real estate: “build it, and they will come.”

Science and Technology Daily 科技日报, another state-run newspaper, described one western city’s intelligent-compute center, built more than 20 kilometers outside town to take advantage of cool weather, cheap land, and clean energy. Its rack-up rate was reportedly below 50%; even the servers already installed had actual utilization below 30%; meanwhile, annual operating costs exceeded RMB 30 million (around $4.44 million). The same article said China had nearly 150 operating intelligent-compute-center projects by late 2024. Nearly 400 more were either under construction or being planned, creating “blindly constructed systems divorced from actual needs” (脱离实际需求的盲目建设).

That is the injustice buried inside the slogan of “Eastern Data, Western Compute.” It promised poorer provinces that they could be more than resource suppliers; they could become meaningful contributors to the AI economy and meaningful beneficiaries from its gains. Perhaps some will — the likes of Guizhou or Inner Mongolia. But many may discover that their data centers are the next installment of ghost buildings rather than tickets into the AI boom.

We built an interactive microsite for this story! Check it out here. It is pretty cool.

• We’re hiring for video talent! Learn more here.

• We’re running an AI + nukes contest with a $20k prize pool. Deadline is July 24th.

• We’re always looking for new contributors to the substack! Learn more here.

A guest post by Mieke Eoyang is a visiting professor at Carnegie Mellon University’s Institute of Security and Technology (CMIST), and the former Deputy Assistant Secretary of Defense for Cyber Policy.

As frontier AI labs race to release the next generation of models that meet or exceed the capabilities of Anthropic’s Mythos, we need to contend with what these powerful tools mean for the world. These models, whether Anthropic’s Mythos/Fable, Open AI’s GPT 5.6, or Z.ai’s GLM 5.2, the latest generation of models have demonstrated tremendous capability to identify previously unknown cyber vulnerabilities.

Though most of the popular focus has been on new hacking vectors, these models also give developers and manufacturers a way to address their tech debt through an avalanche of vulnerability identification and patching. On the other side of the “Patch Apocalypse”, Mythos-class models should change the way we think about cyber risk from Chinese-made tech products, and supply chain risk generally.

The conventional wisdom is that Chinese- made technology products pose a higher national security risk. The fears break down into two scenarios: a) the Chinese government installs secret back doors into the products allowing them to spy or brick products at their time and place of choosing, or b) the superior knowledge of how those products are built will allow them to build or later-install such back doors at their time and place of choosing.

Given the ability of these Mythos-class models to identify the vulnerabilities, could a Chinese developer really create a hidden back door in a product that’s released into the wild that would not be identified by the later scrutiny of these models? Creating a

hidden back door is not easy. Even in the pre-LLM days, attempting to prevent a backdoor from discovery was difficult. For example, in 2013 researchers identified flaws in an encryption algorithm proposed by the NSA. NIST quietly began recommending against using the encryption standard in the wake of the Snowden disclosures, and the resulting uproar led to the development of PGP. The resulting reputational hit on trust and subsequent products haunts the NSA to this day.

These Mythos class models supercharge the discovery process, dramatically increasing the chances that defenders would be able to find an engineered-in back door before the attacker could make use of their hidden portal. As one prominent NSA-er said to me once: Conducting cyber operations will break your heart when the night before you go to execute, your target will have patched your way in. With these models, attackers can’t count on retaining access from unboxing to exploitation, as the speed of identification and patching increases. [Jordan’s take: it’s still unclear whether Mythos 2027 can create new exploits Mythos 2026 wouldn’t be able to find. Perhaps the calculus is different if you’re still regularly applying Chinese software patches.]

Even if governments compel or coerce manufacturers into providing that access, doing so would have clear reputational consequences to the ongoing competitiveness of the company. Once a Mythos-class model finds that vulnerability, if a manufacturer or supplier refuses to address it to maintain government access, the trust in that product and company will be gone forever. Customers will want assurances that a company has a transparent and open process for identifying and patching discovered vulnerabilities found by anyone. Indeed, for enterprises, running Mythos-class models against their own technology stack should become standard practice for risk management against cyber attacks. Likewise, Mythos-class models’ ability to find vulnerabilities in a wide range of technology products, undermines the argument that the original manufacturer’s knowledge of the product gives it superior ability to exploit a vulnerability later.

Chinese products, like US products, and European products, all have vulnerabilities due to the decades and decades of software and hardware development that didn’t prioritize security. And even with the best, most security conscious engineering, there will always be some vulnerability left slightly ajar some unforeseen vulnerability. If these models enable you to identify (and address) any vulnerability that might be out there, then what is the difference, from a security perspective, of Chinese-origin products, relative to US ones?

The US will have to rethink whether security-based import bans on Chinese products, like EVs, drones, and routers, will continue to make sense in the wake of Mythos-class models. Companies around the world will be able to test the security of US and Chinese products for potential vulnerabilities. If a company has just as much insight into the security of a Chinese product as a European or US one, then it may return to making decisions on cost, performance, and availability, rather than national origin of the manufacturer.

[Jordan: many of these bans justified on paper for cyber reasons double as industrial policy programs. The US should try to nurture a manufacturing industry that isn’t reliant on China for strategic industries, and the powers of the FCC + BIS are more effective than tariffs or industrial subsidies on protecting industries in the current moment. Even if drones, humanoids, telecom equipment or routers have verifiably unhackable software, the military implications of not having an industrial base you can scale up in a crisis, or just the leverage in non-wartime negotiations from an adversary having the ability to threaten to cut off supply, are a big deal.]

As my CMU colleagues have written, empowering America’s AI revolution requires a more careful examination of cyber risk in the subcomponents that power the data centers that enable frontier lab development and model usage. If Mythos-class models give vendors the ability to review the vulnerabilities of foreign-made battery management systems or other smart devices that manage power flows, then they can review devices on producer-by-producer basis, rather than banning all Chinese-made products.

That doesn’t mean there should be a free-for-all on installing Chinese products. There are still adjustments that should be made to the Chinese cyber ecosystem as a result of the deployment of these models in the months to come. As Mythos-class models change our thinking about latent cyber vulnerability and supply chain risk, it should also change our thinking about what security and trust look like. In this post-Mythos world, China is going to need to change its approach to cybersecurity to remain globally relevant as a technology producer.

Once Mythos-class models are widely available, the question is not whether a product has a vulnerability, but what has it done to address the vulnerabilities that have been discovered. China’s discouraging of a robust cybersecurity ecosystem, as well as its requirements to disclose vulnerabilities to the government first run contrary to a world where customer trust is based in a manufacturer’s maintaining a transparent process for ingesting and repairing identified vulnerability. Customers and independent researchers alike will be finding these flaws as a result of their own analysis. Vulnerability discovery will take place not just between US and Chinese researchers, but also from users in India, Europe, and throughout the global south. Remaining globally competitive will mean addressing the security concerns of everyone.

US companies are well attuned to this dynamic, having developed bug bounties, vulnerability disclosure policies, and regular cycles of security updates. Chinese companies have not yet engaged in this type of public-facing security architecture. But as Mythos-class models are used to identify vulnerability, customers will demand that Chinese developers also address flaws and a failure to do so in an open way will undermine trust, ultimately resulting in a lack of market share.

Likewise, the US will have to adjust its policy posture. Going forward, broad import bans based on fears of latent exploitation and national origin–as the US FCC has proposed–will drive up costs to companies and consumers for no appreciable benefit to security. Mythos-class models enable a more sophisticated analysis of supply chain risk than just national origins of technology products. And there are certainly benefits that come from a more diversified technology supply chain, or a reinvigoration of a domestic supplier base. It just means that we need to be more specific about how and when to wave the security rationale against a particular technology type. A more sophisticated analysis will allow suppliers more time to evaluate the need for alternative sources for their subcomponents, to balance risk and benefit in a more nuanced way, and ideally deliver best value to their customers.

While we do a ton of excellent written and audio content that reaches hundreds of thousands each week, we haven’t quite cracked video yet. Our channel growth on YouTube is a little embarrassing, up to just 15k subs after years of posting.

We’d love that to change.

We’re open to hiring part time, full time, or working with a production studio. The goal is to make better use of the ChinaTalk interviews for video as well as potentially producing some more 10-20 minute shows based off the research we do for articles. Some channels we admire include Search Party, Perun, Asianometry, Dwarkesh, and Polymatter.

Here are some documentaries we’ve done on peptides and Taiwan’s war on renewables.

Perhaps you’re the right partner for us? Fill out this form here if you’d like to start a conversation!

To understand this foundational piece of the semiconductor ecosystem, I’m joined by ChinaTalk regular and Chip War author Chris Miller, ChinaTalk analyst Aqib Zakaria, and Carl Jackson, co-founder and managing director of SSoT Group. Carl has spent more than 25 years at the center of the specialty gas industry, working across Asia, North America, and Europe.

We’ll do a 101 on the chemicals behind the chips, see how China built a world-class semiconductor gas industry in just 15 years, and talk about vulnerabilites in the global chemical supply chain.

Plus, why Taiwan may be the worst place in the world to host a semiconductor industry and the existential supply-chain risks that come with it and, the surprising appeal of a career in industrial gases.

Listen now on your favorite podcast app.

Chris Miller: Carl, to start, tell us about the role of helium in the chipmaking process and why the Hormuz shutdown has been so disruptive to the industry.

Carl Jackson: On the application side, helium is used by every semiconductor manufacturer in every fab in every location in the world, and it’s primarily used for cooling. A lot of these semiconductor manufacturing processes are quite violent, they’re quite exothermic, even though it looks very calm and silent from the outside. There’s a lot of cooling required to keep the manufacturing process at reasonable and workable temperatures, and that’s the role of helium mainly in these fabs.

In terms of why the Strait of Hormuz issue has been so significant — the Ras Laffan Qatar helium production facility now produces about 15% of the world’s capacity. It was an overnight closure of the tap, essentially, of 15% of the world’s capacity that generally operates at or around equal supply between production and demand.

This is not an example of a molecule that’s got a huge overcapacity that’s very easy to just take up slack from alternative sources and deliver to a fab. It’s also a very difficult molecule to deliver. You need to move this around in the world’s most expensive thermos flask at -269°C. That whole logistics supply chain is extremely complicated as well. Just turning off the ability to move those assets, plus the production that comes from that site in Qatar is a pretty major disruption.

Chris Miller: That’s just one of dozens of types of gases that are used in the chipmaking process. Walk us through at a super high level — what are the gases you need to make a chip?

The Gas Chemistry Behind Semiconductor Manufacturing

Carl Jackson: The role that gases play generally in chip manufacturing is often misunderstood. People don’t realize that semiconductors are basically made using gas.

You look at your phone, and it’s made of plastic and metal. People don’t see the inputs required to make that in these fabs — which is all gas. There are probably 120-odd different chemicals that go into a fab from different suppliers. There are probably 60 unique chemicals that go in, and they’ll all basically go in in gaseous form.

It’s a huge chemistry toolkit that’s been developed and is required to build all chips. This is not just future advanced technology. This is every chip everywhere that needs this chemistry set.

Jordan Schneider: You mentioned cooling as one of the applications. What are other things that the gases do?

Carl Jackson: They build with gases. Everybody’s seen a picture of a silicon wafer — that ultra-high-purity silicon that’s the most perfect structure in the world. That gets delivered as a building block. I always talk about this in terms of a skyscraper to help people imagine what it looks like. This building block arrives as your concrete foundation. Then every floor, every elevator, every piece of wiring that builds the next 300 floors of that skyscraper — which is essentially what a chip looks like — is built using gases.

Silicon gets deposited via a gas called silane. These deposition steps happen over and over again, creating a repetition in how you build these structures. You then need to knock some of that structure out using fluorine-based gases to etch and remove portions. You need doping gases to change the electrical properties of some layers. You need to build the transistors in there, which requires different types of gases and processes.

It’s an extremely complex, almost unbelievable manufacturing process, all made possible by the use of gases. This underlayer — this supply chain and chemistry set — is not very well understood as an input into this industry.

Chris Miller: Can you talk about the categories of gas? As I understand, some are pretty straightforward, produced in bulk, while others are highly specialized to the chip industry. Walk us through the key categories we should think about.

Carl Jackson: I would split them into two categories. Any major fab needs a huge amount of what we call bulk gas. These would be the inerting gases — the gases used to create the clean environments needed to make chips. That would be nitrogen, argon — things that don’t react or essentially don’t do anything in the process apart from keeping things clean.

Then you’ve got the specialty gases. The bulk gases would typically be delivered by gas companies with an on-site air separation system. The volumes required are massive, so they dictate that you would need a plant built next to your fab that would deliver these gases 24/7. We’re talking about tens of millions of dollars of investment by the gas company.

Once they’re built — and they get built wherever the fab goes — there’s a relationship with the gas company that extends over 15 to 20 years because nobody builds a competing air separation unit next to your fab. There’s a relationship anchored around the bulk gases.

On top of that, you’ve got 60 or 70 different process gases that are typically delivered in a range of different containers. You’ve seen these larger containers rolling down the highway, cylinders, and packages smaller than cylinders. They deliver the rest of these specialized gases for deposition (putting new material onto a wafer), ion implantation (which changes the electrical characteristics), and etching (which shapes and helps create the devices you need on silicon).

The toolkit’s very big. A typical fab will be taking these materials from every geography, probably with the exception of China. If you want to talk about Taiwan as one example, they’ll be receiving their bulk gases from the on-site supplier, but the balance of materials they need to run their fabs will probably come from at least another 4 or 5 countries. It’s extremely complicated logistics and supply chains to get that balance of the other 60 to 70 materials needed.

The Quest for Purity

Chris Miller: In terms of the purity level of these gases — we know that silicon wafers are extraordinarily pure silicon. How do we think about the level of purity in each of these types of gases? Is it different between bulk versus specialty, or what’s the way to distinguish the level required?

Carl Jackson: Well, one commonality is that it’s always increasing. There’s a drive on the customer side — TSMC, Samsung, or GlobalFoundries, etc. These companies run very sophisticated SPC and quality assurance systems where they’re consistently tracking everything that comes into their fab.

Obviously, as node sizes decrease and the complexity of these devices increases, they want to ensure that every input into the manufacturing process has higher and higher purity. At the extreme is the silicon wafer at 11 nines, which is industry speak for almost impossible. Only two companies in the world currently can achieve that.

The gases range from parts per million, which is now completely standard. A one-part-per-million impurity is really just the basic requirement to be a supplier of any material into this market. Some gases are now reaching parts per billion, and some even get into parts per trillion.

Parts per trillion is very difficult to imagine. It’s basically a heartbeat in 32,000 years. For the entirety of anything that’s been known to be living on this planet, it’s a heartbeat within that scale of time. That’s the level of impurity we’re talking about. That’s how to think about how difficult this is.

Chris Miller: I’m just adding up zeros in my head, but that’s dramatically more pure than 11 nines for a silicon wafer. Is that right?

Carl Jackson: Yes, and that’s the trajectory. Where it all stops, I don’t know. It’s obviously leveled out. There was a huge push in the early days of semiconductors where big advances were made in the way materials are purified, packaged, delivered, and used. It’s tailed off a little bit, but they still continue trying to improve the purity of all these materials to impossible levels — levels where, really, the difficulty is measuring the impurity, not making the product.

Jordan Schneider: How do you measure parts per trillion?

Carl Jackson: With extremely expensive analytical instruments that hardly anybody knows how to use properly. It gets into ultra-specialist and extremely expensive equipment that’s so sensitive it makes the whole job very difficult. Ultimately, that will be the limitation. There’s a cost implication of increasing the purity of these products, but ultimately it’s the metrology around it that will be cost-prohibitive.

Jordan Schneider: It just seems like when you’re that deep, there’s some Schrödinger’s principle going on where it’s the measuring tool’s fault more than your poor gas refiner.

Chris Miller: How do you get these gases? It’s different for different types of gases, but if you want a gas purified to 1 part per trillion level, how does that happen?

Carl Jackson: Well, that’s the extreme, Chris. But let me give you one example. The other thing to say is quite a lot of this chemistry set is also used in industrial chemistry.

One example is HF, or hydrofluoric acid, which is made in the hundreds of thousands of tons. That material starts life in a mine — it starts life underground as a material called fluorspar. It’s literally dug out of the earth at concentrations that you would never recognize as suitable for a semiconductor fab.

It’s then made into an industrial-grade HF. At that point, it isn’t destined for a semiconductor fab. That’s used, as I say, on a scale of hundreds of thousands of tons industrially.

Then the semiconductor guys get involved. They’ll take a small part of that volume and start to process it using very different types of purification systems that have been developed around the requirements for semiconductors. You start with a typical industrial-grade HF at 3 nines purity, and that’s good enough for almost everything required on Earth that uses HF — for pharmaceuticals, for all of the different industrial applications, but not for semiconductors.

That material then needs to go from 3 nines up to 6, 7 or 8 nines, depending on the process. That’s where the semiconductor supply chain, the people involved in this, really add their magic. It’s how you get and maintain and then ship and have that high-purity product delivered all the way to the wafer through this supply chain. At that point, it becomes very specialized, and you need to rely on a very small number of people in the world who can do that.

Chris Miller: If you get a jar of HF at 3 nines and want it to be 6 nines, what do you physically do to that jar to get to that level?

Carl Jackson: In that case, you put it through a distillation system. It’s a little bit like the petrochemical distillation system. Distillation is distillation, but not all distillation systems can be set up to handle HF and to get to that level of purity required for these semiconductor customers.

It’s advanced distillation with all the steps built into it to take out impurities that you wouldn’t normally need to remove. It might be distillation plus another couple of steps that, as I say, wouldn’t be necessary in industrial processes.

Aqib Zakaria: Is the jump from 3.9 to 6.9 significantly easier than it is from 6.9 to 9.9? We think of the semiconductor industry going from 7 nanometer to 5 nanometer as extremely difficult, a lot more difficult than from 28 to 14. Can you give a picture of how much money it takes to get to the next step and how much time it takes?

Carl Jackson: It depends on the product. Some materials are easier than others to purify. The next 9 — from where we’re at now — is extraordinarily difficult to produce and deliver. Almost nobody is interested in investing to get to that next 9, because there’s no demonstrable evidence that it’s actually required in the process.

There’s an interesting dynamic where the drive from the customer side to get even purer material, consistently improving that purity, isn’t actually required to make the devices. There’s a drive based around math and SPC and quality data that doesn’t necessarily triangulate with what’s needed on the process side. If you understand that and you’re on the production side, then you’re really not motivated to try and increase your whole production system and your supply chain to that next 9. It might give you some competitive advantage, but ultimately nobody’s going to pay for it unless it’s really needed on the manufacturing side.

Aqib Zakaria: You were saying earlier that at some point the semiconductor guys get involved and say, “Hey, I need an extra 9, I’m willing to pay this much money for it.” In that scenario, is it TSMC saying we need better gas, or is it the deposition toolmaker where the gas is used saying we need better gas? What’s the player here?

Carl Jackson: The industry talks about recipes — when these gases are used, they’re used in a recipe just like a cooking recipe. The recipe dictates what materials you need to make a certain type of device at a certain time in the manufacturing process.

Those recipes will typically be developed by the toolmaker. They’ll tell you what chemistry you need and at what purity. They’ll guarantee that you can make a particular device. They say, “Okay, you’ll need five materials and they all need to meet this purity. If you provide that into this tool, you can make this device.” That’s what TSMC will receive.

They will then improve that recipe. They’ve got thousands of people working on optimization of their processes and their tools. There’s a juxtaposition — always between improving, optimizing, but not changing anything. That’s a force that’s always fighting here, but they typically will have people that are optimizing and improving. Part of that will be, “Okay, we need purer material.”

Sometimes it’s just a blind requirement — it just needs to be purer. If you challenge it, they’ll say, “We’re the customer, we’re telling you that it needs to be purer.” Why? “It doesn’t matter. If we’re going to keep buying it from you, it needs to be purer.”

At the toolmaker level, they will typically put performance guarantees around a particular recipe set and a particular purity for that recipe set that will work. When it gets into the hands of the guys who are actually making the devices, then it goes into a different world. Some customers don’t change. Some customers are pretty active in terms of demanding better materials.

But it’s slowing down. The parts per trillion thing is really extreme — it’s an outlier. I’m certain that all of this standard chemistry set isn’t going to get there. Some specialist materials may be for some ultra-specialist devices, but generally people are not going to be demanding that. It will stay in the parts per million, parts per billion.

People forget that parts per million is an unbelievably difficult thing to deliver for a material that is — for example, HF will try to eat the cylinder that it’s being delivered in.

Lethal as the Specification

Chris Miller: Could we dig into that? One of the fascinating parts of this is that it seems like half the materials we’re talking about are either toxic or explosive or both. Why is it that you can’t make chips without all these seemingly awful chemicals? And what challenges does this impose in terms of transporting them?

Carl Jackson: Almost everything in this toolkit is lethal. It either poisons you, explodes instantly upon contact with air, or can kill you quickly or slowly. But lethal here is not a bug — it’s the specification.

If you’ve been to a fab or seen pictures or videos, you see this very quiet, controlled, seemingly delicate process going on to make these chips. People envisage it as delicate. But at the atomic level, it’s an extremely violent process. The violence requires the most violent chemicals and the most reactive chemistries.

This deposition and etching process that builds these layers — you’re trying to drill the equivalent of an elevator shaft down through 300 stories in one go, straighter than anything imaginable physically. You need something that’s extremely reactive in the process to allow you to do that.

Similarly, when you’re doing an ion implant to change the electrical characteristics of the silicon so that it actually works and burns the transistor, you’re firing boron, phosphorus, or arsenic — all of these lovely gases that you need to do some work. You’re firing these things in at 400 kilometers a second into the silicon wafer. Most of the molecule disintegrates when it hits this wafer. It’s like firing a bullet into a concrete block. That bullet needs to go exactly where you need it to in that concrete block — the depth, the angle, everything. Then it disintegrates and leaves something behind that does some work.

The process itself is extremely reactive and violent. That’s why you need these chemistries. Which is unfortunate because making them and delivering them safely is a little counterintuitive.

Semiconductor doping: how a chip is made (A really calm explainer of an extremely violent process.)

The chemistry set that you need, because of the difficulty of the device that you’re making, needs to be large and contain these difficult gases. To your point about how they’re delivered — it’s done in all kinds of weird and wonderful ways. It’s part of the IP of the semiconductor industry and the people that feed into it. How do you get these products to the customer?

Some products are eaten themselves on the way — you make these products, and they start consuming themselves. Some products start to consume the packages, the cylinders that they’re in, and create impurities on the way. They’ve got a very short shelf life. Some of them are so dangerous that you can’t just ship them in a cylinder. You need to ship them in some other container that, in the event there’s damage to that container, these chemicals won’t come out.

There are innovations around the handling, safety, and maintenance of purity in this business that are never really seen, but they’re a huge part of enabling how these devices are made.

Chris Miller: What’s the most dangerous chemical used in chip manufacturing?

Carl Jackson: There are so many dangerous chemicals. The one that people fear most is HF (hydrofluoric acid), largely because there have been a few incidents with it. It’s used in fairly large volumes. What makes it particularly frightening is that it doesn’t immediately kill you — it gets into your skin and then pulls the calcium out of your bones over time, essentially killing you slowly.

I’m not trying to fearmonger here. I’m painting a picture of what could happen if this industry wasn’t as well-managed as it is. In reality, there are barely any incidents these days. On the fab side, there are almost no reportable incidents in a year across every fab worldwide, which is absolutely remarkable when you consider the chemistry set involved. Similarly, in the shipping of these materials, there are barely any incidents. On the production side, there are a few, but nowhere near the amount you would have in a normal industrial process. It’s probably one of the safest businesses, given the risk profile of these materials.

But HF is definitely not a good actor — it’s a difficult one to deal with. Arsine is used on every chip. Every chip requires arsine. It’s essentially arsenic — the same molecule that was the housewife’s poison of choice in the 19th century. The same molecule that makes a transistor work is the one you used to be able to buy to slowly kill your husband. It’s the same chemistry, just doing a different job.

Silane is another dangerous one. This is silicon in gas form. People figured out a long time ago that you need to build chip structures using silicon. You get your starting wafer, but then where do you get the silicon from to build these structures? You can’t just get a matchstick-size piece of silicon and glue it on — you need to build it differently. That’s where silane comes in.

Silane’s great unless it starts to leak. If it leaks from a cylinder, the current safety protocol is actually to leave it leaking. It’s quite counterintuitive. If a valve on a cylinder fails and silane starts to come out, you literally get sand everywhere because it reacts with air — the hydrogen gets replaced with oxygen, and you end up with silicon oxide, which is sand. But if you create some kind of static or some issue that causes it to ignite, you end up with a fireball coming out of the cylinder instead of a load of sand. It’s still not very well understood why this happens, but it does.

When training someone to deal with a silane incident, the instruction is: if it’s leaking, leave it to leak. Don’t close the valve because, in closing the valve, you might create a fire, which is worse than creating sand.

As I say, it’s a big toolkit, and really none of it is very friendly. Even nitrogen — the most benign chemical used in large quantities — is really dangerous. If you’re in a room with just nitrogen... Right now the air is 79% nitrogen, which is fine because you’ve got your oxygen there. But if you start getting nitrogen concentrations any higher than that, you start having issues. Those are actually the most common safety issues you’ll see — nitrogen asphyxiation rather than problems with some of these more dangerous chemicals.

China and the Chemical Supply Chain

Aqib Zakaria: Let’s discuss these potent chemicals that are sometimes delivered at mind-warping purities. Who are the companies actually delivering them? How many are there, and what differentiates one from another?

Carl Jackson: It’s best to look at this over time because the landscape is changing rapidly — almost all of it happening in China.

If we go back 10 to 15 years before China’s Big Fund initiative to localize semiconductors and their supply chains, the market was segregated by chemistry specialty. Japan historically excelled with fluorinated gases — anything with a fluorine backbone. There are probably 15 different fluorine-based chemicals and chemistries used in semiconductors, and Japan was the go-to source for these.

The US had a pretty broad range of manufacturing capabilities since that’s where this industry originated. Twenty years ago, there was near self-sufficiency in almost everything required. While that’s somewhat depleted now at the scale and scope needed, that capability existed then.

Most of these gases were typically made by a few Japanese companies, but the majority came from the majors: Linde, Air Products, and Air Liquide. These companies that built and installed air separation units also provided all the other required chemicals. It was a self-funding, self-financing model — you’d install the air separation unit as the anchor that paid for developing all the other materials. They offered a complete portfolio to each semiconductor customer. Twenty years ago, almost everything came from one of these majors with their comprehensive portfolios.

Chris Miller: Before we discuss the present situation, I’d like to understand what the industry looked like 20 years ago. Would one Intel fab have an air separation unit from Air Liquide while a different one uses Linde? Were these company-to-company relationships or fab-to-fab?

Carl Jackson: Fabs prefer to split risk, though there’s minimal risk with air separation units now — they’re extremely reliable with installations worldwide.

For example, if TSMC were to build a new cluster of three fabs in a new location, it would typically award one ASU per major supplier. Linde might get one or two, Air Liquide might get one or two. There’s very little differentiation in what the customer receives — all provide extremely good reliability and quality, with some negotiation on pricing. No single major would win all the contracts; the split between fabs and awards depends on the customer’s situation or whether one of the majors can offer something particularly compelling. The business is basically split almost evenly among them.

Share

Chris Miller: So you mentioned that the industry is changing. Tell us what’s new in the gas industry and the role of China.

Carl Jackson: The Big Fund is really what’s driving China’s transformation at this level. Big Fund 1 kicked off around 2014, and we could see it was different from the start. Unlike other similar initiatives — the CHIPS Act, for example — the Big Fund targeted all levels of semiconductor manufacturing.

I won’t speak to areas outside my expertise, like design and tool manufacturing infrastructure, but all of that was included in the Big Fund. You’re not just getting new fabs and new IDMs making devices — you’re getting the entire supply chain. They went after everything at once.

There was a lot of money. About $120 billion has been invested in that whole infrastructure ecosystem. It was managed regionally rather than through top-down directives saying “you do this over there, and you do that over there.” While there was an overall top-down directive, implementation was managed on a regional basis.

This led to different regions competing with each other for all the different materials in this chemistry toolkit. Take one product: NF3, which is part of the fluorine family. It’s used for cleaning — it’s like the janitor of the semiconductor world, used by everybody in every process. At the time, China didn’t have any NF3 capability and needed to import it.

Different provinces with different leadership and scopes started developing supply chains for these materials in parallel. This resulted in massive overcapacity. You had new players who had never been active in producing these materials or supply chains. Initially, it was the whole story of China — low-quality materials, questions about trust in where and how things were being made. That took a long time to overcome.

They started out feeding themselves, but because of how the Big Fund was organized, they ended up with huge overcapacity. Now you’ve got China plus the rest of the world in terms of their ability to provide these 60 gases. They’ve developed the capability to make almost everything in China independently and in parallel with everybody else in the world — with a capacity that could almost feed everywhere in the world from China.

We’ve ended up in a situation that is totally different from where we were 15 years ago because of how this Big Fund has been funded and rolled out. There were really no checks and balances on whether it made sense to put this capacity in place. The attitude was: “We don’t care. It doesn’t matter. We’re doing our job.” As a result, there’s now a lot of material available that can feed and compete with this existing supply chain that previously had a very small number of suppliers.

Chris Miller: Do we see Taiwanese or South Korean firms buying gases from the Chinese supply chain?

Carl Jackson: Taiwan is 100% reliant on Chinese supply chains today — 100% reliant. Chris, you talk about missiles, but I could just change that to NF3 as an example. If the Chinese government decided to put an export restriction on NF3, then the Taiwanese fabs would shut down.

Chris Miller: Help us understand how we should think about finding alternative sources of supply. I think back to 2022 at the start of the Russia-Ukraine war — both countries were big suppliers of neon. The Azovstal steel plant, where there was a big siege at the start of the war, was itself a neon supplier. That created a crisis for the semiconductor gas industry, but it didn’t end up impacting chip production much at all. This suggests there was flexibility in the neon market. Is that unique to neon? Is it different from NF3?

Carl Jackson: Helium is probably the best example of a supply chain that’s inelastic, mainly because of the packages. The most basic package to get gases from the supplier to the consumer is a cylinder. These cylinders cost $50, and there’s no shortage — millions circulate around the world.

The helium packages I mentioned before cost $1 million each. They’re extremely complex. You’ve got 45 days to get from production to consumption before you start to lose the product. That’s pretty inelastic.

On the other extreme, take NF3, for example. Chinese domestic consumption requirement for NF3 is about 8,000 tonnes. You have one producer in one province in China making 55,000 tonnes a year. Those guys are busy recruiting international sales and marketing teams and trying to sell their NF3 internationally — into Taiwan, South Korea, the US, anywhere that can take that material. They’ve got almost zero incremental cost of selling anything above what’s required domestically.

You’ve got real extremes on either side. The Iran situation isn’t a minor blip — it’s a pretty big one. But we saw something similar with an NF3 plant in Japan in 2024. An NF3 plant blew up and took one offline. The impact that rippled around the semiconductor world was pretty small. Existing customers needed to requalify and re-establish supply chains. But for a material that’s abundant and easy to deliver because the packages are available, those supply chains can reroute and customers can take new material very quickly.

China’s Dominance and Dependency

Chris Miller: Absent the subsidy and overcapacity story, is there a reason why you’d produce more NF3 in China than elsewhere? Energy costs or some sort of comparative advantage, or is it solely a story of governments paying for it?

Carl Jackson: It’s a mixture — a blend of things. NF3 is part of this fluorine family. Fluorine comes from fluorspar, and China controls 70% of the world’s fluorspar. That whole supply chain is integrated inside of China, allowing them to do everything internally at an extremely low cost.

Some of this comes down to economies of scale. Some of it involves ignoring domestic requirements and building out huge capacity for some future need. Some decisions simply don’t make economic sense from a Western point of view.

When Linde invests in a plant, using the Western example, the amount of due diligence required around profitability, customer identification, and production ramp is like a finely polished machine. The Chinese approach sometimes doesn’t follow the same economic logic. They build it because there’s a directive to build it. If they need 10,000 units, they’ll build 25,000. It’s that kind of math — no exaggeration.

Someone in a province 1,000 kilometers away might be making that same decision because it doesn’t matter. The availability of these products internally is far more important than any return on investment for an individual plant. The result of that internationally is what we’re starting to see now.

Chris Miller: You mentioned fluorspar, which is critical for a bunch of different gases, with 70% mined in China. I understand some is mined in Mexico. Can you walk us through the process from fluorspar mining to gas? This seems like one of the key areas in terms of China’s market position.

Carl Jackson: The process literally starts with rocks dug out of the ground. fluorspar just looks like a rock. Those rocks go into an autoclave — basically a massive hot reactor. China has hundreds of thousands of tons of capacity in these reactors, while semiconductors only require a few thousand tons. It’s a tiny proportion of what’s made industrially.

You dig the rock out of the ground, send it to an industrial HF producer — there are probably 10 world-scale producers in China. That gets converted into industrial HF. As I mentioned before, the semiconductor business requirement is to siphon off a tiny proportion and make it an ultra-high-purity material.

The key point is these chemical families. HF is used to make fluorine. Fluorine is then used to make C4F6 and sometimes NF3. There’s a whole family of chemistry built off HF or fluorine that requires additional processing but ultimately comes back to that fluorspar mine.

Aqib Zakaria: I’m curious if there are any cases of it going the other way. We’re pretty accustomed to stories like rare earth minerals and now fluorspar where everything comes from China, creating great dependency. But are there any cases among these 60 or so gases where the mine isn’t in China — maybe it’s in Peru or somewhere else — and China has to import all of it? Are there cases where the dependency goes both ways?

Carl Jackson: Helium is an interesting challenge for China. From a Chinese perspective, helium represents the biggest gap and most high-value, important material for semiconductors. The reason they don’t have enough helium is purely geological. The US is extremely lucky with great geology for extracting helium, and anyone making LNG also has great capacity for making helium.

Helium typically isn’t produced on its own — it’s a co-product from something else. Until recently, nobody mined or extracted helium specifically. You extracted LNG, and helium happened to come along with it. When helium started gaining high-value, important applications like MRI and semiconductors, people became more interested in separating and producing that helium.

China has a challenge — the amounts of helium in their LNG production are extremely small. Where they produce it is basically all in the wrong place, distributed around a couple hundred different small sites. Their LNG infrastructure isn’t conducive to making helium, so they’re still a net importer of helium.

That’s being fixed. In the Chinese way, they’ve figured out ways to produce their own helium despite these restrictions. But it’s going to take time to develop the infrastructure they need to stop being a net importer. At the moment, that’s probably their biggest vulnerability for the semiconductor supply chain. But as I said, it will get fixed — they’re busy trying to figure that out in really smart ways.

Taiwan and Korea are in a similar position where the geology doesn’t allow economically for helium production. They will always be 100% reliant on helium imports, plus almost every other chemical needed to run these fabs forever, unless they do something they haven’t yet done in 30 years. Their supply chain vulnerabilities will remain when China’s fully fitted out.

The Big Fund vs the CHIPs Act

Jordan Schneider: Carl, given your vantage point, how does the comparison between the Big Fund and America’s CHIPS Act look to you?

Carl Jackson: At the fab level — reshoring the actual device manufacturing — they look somewhat similar. The CHIPS Act is about reshoring the manufacturing of semiconductors, typically with existing manufacturers. The CHIPS Act successfully got TSMC to make a new investment in Arizona and got Samsung to make a new investment in the US. At that level, it’s worked because there’s going to be more domestic production of semiconductors.

Where it stops short, in my view, is anything underneath that layer required to support that domestic production. I haven’t seen anything at all yet about how the supply chain and the security of the supply chain and the chemistry supply chain that feed those fabs and any new fabs. All of these domestic production requirements are being announced, but without any of the supply chain underneath. If nothing changes from the US perspective, then there will be total reliance on importing almost every chemical required to build those new capacities.

On the Chinese side, they’ve done it differently. As we mentioned before, they’ve set out to put the entire ecosystem in all at once. It’s not just new fabs built by existing suppliers — it’s new companies entirely building new fabs, creating new technology, and underneath that is the full and complete supply chain to put all the materials required to build those components in, fully domesticated.

The scope and scale of the Big Fund are much larger and, while not perfect by any stretch, it’s certainly more integrated and forward-thinking than CHIPS in its current form, which looks like it’s either ignored, not quite gotten to yet, or isn’t interested in the supply chain vulnerabilities underneath those new fabs being installed.

Aqib Zakaria: Everything you said is right. For me, history keeps rhyming. With fluorspar — every new supply chain I look at tells the same story. You could just copy and paste, swapping out words. Gallium, rare earths — there are just so many examples.

It’s fascinating and needs to be better studied. Western economics makes you prioritize things that are later down the value chain, even though they’re dependent on things that are way more upstream. For one reason or another, China has become dominant in those upstream sectors, whether it’s because of state policy or because of their perceptions of secure supply chains.

That sometimes poses a threat for us in the United States or the West. I might be making the $1,000 iPhone, but it’s still dependent on the $1 gas or $1 mineral coming from China, which leads to some pretty disruptive or strange geopolitical topics. I’m curious what you think about this, Chris.

Chris Miller: Carl, if a chip costs $100 to make, what share of that is gases?

Carl Jackson: The numbers typically talked about are around 10% — that’s the bill of materials that contributes to the value of a chip. That’s all 60 gases combined. It depends on the chip.

I did some work on this maybe 10 years ago when we were trying to figure out the value of the products that we were selling into an iPhone. For each iPhone manufactured, how much value of gas is in it? What’s the weight of the gas in there? Because you’re trying to explain that this is made with gas, and people say, “It can’t be. It’s solid. What are you talking about?” We tried to figure out how we could contextualize it a little bit more. Then we asked how much is in a Tesla? It’s about 10%. But again, it depends on the structures and devices you’re trying to make. That’s one way to look at it.

The other way to look at it is — which one of those 60 can you live without and still manufacture your products? The answer is zero. None of them. You need all of them — exactly when you need them, exactly at the spec that you need them.

The vulnerability is there. It’s not about how much you need or what the breakdown is or what it’s worth. The question is, what do you do if one of those is missing? How do you run your fab?

It looks even worse because if you break the 60% down into 10%, you might have a line item in your materials that’s $100,000 a year for one gas that is barely used. If you’re missing $100,000 worth of gas, you can’t run your fab. The output cost — the opportunity cost or the downside — is huge in relative terms to the input gases. It’s more important that you get them there and in spec than what you pay for them ultimately.

Chris Miller: Given potential supply disruptions, one might expect chip companies to stockpile gases on site. You mentioned certain gases can’t be stored because they corrode their containers. How do chip companies approach on-site inventory management?

Carl Jackson: Credit where it’s due to the gas companies — typically, all materials arrive on site with barely any safety stock. Some materials physically can’t be stockpiled because fabs don’t have the necessary permits to store large quantities.

Take HF, for example. You can’t just store a few tons of HF in the fab as a precaution against supply chain disruptions. You can only order and store what you need for the next month of production. There’s no stockpiling capacity.

These supply chains work miraculously. Despite their complexity, they deliver all these materials and molecules through customers’ gates exactly when needed. But this reliability has become part of the problem — if there’s any real disruption, there’s no facility in place to cope with it well, particularly when materials come from overseas.

This is where Chinese infrastructure will demonstrate its power. It’s not just about cost. They have multiple domestic options, all at the right capacity, all manufactured domestically. While it might be a few thousand kilometers away, there’s no import-export barrier requiring ships, planes, or export restrictions that could disrupt supply. That’s the real vulnerability in this business.

Chris Miller: That’s pretty ominous — the implications are obvious.

Carl Jackson: Taiwan — and I say this as someone who lives there and loves the country — is arguably the single worst location you could pick for semiconductor fabs. It has no natural resources, water is an issue, and there are significant geopolitical risks. They had a brilliant industry pioneer in Morris Chang, but if you were to look at a globe and choose where to drop a semiconductor industry, Taiwan probably wouldn’t make the list. It’s in seismic zone 4 — the earth moves regularly there. Not ideal for making these kinds of devices. But it is what it is, and they face existential supply chain risks.

The R&D Death Spiral

Aqib Zakaria: You mentioned 10 to 15-year contracts for gas suppliers. That doesn’t seem particularly healthy. Could pure gas become the next bottleneck for future nodes? Should we worry that it won’t happen because the investment isn’t there?

Carl Jackson: Fifteen years has become the norm, based on a sensible premise: these plants are incredibly reliable. The 15-year operational lifetime of air separation plants has been proven repeatedly.

Some of these air separation units cost $50 to $60 million or more. No gas supplier will invest in one with just a one-year contract — it’s pointless. You can’t move it. The piping required to transport gases from the separation unit to the fab runs underground. All the infrastructure is permanent and built-in, so there’s no reason not to establish a long-term relationship.

The problems emerge on the special gases side — what we call the “balance of plant.” These typically involve one-year contracts, sometimes longer. They’re often rolling contracts –- you get a year, and if nothing goes wrong, you might get another year. If you improve quality, you might secure two years — but definitely not ten.

The race to the bottom on pricing is in full force for almost all these materials. They’re fully commoditized and extremely difficult to differentiate when customers demand identical products from every supplier. What can you do when they want silane at six nines purity? Deliver it in a different color bottle?

This standardization benefits buyers who see no difference except pricing and supplier reputation. It’s increasingly difficult to remain an unchallenged incumbent. This race to the bottom means major companies like Air Liquide, Linde, and Merck have less money for R&D and are less prepared for the next materials required by OEMs developing new recipes in Santa Clara or Tohoku. Those materials — and their supply chains — may not be ready.

Fifteen years ago, we had well-integrated programs like Tech Set. During my time at Linde, we developed many materials with a high risk of never being used. We were happy to invest in R&D for products we thought might go into next-generation devices — that was part of the program.

But when your entire portfolio faces permanent cost pressure that keeps worsening, you enter a death spiral that makes R&D economics extremely difficult. R&D becomes the first casualty. You won’t maintain ten different research programs when your entire portfolio faces pricing pressure. Eventually, you decide to compete purely on scale and price rather than frontier innovation.

This shift has accelerated over the past decade and will likely continue. We’ll see a growing disconnect where new material supply chains take longer to develop.

Chris Miller: Can you explain what an air separation unit is and how it works?

Carl Jackson: You get different flavors of air separation units, but basically all the components in air — nitrogen, oxygen, and in some cases argon and the rare gases like neon, krypton, and xenon — exist at various concentrations. The only way to economically get the amount of nitrogen or oxygen that a fab needs is to separate the air right next to the fab.

You’ll have seen these things without recognizing them. They’re tall boxes, about 30 to 40 meters high, sitting close to fabs. They have huge compressors that suck air from the atmosphere and cryogenically separate it. They cool the air down to liquefy it, then separate different components at their various boiling points. These purified gases are delivered straight to the customer.

Semiconductor customers need massive amounts of nitrogen for inertia and environmental control. It would be physically impossible to ship that quantity in — you have to make it on site. The only way to produce those volumes is through air separation.

It’s an unbelievable thing to do, but it’s now so commonplace that separating air cryogenically and feeding it into a fab 24/7 for 15 years with zero interruption and no quality problems. It’s just Tuesday at a gas company.

Chris Miller: You’ve mentioned that the Japanese are experts in fluorinated gases and lead in many semiconductor manufacturing materials. But the three key companies that have historically provided gases — Air Products, Air Liquide, and Linde — are US or European. Why isn’t there a big Japanese gas provider?

Carl Jackson: There are large Japanese gas companies, but they haven’t achieved the scale and scope of the big three you mentioned. Some have joint ventures and international operations, but they’ve maintained a primarily domestic focus — Japanese companies set up to serve the Japanese market.

There’s also a portfolio mix issue. The companies you mentioned are very strong in air separation, which requires building a global footprint. If you need an air separation unit somewhere, you must physically build it there rather than making it in one place and shipping the material. This naturally creates a global presence.

The Japanese companies aren’t as historically strong in air separation. Like most complicated questions, there’s no single reason why there isn’t a Japanese equivalent of Linde. However, if you go to Japan, there are dominant and very large-scale industrial gas and special gas manufacturers with a high degree of self-sufficiency, especially compared to somewhere like Taiwan.

Chris Miller: Let’s turn to the supply chain question. Do companies know if they’re getting NF3 into their fabs where the fluorine is coming from?

Carl Jackson: In most cases, no. There are probably outlying examples, maybe from companies like TSMC, which is at the extreme end of their quality systems and supply chain tracking. They may know one or two steps, but typically, the purchaser of those materials at a fab will be interested in the specification, how much you can provide, and what the price is.

They won’t go back one, two, or even sometimes six steps in these manufacturing processes to understand where their supply chain risks converge. Sometimes it matters, sometimes it doesn’t. The reason it mostly wouldn’t matter is that the volumes required for most of these gases are so small compared with the industrial volumes being made that there’s not a huge amount of risk.

Where people aren’t looking is where these materials are coming from and whether there are any issues in terms of those countries being able to turn off those taps via export controls. It’s more the geopolitical risks than the actual supply chain that may converge into a single mine somewhere for 10 or 12 of those materials. The mine is so big, the industrial base is so big that it isn’t a question of running out of capacity. It’s if something geopolitically happens and I’m unable to access this, then what happens? That’s the more important question.

Chris Miller: Seems like a highly relevant question to be asking.

Career Paths in Industrial Gases

Jordan Schneider: Carl, what are these people like? Who gets into chemicals? How would you distinguish them from other wonderful people?

Carl Jackson: People in this business are wonderful. Everyone’s got a strange story coming into industrial gases. No one’s going to school saying, “I want to be in the industrial gas business.” There’s a wide mixture of engineers and chemical engineers that end up on a project that may involve some gases and get interested in production methods or metrology or some element to do with this supply chain, and then they stick around for a long time.

The number of people I’ve met in my career who have been involved in this for 20, 30, 40 years is amazing. People are not in and out of this. They come in, and they stay. Don’t ask me why. I’m not sure.

Jordan Schneider: I have to ask you why. What’s the appeal? We’ve got a lot of kids listening. Why should they get into industrial gases, Carl?

Carl Jackson: As I said before, if you start to understand what these materials are required for, the scope of the different chemistries underneath this and the number of different projects that you can work on is almost limitless. You’re at the frontier of everything, if you want to look at it that way. If these chemistries don’t fulfill the requirements of the next device that’s being manufactured, there’s no device.

You can look at it as, “Well, it’s just a gas. Gas is a gas. Who’s interested in that?” If you want to frame it in terms of enabling everything, then you can. You just need a bit of imagination of how you talk about this industry, this business. If it’s logistics of moving a cylinder around the world — yawn. But what’s in that cylinder? How’s it been made? What’s it going to get used for and by whom? And how has that all happened? And where’s it going to go in the future? From my point of view, that’s pretty interesting stuff to be working on.

That’s what keeps me going. There are challenges of taming these materials, there are challenges of developing new ones. The customers that you’re selling into are changing all the time. Nothing’s still in this business. Nothing is the same from day to day. But on the outside, it looks pretty boring. This is like, “Wow, I see a gas truck going up the highway now and again. I’m not sure what’s in it, but whatever.” But when you get on the inside, it paints a different picture.

Jordan Schneider: Carl, thank you so much for giving us all this insight as well as an inspirational closure. I did not think that we were going to end this podcast with a bit of industrial gas romance. But here we are. Thank you so much for being a part of ChinaTalk.

20 Takeaways from H1 2026

1. Xi is willing to purge trusted allies, not just rivals.

Zhang Youxia was purged in January, and no one knows why exactly. But we do know that the PLA’s command structure has never had less experience with actual warfare than it does right now.

From Xi’s Military Meltdown with Jon Czin:

Jordan Schneider: What does this mean for Taiwan contingencies?

Jon Czin: I’ve actually been turning this question on its head. This isn’t the core driver of what’s going on, but Xi’s willingness to totally clean house — renovate the military, strip the high command down to its studs — shows he feels pretty comfortable about the external environment and the cross-strait environment in particular.

There are three big reasons for that. First, President Trump doesn’t seem personally invested in the Taiwan issue. The national defense strategy doesn’t even mention Taiwan, and they’re reading that signal pretty clearly. Second, President Lai Ching-te, whom they loathe, is in political trouble at home after the failed recall campaign this summer. There’s going to be an election in 2028, and the opposition KMT’s new leadership is saying very favorable things about Beijing. From their perspective, they’ve got breathing room, and 2028 is probably the next big pivot point where they sense a real opportunity to shape and shift the dynamic.

2. Manufacturing leaves a legacy, even if its revival is jagged.

From How Ukraine Built Drones with Cat Buchatskiy: Civilian talent is portable, and having a historically strong manufacturing base can be an asset even if the capabilities have been dormant for decades. In Ukraine, lapsed manufacturing and support capacity for traditional systems both forced and facilitated the pivot to drones.

Cat Buchatskiy: Although we had the history, and Ukraine has historically been full of engineering talent with a lot of that knowledge, the manufacturing was not maintained to the extent that it should have been. Most of our legacy exquisite systems were completely out of date, in need of repairs, and basically unusable. One of the huge reasons that we had to start using USVs and sea drones was because our fleet was in complete shambles and complete disrepair. Even though we had some ships, it just wasn’t realistic to use them in a wartime scenario at all.

A lot of the tech talent in Ukraine wasn’t actually working in the defense industrial base at the time. Ukraine was famous for its IT industry, software, and computer science. When the full-scale invasion began, harnessing civilian talent was one of the big things that kept us in this fight. Many people who were previously working in the software industry, in consumer goods and technologies, completely shifted.

3. Chinese developers have routed around every Claude crackdown — and that process is making Chinese models better.

Anthropic restricts Claude access in China, but Chinese developers are all using it anyway. In How to Buy Cheap Claude Tokens in China, Zilan discovered that Chinese companies are using Singaporean entities to serve up Claude tokens at a 90% discount, alongside several other layers of workarounds that have evolved in response to Anthropic’s iterated restrictions. The conversations users have with Claude through these discounted API calls are then used to train Chinese models.

4. On-the-ground reporting is underrated.

The audience reception to our travelogues has been exceptional, and we’ve only recently started to appreciate how much value this type of reporting unlocks. The knowledge we collect from spending time in a new country isn’t always tangible or actionable, but insights accumulate and often become useful later in ways we didn’t expect.

From Nick’s Notes on Egypt:

The Japans and Norways of the world can still build metro stations and decent housing, so America’s particular impotence isn’t a universal feature of democratic governance. But those countries are also quite wealthy, and the more striking pattern runs in the other direction. Egypt, Kazakhstan, and yes, even China (once you step beyond the tier 1 and 2 cities) all retain a capacity to build that feels out of proportion to their material means. Their GDP doesn’t fully explain what they’re able to put in the ground.

The actual dividing line, therefore, might not be China versus everybody else, but between countries with entrenched rule of law and weak political consolidation on one side, and countries without either constraint on the other. One model can build a city for six million people and leave it empty. The other struggles to build enough housing for the people already there. Somewhere between the autocratic state that builds too much too fast and bankrupts itself chasing a pharaoh’s legacy, and the democratic state that can’t break ground on an apartment block without a decade of environmental review, there has to be a better answer. Finding it may be the central urban planning problem of this century (the one Abundance and Breakneck and YIMBYs are all trying to figure out), and, depending on how you look at it, the algorithm for how civilizations grow or decay.

5. Social technologies like mascots are cheap and underrated!

Egypt spent ~US$58 billion building a revolution-proof capital city that no one wants to live in, while Taiwan rebuilt civic trust after decades of martial law with cute mascots. From Lily’s Goodbye, Taiwan post:

(If you’re a new subscriber, you can read our old travelogues from Kyrgyzstan, Korea, and beyond here. Lily also wrote this post after spending 10 days touring Chinese AI labs.)

6. China’s top AI leadership is pessimistic about catching up.

We translated an all-star panel discussion about the state of Chinese AI, in which panelists argue that China’s success with open source cannot compensate for the persistent and widening gap with US closed models. Qwen’s former tech lead estimated that there’s only a ~20% chance (“already very optimistic”) that the leading AI model in 3-5 years will be a Chinese one.

7. China has no interest in confronting the US Navy… and if they do, something will have gone very wrong.

From Ed Luttwak on Military Revolutions: China challenging the US Navy would be “high-grade idiocy” because American naval power protects Chinese seaborne commerce. Unfortunately, that same economic incentive didn’t stop Imperial Germany from building a fleet to take on the British Royal Navy…

Edward Luttwak: That’s exactly what Xi Jinping says in China, “We have to challenge the US Navy, because if we don’t, one day, they will suddenly shut down everything.” This is not an ordinary error — I made mistakes. I made them because I’m stupid various times. I made a mistake yesterday. I made a little mistake because of some foolish calculations. You need intellectuals for this. To make that level of error that Germany made in challenging the Royal Navy, beginning the whole competition, causing the British to start organizing a coalition against them — for that you need the intellectuals.

Edward Luttwak: As I say, his luck was von Moltke, the chief of the general staff. It was intellectuals, naturally, professors and so on, who said, “We have the most powerful army in Europe. We have unredeemed Germany, because we have all these Germans who are not in Germany. Everybody else is unifying, the Danes, the Italians, the damned Portuguese. Only we have Germans stranded outside Germany. So, we have to use our power to unify Germany like everybody else. Are we racially inferior that we can’t be unified?” These were very strong arguments. They were not advanced by hotheads in pubs. They were advanced by university professors.

If you already listened to the Luttwak show, you’ll also enjoy our less-viral show with Paul Kennedy:

Paul Kennedy: It would take a big concession, an act of inordinate political wisdom for the rising successful number two to say, “I understand the neuralgia of the number one as we become more and more a success story. I am going to be superbly clever here, and I’m going to temper down the shape and the size of the imperial German navy.”

Paul Kennedy: So I do wonder whether distance in the first place, then cultural and ideological similarity or antagonism, and then trade rivalries and other things, put together form the complex explanation as to why Britain found it so much more difficult to deal with and be reconciled to a rising Germany than to a rising United States.

8. China holds roughly one-eighth of global AI compute.

Nick and Aqib used two independent methods to estimate how much compute China has in total, which converged on the same number. Both supply-side and demand-side calculations indicate that China has ~2.7–2.8 million H100-equivalents.

Check out the full posts and appreciate the math in all its glory:

9. With AI chips, quantity is not a substitute for quality.

Export controls are working. From No Jensen, Not All Compute Is Created Equal: Networking limits mean China can’t compensate for weak chips by using more of them. Surely, the CEO of Nvidia should know this…

10. The US should think about buying Chinese DRAM.

The memory crunch has received a lot of coverage in the past few weeks. The simple solution, according to Aqib, is to let American companies buy DRAM from China. “That sounds like a national security risk,” the China hawk on your shoulder might protest. He’s wrong. If Chinese companies are incentivized to produce commodity DRAM for us, Aqib argues, they’ll have less capacity to build AI memory (HBM) for themselves. And of course, buying Chinese memory would unlock more supply and better prices for American consumers. Read Aqib’s full takedown of the natsec objections here, or leave a comment telling him he’s wrong.

11. A dollar buys more AI compute in America than in China.

How Much AI Does $1 Get You in China vs America? In February, Aqib calculated that American chips are so much better than Chinese ones that, even with cheaper Chinese electricity and construction, $1 gets you more FLOPS in the U.S. (~24 GFLOPS/$) compared to China (~14 GFLOPS/$). We found that construction and chips account for almost all of the cost of running a data center, and other bills like electricity and water are basically rounding errors.

12. National champion status is strangling DeepSeek.

DeepSeek researchers worked overtime to launch their new model in time for the Labor Day holiday, but there were no karaoke or cocktail parties thrown to celebrate. The anticlimactic capabilities of V4 stemmed from failure modes that ChinaTalk predicted back in February of 2025 — national champion status caused DeepSeek to bleed talent alongside new directives to reduce reliance on Nvidia chips and CUDA.

Irene documented exactly how these limitations played out with some brilliant open source research:

With V4 out now, DeepSeek is in the throes of a dilemma that cuts to the center of its tripartite mission. While OpenAI’s large-scale marketing of consumer and enterprise products smoothed its transition into a for-profit company, DeepSeek missed out on a golden period of market development inside China. Between V3 and V4, ByteDance’s Doubao became China’s most-downloaded chatbot; vertical-specific AI products — like Alibaba’s health app Afu — achieving groundbreaking success; and MiniMax and Z.ai, two pure-play model makers, went public and broke into international markets. DeepSeek, arguably, came late to realizing the importance of revenue under the Chinese market’s capital constraints.

13. Unitree’s robots have escaped containment.

From Unitree Goes Public: Unitree, China’s most notable robotics company, has a customer base rapidly diversifying beyond universities and research institutions. Revenue from commercial and industrial sales saw triple-digit growth from 2024 to 2025, with non-research applications now driving nearly 70% of earnings from quadrupeds and more than a quarter from humanoids.

14. China’s AI optimism isn’t what it seems.

Polls don’t capture the fact that China’s AI adoption is driven by fear and a “last bus” mentality as opposed to sunny techno-optimism. Another brilliant post by Zilan:

The polling data is striking: Stanford University’s 2026 AI Index Report shows that more than 85% of Chinese respondents see AI as more beneficial than harmful, compared to less than 45% of respondents in the United States. A 2025 report published by the University of Queensland and KPMG Australia revealed that 73% of Chinese respondents are willing to trust AI system outputs and share relevant information with AI at work, and 88% intentionally use the technology, compared to 52% and 48% of Americans, respectively.

Why does Chinese society, which suffers from acute job loss and a youth unemployment rate close to 17%, embrace a technology it knows is likely to take away more jobs?

The question was answered three decades ago. The answer is not a narrative about AI, but about an earlier transformation also perceived as inevitable. It is a story about how Chinese society has learned, through repeated upheaval, what it believes to be the only permissible response to disruption. Accurately interpreting that response — which is often misleadingly called “enthusiasm” — is essential to understanding that worried Americans watching China’s AI frenzy might not be looking at a rival but into a mirror.

As a corollary, the CCP isn’t sure what to do about job displacement. Chinese courts have ruled that firing workers made obsolete by AI is illegal, but the government is still worried that automation could destabilize society. Collapsing local government budgets will likely limit the party’s ability to cushion AI-induced displacement the way it once accommodated laid-off coal workers. (For more, check out China on AI Job Loss: “No ‘Matrix’ for us, thanks.”)

15. China is all-in on AI for education — and it could be a boondoggle anyway.

China’s education system is deeply unequal and ruthlessly zero-sum. Entrenched interests mean that addressing the root causes of this inequality (like the gaokao and hukou systems) is a non-starter politically, but improving the appearance of educational institutions with shiny new hardware is a proven method for boosting the party’s popularity. The Ministry of Education’s grand AI initiative is at risk of falling victim to these perverse incentives, boosting surveillance more than improving educational outcomes. From Lily’s deep dive, supported by a grant from Tarbell:

[T]he April 2025 opinion instructs schools to “strengthen the overall planning of funds to ensure expenditures on digital education” (学校加强经费统筹，保障教育数字化支出), but budgets are already stretched thin. One school in Gansu province (官鹅沟小学) reportedly devoted two-thirds of its budget to internet fees in order to meet the basic connectivity requirement, which left no money for maintenance. The government has been burned before by the disease of “emphasizing construction while neglecting application” (“重建设、轻应用”) and is thus hesitant to shell out its own funds this time. In the early 2000s, China’s central and local governments spent nearly 2 billion RMB altogether buying DVD players, satellite receivers, and internet infrastructure for rural pilot schools — but left teacher training, maintenance, and operational costs up to the schools. Officials fulfilled their KPIs, but the equipment often fell into disuse once they left.

Finally, there’s the problem of scandals. A pattern of schools forcing families to buy overpriced tablets has poisoned the well for device distribution programs. For example, a middle school in Anhui’s Wuhe County charged students ¥5,800 (~US$841) per tablet; the principal was removed in response. CCTV warned administrators, “Do not use ‘educational informatization’ as a pretext to turn students into profit-making tools,” and the MOE issued a 2022 directive explicitly prohibiting schools from forcing tablet purchases in response.

But the reality is that Chinese public school students have basically no expectation of information privacy. Schools routinely expect (and pressure) parents to sign agreements granting the school expansive powers over student data, and tons of personal information is shared via WeChat as opposed to a secure portal. Adding AI-powered smart classrooms to this environment is a recipe for dystopian scandals — collecting and analyzing student biometrics is an explicitly advertised functionality of this hardware. I don’t just mean faces or fingerprints either — these systems are being designed to film classrooms and analyze student body language/facial expressions to determine which students aren’t paying attention. What happens when administrators decide to sell that data to the private sector?

I worry about the fact that the MOE is so heavily emphasizing mental health interventions as a use case. The white paper greenlights using AI to decide if a student needs psychological help, but it doesn’t provide guidance on what that help should look like. Imagine how humiliating it would be to be pulled out of class to be questioned about your emotions because an administrator needed to fulfill a KPI.

16. Old industrial policy tools can’t unseat China’s biotech dominance.

From The Biotech Empire of Wuxi: The key to the Wuxi companies’ biotech success is their business model. They vertically integrated the entire pipeline for contracted drug development from R&D to manufacturing, and with such efficacy that WuXi AppTec alone is now involved in manufacturing a quarter of all drugs consumed in the US. The US is trying to slow down Chinese biotechs like Wuxi by using a similar playbook to AI export restrictions, which probably won’t work. You’ll love this piece if you enjoyed the peptides radio show.

WuXi doubles down on this model by targeting a “long tail” of biotech customers. Rather than limiting themselves to massive deals with the pharmaceutical giants, they target many small- and medium-sized firms. With more limited resources, these small companies benefit particularly from the cost efficiency of WuXi’s end-to-end services, which then locks them into the pipeline. Their sheer number and diversity also diffuse the risk of major damage from any one customer pulling out. Furthermore, research by consultancy firms has shown that these smaller companies tend to produce more innovative drug leads than their big pharma counterparts. WuXi is therefore able to link itself to these disruptive — and therefore lucrative — products early on. These strategic decisions have given WuXi a “strong, diverse, and sticky customer base.”

The U.S doesn’t have an easy way to address this. China’s specific advantages in biotech look less like control over a single node and more like what it achieved with its manufacturing sector. It is about process expertise, cost efficiency, labor and talent, and deep integration into global supply chains — perhaps more like BYD’s success in the EV sector. These are not easily reducible to export-controllable chokepoints.

17. China’s clinical trial abundance is producing breakthrough medical innovations.

In China, individual doctors can set up clinical trials with real patients without approval from a centralized authority like the FDA. This system produced the blockbuster cancer drug Carvykti, as well as a wave of experimental therapies available only in China. Jacob Stern explains the mechanics of this system based on his experience on the ground:

One of the defining features of IITs is that reputation acts as the primary coordination mechanism, and that in turn helps enforce safety. Investigators are highly attuned to reputational risk. Because a death or a serious adverse event can have lasting professional consequences, they design protocols carefully, demand strong supporting data, and prioritize projects they believe are both safe and scientifically credible. At the same time, relationships play a central role. Since every trial involves uncertainty, investigators tend to work with collaborators they trust from prior experience. As a result, trial opportunities are allocated less by price and more by a combination of trust, track record, and perceived scientific promise. And the balance of supply and demand is such that academics with the platforms to do IITs and recruit patients quickly have many options, with both local and global biotechs approaching them with ideas, so they can be choosy.

On my trip, I was repeatedly quoted a timeline of 18 months from a company having an idea for a therapy to testing it in a patient. My lived experience from my week on the ground backs up this speed. I experienced a sense of urgency at every level. Not just from start-up companies themselves, but also the ecosystem of third-party vendors that perform services for these companies.

During a visit with a CDMO focused on cell therapy manufacturing in Suzhou, I asked the business development rep giving the presentation about the company’s experience with non-viral gene editing. He picked up his phone. As we were preparing to leave 10 minutes later, the principal scientist responsible for the non-viral editing platform caught us by the door. He answered my questions, and we figured out the next steps to evaluate the suitability of their platform for the non-viral editing approach our collaborator is using.

18. China’s power lead is structural.

China’s grid was built under a centralized authority with newer tech compared to the US grid. In Transmission Dominance With Chinese Characteristics, Dana Golden compares US electrical buildout:

As electric power in the early days of electrification was highly tied to political and economic power, companies providing the power benefited from creating monopolistic fiefdoms. The utilities had little incentive to interconnect beyond what reliability required. The current structure of the grid even ensures that different utilities can have different incentives, with some utilities with power plants even benefiting financially from higher wholesale prices even as other utilities have their lunches eaten. Even after much of the market structure has been improved, the continuing result is a web of seams: jurisdictional boundaries that create physical and economic bottlenecks wherever one utility’s territory ends and another’s begins.

vs China’s:

China’s system is built on a different logic entirely. For decades, electricity was allocated through government-set benchmark prices and administrative dispatch. Generators received a guaranteed number of operating hours at regulated tariffs, and the grid companies (SGCC and China Southern) handled transmission and distribution under regulated cost-of-service models. There was no real-time price formation, no congestion pricing, and no market signal telling investors where transmission was needed. The central plan told them.

19. Yes, transformers are a problem, but so are converters, inverters, and MMCs!

Dana wrote an incredible deep dive on the multi-tiered obstacles holding back grid modernization:

The creation of GOES is highly specialized and concentrated with only a few producers worldwide, such as Baowu Group, Nippon Steel, and POSCO. China alone produces 56% of GOES. Currently, the US produces only 12.5% of global GOES, with major declines in a once-dominant industry due to underinvestment and foreign subsidization. Cleveland-Cliffs in Ohio is the only producer of GOES in North America, with production capacity for only 240 thousand tons compared to estimated North American consumption of 489 thousand tonnes. The perpetual lack of domestic supply comes despite over a decade of handwringing over the need for US steel. As President Trump said back in 2018, “if you don’t have steel, you almost don’t have much of a country…” Policymakers promise to bring back US steel, David Riccardo be damned.

I have another thought experiment for you. It’s 2027, and China has carrier groups in the South China Sea headed for Taiwan. The current US administration has chosen to escalate. Maybe the escalation looks like sanctions. Maybe this is a full-scale military fight. Either way, the Shanghai Metals Exchange is closed for business to the West. Suddenly, analysts get on TV shouting about our dependence on Beijing for rare earth elements and — of course — critical minerals! So critical. Important minerals.

Talking heads from MSNBC to Fox cannot agree on whether we should be defending the ROC against the PRC, but everyone seems to agree that we are effectively up shit’s creek without a paddle. We have brought a shovel to a drill fight and cannot wage war without the ability to purchase critical minerals. Suddenly, prices from copper to lithium to ytterbium shoot to the moon. Weeks later, prices for almost all metals have come back down. We discover that only a small subset of metals could not be procured in the medium run without trade with the Middle Kingdom. Turns out our vulnerabilities in this space are almost entirely restricted to a few elements — some notable rare earths, and importantly, gallium.

For a full deep-dive on gallium-nitride, check out Aqib’s post:

20. Not all minerals are equally critical!

Essay contest winner Farrell Gregory devised a system for allocating critical mineral spending more efficiently. From Critical Mineral Security: The Endgame:

We remain reliant on China because U.S. critical mineral policy has recently taken a categorical approach, rather than prioritized strategy. These policies treated critical minerals as an undifferentiated category, deploying uniform benefits that apply equally across a wide range of materials regardless of their varying exposure and likelihood for interference.

Since the United States Geological Survey list was established in 2018, it has grown from 35 to 50 to 60 materials, with many of those latter materials marginally exposed to China. Because the USGS list is the reference point established by the 2020 ENERGY Act and cited as a definition by many subsequent federal and state laws, broadening the list affects policy and dilutes any categorical benefits across a wider list of critical minerals. Because these categorical benefits do not distinguish between the relative criticality of different materials, they offer benefits to the most consequential and insignificant minerals alike.

What is needed instead is a shorter list that prioritizes materials for targeted investment, defined by proven Chinese ability and willingness to interfere with those specific supply chains, as well as the possibility for government support to make a meaningful difference in the short term. This prioritized list translates into quantifiable KPIs that can be used to measure our success in decreasing reliance on China. America should be well past the point of treating critical mineral risk as a hypothetical — the Chinese certainly are. Our policy should reflect this reality by concentrating active government support for the few materials where China has imposed export controls.

Special thanks to for the style inspiration!

We hope ChinaTalk’s reporting has added value to your life! If you’d like to support us, you can check out our merch store, upgrade to a paid subscription, or share one of our posts with a friend.

Subscribe now

If you have feedback for our team, we’d love to hear it! Respond to this email or leave us a comment to share your thoughts.

Thanks so much for being a member of this community.

Our conversation covers:

• Why robots are the real general-purpose technology because, for the first time in history, we can decouple capital from human labor.

• How everyone keeps underestimating Unitree: the DJI and BYD playbooks, the danger of dismissing “robot dogs,” and why iteration speed matters more than dancing demos.

• China’s edge: vertical integration, actuator manufacturing, and supply chains that make Chinese humanoids much cheaper than American ones.

• Why you can’t AI your way out of a hardware problem or from a supply chain that makes your robot far more expensive to build.

• What America should do next: allied supply chains, special economic zones, and industrial policy.

Why We Should Care About Robots Now

Jordan Schneider: Niko, let me start with you. Why should people care about robots now?

Niko Ciminelli: We’re in a world where everyone’s excited about digital AI — Claude Code, Codex, ChatGPT, etc. These tools are boosting productivity in white-collar cognitive work, but today they remain very much complements to people. Whether that stays true remains to be seen.

Robots are different. They’re a true general-purpose technology, the same way you’d think about AGI. For any subset of tasks where you can deploy a robot, you’re absolutely replacing someone you’d otherwise have needed. These jobs are usually fairly harsh, not well paid, and increasingly scarce in the West.

What’s interesting is that this is the first time in the physical world where a general-purpose robot — at the level of dexterity and capability we’ve written about — lets you fully decouple capital from labor in a profound way. Your production capacity becomes deeply tied to how many robots you can make. In a future where you want to spin up a small business or a manufacturing company, or where the US has a strong prerogative to make X, Y, and Z, robust robot capacity lets you do things you couldn’t previously imagine or resource. This is the first real general-purpose technology we’ve seen in the physical world in probably a hundred years, and that’s why it’s going to be as profound as digital AI in the limit.

Much of the data we want these magical AI returns from is scarce and physical-world oriented. If we want to build the Dyson sphere, build robots that build robots, build all these unique technologies, we need general-purpose robots. In pragmatic terms — to fix the labor shortage, to raise quality of life, and to do the things the West has veered away from — national-security-oriented, sure, but really the ability to improve citizens’ quality of life — that’s a very real possibility.

Jordan Schneider: Let’s level-set. It’s late June 2026 — how far are we from these things being genuinely economically productive? There’s this wondrous sci-fi vision, but you don’t get there until there’s a market for the crappy version: something just barely better than a human at some economically useful activity, enough that it’s worth scaling from making 1,000 to 10,000 to 100,000 units. Where would you put these humanoids today?

Niko Ciminelli: I’d take a slightly different perspective, because this gets really hard. Part of my motivation for helping SemiAnalysis cover this space is that we tend to view industrial automation through two lenses: bespoke technologies or sci-fi technologies.

Our first article was “Levels of Autonomy,” focused broadly on general-purpose robotics. In many ways, these systems have improved at a gradual rate that tracks progress in AI. A few things are different about robots: the data is harder to get, the unit economics differ, and the deployment constraints differ — how performant do you actually have to be before you can deploy in the real world?

This year isn’t one where we won’t see many deployments — we absolutely will, just for a very specific subset of tasks. You could argue that over the last year or two, pick-and-place saw enormous improvements, and we’re still watching those capabilities speed up and deploy across different parts of the world.

I caution people against framing this as a binary ChatGPT moment, because frankly I don’t believe in it. Certain results over the last six to nine months could be coined that way, but the label becomes uninteresting, because there have been multiple inflection points showing that certain approaches are scalable. Over the next two to three years, these will make a large swath of coarse manipulation tasks deployable at a scalable level. Within two to four years, we do see coarse manipulation with mobile manipulators becoming available to everybody.

Unitree and the DJI Playbook

Jordan Schneider: Let’s talk about the DJI analogy — going from making a drone that people actually wanted to buy, even though it couldn’t do everything you wanted, to being good enough to get the flywheel running, so that today we have this giant commercial drone ecosystem. Tell that story.

Reyk Knuhtsen: The DJI story is quite a tale to look back on when you examine what’s happening with Unitree right now.

People look at the drone market today and say DJI owns it — roughly 70% market share. Everyone you know who uses a drone probably uses a DJI; every filmmaker does. It seems ubiquitous now, and most people take it for granted without knowing how it got there.

Before the first Phantom, back around 2013, the consumer drone market basically didn’t exist. A competent drone cost around $20,000. Otherwise you’d buy the components and build one yourself for a couple thousand dollars, attach your own camera, and end up with something unstable, with poor flight time, that would crash. It was a mess and expensive all around.

Then DJI shipped a moderately better drone. It wasn’t exceptional — no 4K camera, still not very stable, flight time as short as ten minutes — but it was a fraction of the price, and that was enough to get people buying. It worked well enough to film something, to do things people couldn’t do before. That was enough to jump the company from about $4 million in revenue to roughly $130 million when the Phantom 1 came out in 2013. That’s the creation of the consumer drone market in real time. Suddenly there’s a whole market that didn’t exist before, one nobody had considered because it wasn’t on the table — and now it’s there for the taking. DJI took the money they made and kept scaling it into their drones: better stabilizers, longer flight times, better cameras, easier to use, cheaper as production scaled. Those economies of scale compounded, and now most people know what a DJI drone is and acknowledge it’s just the drone to buy. But there’s a huge, fascinating gap in how that came to be — the drone market was created out of thin air because DJI made something good enough that people would buy it and fund the market’s creation.

Lily Ottinger: Unitree’s IPO prospectus indicates that a growing section of their customer base is now commercial users — and it seems most of those are deploying the robot to dance in an entertainment setting. That’s not exactly what you’d imagine robots doing in the techno-optimist future. Is that going to be enough to set off this flywheel effect? Is this the birth of the consumer robotics market?

Niko Ciminelli: It’s unclear whether the current form factors are that. When Reyk and I looked back at how we so often get surprised by China’s manufacturing capacity, we found it comes in waves — not one complete unlock. Unitree’s revenue is very analogous in this regard.

Yes, this trajectory will continue, but it’s been heavily research-driven; they started in research. I wouldn’t call this the commercial moment where everyone buys something for their home for fun — these robots aren’t at the cost or quality level to make you feel safe doing that. But Unitree has the R1, which is about as cheap as a decent computer. That’s the mind-boggling part: it’s not actually a high-quality robot, but if you have decent spending power as an upper-middle-class white-collar worker, it’s within reach. That doesn’t mean one goes in everybody’s house — maybe they’re not there yet.

But first, it’s impressive they got there. Second, small businesses may use them for the things Reyk described. It becomes a bit of a meme, a small amount of usefulness, a novelty. This kind of thing captures the public eye.

Reyk Knuhtsen: To expand on Niko’s point, more and more people are going to buy them — for entertainment, for research, or, in the case of the very few, deploying them in industrial settings for actual job-based tasks.

The fact is they’re selling more, and they’re slowly eking into actual job-like settings. This can keep happening incrementally. The question, as Niko said, is whether this is the right form factor, whether we’ve found the right task, and whether the AI is there. But as long as this continues, the robot will get better, get cheaper, and become more accessible, which might improve it in other ways — making it more useful for specific tasks, unlocking new tasks, or fueling a research flywheel that makes the AI unlocks easier and the robots more competent generally.

Niko Ciminelli: The core kernel of our last paper — not to harp on gradual change — is simply that Unitree was a quadruped company no more than a handful of years ago. People don’t tend to remember; our memories are short. So much is going on in the political economy, in AI, in everything, that people forget Unitree was just a robot-dog company. And I mean “just” the way it was used — almost pejoratively, as if that isn’t a market in itself.

In that rapid pace of progress, they reached something like the H2, a far more performant robot in payload capacity. They clearly have deeper ambitions — the cheaper consumer robot is within their sights, and they’ll also compete on robots that do economically valuable work. The G1 is an in-between robot I’d call a research robot; it’s used for entertainment, but we’ll see much better entertainment robots as they expand. The H2 is okay, but impressive that they got there so quickly — and there’s no reason to believe they’ll stop.

Some people point to their entertainment strategy a lot these days, but their iteration speed is what people should focus on.

Reyk Knuhtsen: The trajectory is what we were mainly trying to highlight. We don’t have to argue that the G1s are the ones you put in the factory right now, that they’ll be it, that they’ll do everything. No — these bots didn’t really matter before; they were research tools or toys. Now people are trying them out, taking them seriously, and trying to deploy them, and the pace at which that happened is remarkable. Then the H2 comes — a more performant bot. If this trajectory continues, it’s exciting.

How Much Better Are the Robots Now?

Jordan Schneider: Can you talk about the dimensions on which these robots have gotten, and will keep getting, better? How long they can run without a human fixing them up, how long they can hold something at ten pounds without overheating or tiring out — what do you need, besides more kung-fu moves, to start filling in for an actual human currently doing the job?

Reyk Knuhtsen: On overheating: look at the original G1s from a year or two ago, used in research tasks where they could carry a box of a couple kilograms for maybe five minutes at most.

Then it would overheat, and you’d have to let it sit in the corner for 30 minutes — sometimes a full hour — before doing another five minutes. They’ve clearly iterated on this. As we showed in our paper and heard from people, you can now get around five minutes of work with ten minutes of rest, even up to ten minutes of work with ten to fifteen minutes of rest. It varies by task, but it’s a vector they’ve clearly improved.

Other dimensions are interesting. Right now, admittedly — and we point this out in the paper — a lot of what we’re discussing is specific to one task: you pick up a box and move it from A to B. We’re not doing tactile manipulation, nothing that requires an understanding of force, like plugging something in. It’s just moving a box. In that regard, you could see battery life improving, stability improving, general controls improving so it functions better in the environment and is safer around people and objects. But moving beyond that task raises different questions depending on which task you look at.

Jordan Schneider: Let’s talk about hands. How do we get my data-center electrician?

Niko Ciminelli: The framing is that you have two main axes to improve. Hardware has to improve — cooling, more performant actuators, internal tracking, and, as Reyk mentioned, battery life. Those are specific to Unitree; every robot involves different design trade-offs. On the AI side, we’re not yet at the data scale to get very dexterous tasks unless you build more custom models.

On hands and data centers: for hands, we have companies like Wuji and Sharpa, and a few in the US — Tesla is doing interesting things. We’re making a lot of progress. But the important question is whether hands matter for the data-center electrician, and I’d argue no. A handful of startups, and quite a few non-startup projects, are increasingly interested in data centers. The cost of data centers is rising super-nonlinearly, especially as the buildout continues. You can make a custom end effector and a custom model that executes this well.

Also, you can’t teleoperate when the data center isn’t fully brought up. Even a gen-AI data center may be air-gapped — they often are — so you can’t always use teleoperation. We will get robots doing electrician work, at least cabling, as early as this year, deployed more materially over the next few years.

So we’ll have a very jagged frontier. Even though cabling is dexterous and has catastrophic failure modes, clever engineering — harnesses, safety backstops, extra code surrounding the model, the same way you’d approach gen AI — gets you a lot further than you’d think. Think of it like tool use in AI: a few clever application-layer companies are 3D-printing tools to surround a specific object — a spoon, or something a two-finger end effector can’t hold easily — which lets you jump the dexterity scale massively. That’s why these deployments will be jagged and gradual: there are many ways to get much further than the general model gives you naturally. In a bespoke domain, custom data and distribution data are really all that matters. That doesn’t mean the general models won’t come and unlock massive value, but robots will get deployed by many means over the next half-decade.

Jordan Schneider: Speaking of custom data — I just got served an ad offering to clean my house in New York City for free, because the cleaner wears a headband with a GoPro on it, and apparently that’s data. How is that economical?

Niko Ciminelli: I don’t know. These startups exist, and some have more interesting unit economics than others. But we’re all going to be collecting data and doing a lot of this pretty soon. This is the beginning, not the end.

Reyk Knuhtsen: We’re soon to be all data labelers. It’ll be a really good time when we have all the robots taking everything for us.

The BYD Warning

Jordan Schneider: Let’s come back to the Unitree arc and the BYD analogy. How are they running the playbook of internalizing more and more of their bill of materials, so they don’t rely as much on outside suppliers?

Niko Ciminelli: The most important thing to note is that BYD is a proof point. It was an underrated company in the West for a long time. We didn’t examine their quality at all. We use them because it’s a more intelligible example than what China has also done in cobots.

China’s arms were long considered low quality. BYD’s cars were considered low quality for a very long time. These things sneak up on you because of extreme economies of scale — and because you can close the loop tightly while still in R&D.

Right now, the vast majority of robots in China are still hand-assembled. They’re not at scale; this is still very high-mix manufacturing. One reason Unitree benefits so much from in-housing actuator production is that the designs aren’t set. That early decision — not to go to contract manufacturers until you have a high-volume, settled design — matters. They have used CMs before; I believe the G1 now uses one for specific parts.

If you don’t control every level of the stack, you don’t get the internal feedback loop or see all the tiny details. They go from rare-earth materials to a full robot at full pace, with suppliers down the street. That iteration loop is what let them improve so fast.

Typically, once you reach some revenue scale, the West demarcates you as one thing. They labeled Unitree a quadruped company; now they’re labeling it an entertainment-robot company. It’s a warning sign — we’ve seen this before. Look at BYD. We spent a decade saying China makes crappy cars. Now they’re extraordinarily cost-competitive at higher quality than many Western vehicles. They do this through extreme verticalization within a very dense external ecosystem. That’s the core kernel.

Jordan Schneider: Actuators — what are they, and why do they matter?

Reyk Knuhtsen: Simply put, the actuator is the part of the robot that generates motion. It’s easy to overlook as something simple, but you have to figure out how fast to move an arm, how to move it appropriately, how much torque to apply. If it hits something, should it stop? How does it stop — do you reverse everything at once? There are many configurations for designing an actuator, but the point is to generate motion accordingly.

Jordan Schneider: Why is everyone freaking out that we don’t make good ones in America?

Reyk Knuhtsen: The actuator generates motion, but it’s also the largest part of the robot’s bill of materials. We’ve done internal BOM costing for a number of humanoids — the G1 is the one we published — and the actuators run around 50 to 70% of the BOM. Without them, your robot doesn’t generate motion and doesn’t function.

If you source them externally, you pay all the extra costs on top: shipping, other people’s margins, some value-based pricing. So a robot built in the US — where manufacturing actuators is challenging — ends up structurally more expensive than one built on the same supply chain a Unitree or many Chinese robots use. You’re stuck. To compete on cost, you have to compete on the same actuator supply chain, which is extremely difficult. Our first paper covered this a lot — the many components involved, how hard they are to manufacture, and how they’re just not really made in the US.

Niko Ciminelli: We frankly don’t have the capacity to spin these up easily either; it just doesn’t exist. Many US companies do the winding of their actuators, but all the components come from outside — the aluminum, the magnets. Magnets are rare-earth-based, which isn’t always understood. We have some neodymium capacity, but we don’t process it; the processing goes back to China, and the chemicals for processing are in China. Some of these are multi-decade buildouts serving other industries, giving China a network effect. If we wanted to bring this home, an absurd amount of investment would have to happen today — or earlier — even to be a quarter of the way caught up in a decade.

And it’s not the only place this happens — Malaysia does some processing, Australia has some of these metals, the US produces some. But it’s a very complex, very sparse supply chain outside China, and China is really the only place that can do this at true scale today.

Share

The Need for Production Capacity

Jordan Schneider: Sometimes I hear the argument: we have better software, our robots will be smarter, maybe we pay an extra 20% to build these, but they’ll have an OpenAI GPT-7 brain that your Unitree won’t. Niko’s shaking his head.

Niko Ciminelli: I just don’t know what that means. There’s a lot of obnoxious extrapolation in AI. Yes, progress is fast, but hardware systems integration is a different thing, and hardware production is a different thing.

The West said it didn’t need a manufacturing base — we’d just have good consultants and good banks and do services. That was extraordinarily naive. If you believe in AI progress and that you need production capacity, then deflating the cost of human cognitive labor makes production capacity more vital, because your ability to produce chips, cars, and everything else you need to extract value from intelligence lives in the manufacturing domain. You can discover a drug, but if you can’t manufacture it, you’re out of luck.

Second, what happens when China stops offering its supply chain at all? I don’t know why they would keep offering it. We’re not giving them chips under certain regimes and policies — I don’t know why they wouldn’t negotiate with robots. This is an enormously powerful technology in the limit.

On the AI argument — I don’t know why your robot needs Einstein-level intelligence if it’s 15x cheaper. That may sound like an overclaim, but look at the margins — Unitree can go quite low from where they are today and still make good money. China isn’t known for operating at these margins so early in a company’s life; they’re very willing to run at low margins. We’re not. That’s why people claim their humanoid will cost $100,000. Is that going to hold when Unitree ships a robot at around $30,000 with 80% of the capability? That doesn’t check out. And if they don’t want us to make robots, we can’t make robots.

Jordan Schneider: This is an interesting contrast with “oh, we’ll just ban all Chinese EVs and be fine.” America still gets cars. You don’t need BYD materials to drive around your country, because you already have these hundred-year ecosystems. You pay more, but you’re not leaving an entire vector of economic productivity on the table. If what you’re saying is true and this is an entirely new economic paradigm about to unlock, then you’re missing out on all of it — not just cars that are quieter and a little cheaper.

Niko Ciminelli: The thing constantly left out is memory. Everyone’s excited about the memory story, and it gets even crazier with robots. A lot more memory is going to go on these bots than in the past.

When you get a new AI capability, you find some absurd amount of demand — a shock to supply, a shock to demand; you can model it either way. That may not be fully true, but with a new capability comes extraordinary demand. In semiconductors, you at least had a mature supply chain — not an AI-mature one, but mature. The robot supply chain is extraordinarily nascent.

You might say: we have Nabtesco, we have Harmonic Drive — these things are made. But harmonic drives aren’t made the way quasi-direct-drive actuators are, which is the core thing we keep pointing out about Unitree. One reason people thought they were just a dog company is that they used an actuator considered non-performant. But then their iteration cycles — making it bigger, adding cooling mechanisms, generally better designs — made their robots more performant at lower cost. That’s really why they’ve done so well.

This gets scary because if you don’t invest early, you won’t have it later. We can’t easily catch up to some of China’s electronics ecosystems; for some things it’s simply not economically viable. The belief that we’ll “AI our way out of this” is enormously naive in a way that doesn’t apply the way it did before, like we have FSD — well, now China has FSD too. When are Chinese researchers not good at deep learning?

And we don’t have the industries to do data collection en masse the way they do. Many of our companies doing data collection do it in Southeast Asia, so I don’t always know where these arguments come from. I do think data quality is currently quite a bit lower in China — but again, these things improve rapidly over time. They’ll catch up.

Does America Have a Chance?

Lily Ottinger: Do you think America has a chance? What can be done? Jordan has said before that as soon as China does zero-to-one innovation at the cutting edge, the US will suddenly get really good at process espionage the way China has —

Niko Ciminelli:I love this.

Lily Ottinger: They’re selling these robots, so we can disassemble them and figure out what makes them good. If Unitree has to spend all this money on zero-to-one innovation, is that an edge for US companies that learn to be fast followers? What else can we do to give ourselves a fighting chance? And what happens if we don’t get it together?

Niko Ciminelli: It touches on the same point as last time –- we need to use our allies to build a supply chain, and we need to become much better hardware engineers. We have phenomenal AI research, and we’re very good at deployment, but this is a real gap.

Reyk Knuhtsen: Niko and I have a paper we hope to release soon on what it would actually take — how difficult the US supply chain really is. It’s quite difficult; there are a lot of holes. Look at PCBs: you don’t make copper foil in the US anymore. CNC machines are typically imported. A lot of ordinary industrial machinery nobody wanted to think about, we now can’t make without importing it or depending straight on Chinese product at some point.

Rebuilding this supply chain is far more daunting than most people realize. Say we do a teardown of the G1 and figure out exactly how to build it correctly — great, but who’s going to build it? Who am I going to pay? Most VCs won’t look at a CNC-shop company, because it won’t give them SaaS returns; it’s a hard bet. So you’re left with a strange pocket of businesses that aren’t sexy but are completely necessary, and few people want to fund them.

It’s fun to shout “US reindustrialization” — fun to talk about, fun to hope for, and I do hope it happens. But it’s far more complex than a motto. You’re probably going to need industrial policy to force it, because ordinary capital incentives won’t be enough.

Niko Ciminelli: It’s important to explain what Reyk and I mean by the supply chain. Yes, we don’t make tools in the US. You might say, doesn’t Germany, doesn’t Japan? We have Okuma and DMG Mori for CNCs, and other companies. Let’s pretend those are applicable, which I’d argue they often aren’t, for cost-structure and performance reasons we can get into. We aren’t even good at getting decent yield out of these machines.

We don’t have the tacit knowledge to run these at scale. Manufacturing tacit knowledge is a buzzword now, but it’s deeply true. Take materials processing: you need rare-earth magnets — neodymium, dysprosium, boron, and others. You need to process them, and you need the chemicals to do it. The machines involved are, like CNC machines, very complex, and we lack deep knowledge of making the details work. We talk about this with transformers in AI — the details matter; some layer norm blows up, and the infrastructure doesn’t work at scale for a given configuration. The same is true for these machines.

The hard part for us is the component supply chain. Everyone says we’re funding humanoids, funding robots, funding arms — but these arms and robots aren’t useful without an ecosystem. Venture capital is funding a lot of robotics; there’s enormous excitement. Some things can be vertically integrated — you can wind your actuators. The problem is the second-order companies. When people say “supply chain” and “we build our own actuators,” that’s often not fully true; they do it as best they can.

We likely need economic zones with special environmental policies, favorable tax treatment, and favorable permitting, where you can build quickly and run more experimental attempts. And we need to use Malaysia, Vietnam, Australia, and Japan for what they’re each good at, with trade agreements in place. The only way to do this is with the countries outside China who also worry about their industrial capacity and would bet on benefiting from US commercial strength. The only way up is through: dampen the pain and use other people’s strengths, because getting the tacit knowledge won’t be easy, and getting people to make serious infrastructure investment that isn’t there today won’t be easy either.

Jordan Schneider: The most proximate Washington debate right now is around imports. To what extent should whole Unitree robots, or their components, be allowed to flow into this ecosystem?

Reyk Knuhtsen: On the bots: banning Unitree imports will be challenging in a way people don’t realize. The main driver of US humanoid progress over the past couple of years has been access to Unitrees.

Before, you couldn’t really buy robots that were reasonably priced, reasonably accessible, and not entirely bespoke. Unitrees are fairly standardized — from one G1 to another. If it breaks, you get another at the same price. Back in 2024, you didn’t have many options; Unitree entering the market was a saving grace that created all this humanoid research. No other non-Chinese company has come in to supply these humanoids. There are a few US companies making open-source bots, which are nice, but those typically rely on Chinese parts too if you check the bill of materials. So if you ban these bots, you remove the researchers’ platform of choice and inevitably stall humanoid research.

Niko Ciminelli: Many American companies haven’t been hardware suppliers — the US unit economics and capital markets don’t favor it; it often doesn’t excite investors. Reyk’s point about the researcher’s first choice is definitively true. We’re not thrilled that’s the case, but I agree.

The bigger problem comes next. Yes, we may get our own humanoids soon — Optimus, Figure, Happy Droid, Apptronik, and others, with wildly variable quality and production capacity we won’t get into. But the supply chains and bills of materials are Chinese. We should be careful, because no one here has the cost structure yet. And what happens after a ban? They’ve already put an export control on neodymium, forcing us to buy their full actuator. If we ban their robots, we’re not in a high-leverage position for these negotiations.

You can ban a robot, but that doesn’t mean we can actually do anything. I’m not saying we should have Chinese robots potentially sending data back and handing China a massive advantage — but we need to be more thoughtful about our playbook. Are we doing this just because we think chips pressured them? We have TSMC for that; we have enormous supply chains outside China for chips. But cut off the component supply chain and we don’t have robots. They’ve already threatened this — CNC-machine exports to us are already dropping, and they’ve put export controls on the rare earths this requires. We might make some neodymium, but we can’t process it, and we have no dysprosium or boron — so where’s that coming from? They can cut us off in more places than we can pivot, especially this early. That would cut us off from our own progress. These are complex things to navigate. I don’t have clean answers, but we need to be very conscientious.

Jordan Schneider: On a less dour note — Niko, what are you most excited for these robots to do in the coming years?

Niko Ciminelli: I’m earnestly, wildly optimistic about them doing my laundry and cooking my food. The wondrous part of robotics is that we’ll see the deflation the AGI people promised, but in physical form. You could compress the cost of a lawyer, get affordable legal advice, get primary care — in the universe where that’s true. If a robot can do your yard work or your plumbing, all these severe “taxes” — where people just can’t afford good services — come down. We’ll see a wild quality-of-life increase from personal robots.

We’ll also see US manufacturers operate at healthy margins. With robot labor helping out, there will still be humans, but with robot labor alongside them. There’s a wild revitalization here — both quality of life and manufacturing capacity — where we can do things greener, cheaper, and at higher quality.

I have an investment in a construction-robotics company whose whole goal is that if you bring down the cost of construction, you can make beautiful buildings again — the aesthetics get reborn. These second-order effects of lowering the cost of physical-world labor will expand our capacity to do really amazing things. That’s genuinely exciting. That’s where my head is.

Reyk Knuhtsen: I’m on board with Niko — the cleaning robots will be very nice when that comes to play. Beyond that, I’m excited to see them enter the public sphere generally. It’ll bring more life to scenes with friends if a robot is helping around the house. It could help in an elder-care facility, making the bed or cleaning the kitchen. There are many areas where it could be very helpful.

The driving benefit is that it mitigates the manual-labor shortfall around the globe — from birth rates or whatever you want to point to. You free up, or enable, a lot more capacity for production or other benefits that weren’t available before, letting people do far more than they could.

But personally? It would just be pretty exciting to have a six-foot robot in my house. I think that’d be comical. I think that’s really, really funny.

Shenanigans abound in this episode, including:

• The lights-out scenario — Taiwan only holds 11 days of LNG reserves, and 97% of the island’s energy is imported, but the ruling party phased out nuclear and botched the renewable rollout anyway.

• The offshore wind graveyard — how made-in-Taiwan components drove developers to abandon the world’s best offshore wind sites,

• The Taipower unbundling reversal — and the Kafkaesque system that keeps electricity prices dirt cheap despite the Iran war.

• “Green energy cockroaches” — why corruption is Taiwan’s dirtiest secret, and how the Taiwanese public came to associate renewables with scandal,

• The nuclear U-turn — How President Lai Ching-te walked back forty years of “Non-Nuclear Homeland” orthodoxy to restart Taiwan’s nuclear reactors.

Special thanks to “Jason Feng,” Angelica Oung, Ricky Huang, Tsaiying Lu (DSET), and Yu-Hsuan Yeh (formerly of CSIS and DSET) for their time and expertise. Everyone’s views are their own.

You can watch this episode on YouTube or listen on your favorite podcast app.

Why Taiwan Isn’t Ready for a Blockade

Lily Ottinger: Imagine yourself on the streets of Taipei. It’s August, 2028 — aunties are pulling their wheeled grocery carts through the vegetable markets, and commuters on motorcycles weave through traffic under the brutal sun. But something is missing — the AC isn’t humming, and the lights are out. Eight days earlier, Chinese warships encircled the island, attempting to strangle Taiwan’s economy without firing a single shot.

97% of Taiwan’s energy comes from imported sources — the highest rate of any major economy. The island runs on liquefied natural gas shipped through sea lanes that China’s navy — the world’s largest — can cut at will. Taiwan holds roughly 11 days of LNG reserves. Coal lasts about 42. After those are used up, the lights go out.

Aqib Zakaria: This isn’t a hypothetical war game. Right now, the war in Iran has closed the Strait of Hormuz and damaged Qatar’s largest LNG complex. Asian spot prices for natural gas have surged over 140%. And Taiwan — which gets roughly a third of its LNG from Qatar — is paying the price.

Angelica Oung: “Taiwan should send a thank you note and a fruit basket to Beijing right now because the reason we’re not having an energy crisis is because of China. The Strait of Hormuz has been blocked for a couple of months now. We’re supposed to get 35% of our LNG from Qatar. We’ve just been buying it on the spot market and we’ve been fine because we have money.

 But the fact that there are all these excess cargoes floating around is because of China, because there’s been tremendous natural gas demand destruction in China. They have all this excess capacity in coal plants that they don’t use, and in the crisis they just crank those coal plants on. It does temporarily create more pollution, but that loosened up supply on the LNG spot market.”

Aqib Zakaria: That was Angelica Oung, an energy journalist in Taipei. She started off as a renewables reporter and is now one of the island’s most prominent nuclear advocates.

Lily Ottinger: Taiwan’s politicians have known about their energy vulnerability for decades. And instead of fixing it, they’ve been systematically dismantling every domestic energy source the island has. They killed nuclear. They sabotaged renewables. They doubled down on imported natural gas. But why did this happen?

Welcome to ChinaTalk. Today, with the help of industry insiders, energy journalists, and policy experts, Aqib and I are telling the story of Taiwan’s war on renewable energy.

Ricky Huang: “Multiple forces have united against renewables in Taiwan, each with different motivations. This reminds me of cross-strait relations, where we say ’one China, multiple interpretations.’ Here it’s one anti-renewable sentiment with countless different rationales.”

Lily Ottinger: This story is a fantastic case study in the ugly side of industrial policy, and the West can learn a lot from Taiwan’s renewable energy failures. And yet — this is also a story about how to build progress in a democratic society, how to change public opinion on polarized issues — and I find myself cautiously optimistic about how this story will end. And if things don’t improve… well… Taiwan’s energy poverty might just be able to resurrect the electoral chances of the Kuomintang — even though the public broadly dislikes their ties to China.

Aqib Zakaria: Please let us know if you enjoy this format and want more episodes like this! This is only the second radio show we’ve produced and would love your feedback.

“No Nuclear Homeland” politicized the grid

Aqib Zakaria: To understand why Taiwan is in this mess, you have to understand the party that wrote the policy. The DPP — the Democratic Progressive Party — is Taiwan’s ruling party. President Lai Ching-te represents the DPP. His predecessor Tsai Ing-wen was also DPP. The party was founded in 1986, the same year as the Chernobyl disaster. The party was organized by dissidents against the authoritarian Kuomintang government that ran Taiwan during martial law. From the very beginning, two causes were welded together in the DPP’s DNA: democracy, and opposition to nuclear power.

Yu-Hsuan Yeh: “Taiwan started construction of nuclear power plants in 1971 during the energy crisis. All four nuclear power plants, including the fourth one that never became operational, plus the Lanyu storage site for low-level nuclear waste, went through their siting and construction processes during the martial law era. This means Taiwanese people never got to be consulted — they never had a say on whether they wanted nuclear power plants in their backyard.”

Aqib Zakaria: That was Yu-Hsuan Yeh, an energy policy researcher formerly at CSIS and DSET.

Yu-Hsuan Yeh: “The DPP, formed from the broader anti-authoritarian democratic movement from the mid-1980s to late 1980s. They developed very close relationships with civil society groups organizing anti-nuclear protests. This relationship continues to have political implications today.”

Aqib Zakaria: The specific slogan “Non-Nuclear Homeland” — in Mandarin 非核家園 — actually came later. It emerged around the year 2000, after the DPP first took power under President Chen Shui-bian, and was written into Taiwan’s Basic Environment Act in 2002. When Tsai Ing-wen ran for president in 2012, she attached a specific deadline to it: 2025. By the time she enshrined “Non-Nuclear Homeland” into the Electricity Act in 2017, it had become the DPP’s single most identifying piece of energy branding.

Aqib Zakaria: When Tsai won the 2016 presidential election, the DPP didn’t just win the executive branch — they won a legislative majority for the first time in the party’s history. That gave them the political room to pass one of the most ambitious energy reforms in Asia. The plan had two pillars. Pillar one: phase out all four nuclear plants by 2025 and replace them with renewable energy. Pillar two: break up the state-owned electricity monopoly. We’ll come back to pillar two in a bit and focus on pillar one for now. The plan for pillar one was to shut down every reactor on the island, and replace them with wind, solar, and battery storage. The targets were aggressive. By 2025, the energy mix was supposed to be 20% renewables, 30% coal, and 50% liquefied natural gas. By 2030, the renewables share is supposed to rise to 30%. The whole framework is called the “5-3-2 plan.”

Plant 1 was decommissioned in 2018-2019. Plant 4 — Lungmen, the one that was never completed — was mothballed in 2014 under KMT President Ma Ying-jeou in response to public outcry after Fukushima. A 2021 referendum permanently put an end to construction. Plants 2 and 3 went offline on schedule. Taiwan officially achieved its “Non-Nuclear Homeland” goal when the last reactor at Maanshan went dark on May 17th, 2025. But Taiwan did not meet its 20% renewables target to compensate. As of 2025, renewables only provide about 12% of Taiwan’s electricity, with about 6% coming from solar and 4% coming from wind.

Lily Ottinger: Before we go further, we need to look at the partisan structure that this debate lives inside. The governing DPP is pro-renewables — but historically anti-nuclear because of its authoritarian connotations. The two opposition parties — the KMT and the TPP — are pro-nuclear, but increasingly anti-renewables.

Aqib Zakaria: Quick crash course on the opposition. The KMT — the Kuomintang — is the party that fled mainland China in 1949 and ran Taiwan as a one-party state until democratization. They’re pro-nuclear, broadly friendlier to Beijing, and they currently control the legislature in coalition with the third party, the TPP. The TPP — the Taiwan People’s Party — is the newest party, founded in 2019 by the former Taipei mayor Ko Wen-je. Angelica Oung describes the typical TPP voter as someone temperamentally Republican but socially liberal — politically homeless and focused on bread-and-butter issues. They caucus with the KMT on most things, including nuclear energy.

So nuclear and renewables are on opposite teams, and posited as mutually exclusive energy paths instead of compliments. That means the pressure to combat climate change doesn’t actually motivate crosspartisan cooperation. The people we talked to broadly preferred an all-of-the-above approach, but that constituency is by and large unrepresented in Taiwan’s legislature.

Yu-Hsuan Yeh: “Political polarization has been a huge problem in all aspects. In the past, there were rooms for both sides to support each other’s bills, but it has been harder and harder over the years — probably similar to the US situation right now. This polarization has resulted in the lack of cross-party alliances or consensus.

Especially on energy, which has been a huge topic that people are aware of, the parties have already built up this image that the major two parties and the third smaller party have different stances. Because this has been part of their branding, they don’t find it attractive to work with the other side since they have already attacked the other side on this issue so much.”

Lily Ottinger: To me, it seems that the pressure to “pick a side” on the nuclear vs renewables debate emerged as a result of President Tsai’s initiative to bring nuclear power to zero percent, and her explicit framing that wind and solar were suitable replacements. Angelica argued that the KMT wouldn’t have made this mistake if the tables were turned.

Angelica Oung: “The KMT are not anti-renewables. If they were in power, they’d be pushing ahead with those renewable projects and probably be much more successful because they just have more capacity. I hate to say it — I know it’s a cliché — but the DPP are really more the party of lawyers and activists. When they formed, they were the dissidents against KMT one-party rule. The KMT is more the party of engineers because throughout the authoritarian period, they were a one-party technocracy. There’s a tradition in the party of being good with numbers and respecting the tangible, pragmatic aspects of projects.”

Lily Ottinger: But that hasn’t stopped the KMT from criticizing renewable energy projects on proceduralist grounds. Perhaps the KMT’s real complaints are about cost or corruption, but they have also amplified narratives that renewables are bad for local ecosystems and infringe on the rights of farmers. Both parties have resorted to sensationalist political stunts and smear campaigns.

Angelica Oung: “They’re not anti-renewables, they’re anti-how they perceive the DPP is not doing the renewables projects in a way that is not up to their standards. I do admit they do get into demagoguery because they feel like, “Well, if you’re gonna smear our nuclear power plants, we’re gonna smear your solar panels,” and that’s not worthy of them. But I don’t think that’s behavior that can be solved by a simple political deal.”

Lily Ottinger: Now, Taiwan has a DPP president, and a KMT/TPP-controlled legislature. This is only the second period of divided government in Taiwan’s democratic history. The first period of divided government ran from 2000 to 2008, but there was far less partisan gridlock thanks to Wang Jin-pyng 王金平 of the KMT, a legendary dealmaker who served as president of the legislature. Notably, he facilitated a quid-pro-quo deal that restarted construction on the 4th nuclear power plant despite the DPP holding the presidency. There is no Wang Jin-pyng today.

Aqib Zakaria: Only about 16% of Taiwan’s electricity comes from domestic renewable or nuclear power — and TSMC alone consumes 8 to 10 percent of Taiwan’s electricity. The rest — 84% — comes from burning coal, oil, and natural gas, almost all of which is shipped in by sea. To cut Taiwan’s gas supply, China would only need to blockade three LNG terminals.

Angelica Oung: “In terms of the blockade scenario, you can’t store natural gas. The more natural gas you have in your grid, naturally, the more vulnerable you are, both to a blockade scenario or to just stuff happening — like the Iran situation.”

… “Obviously the Iranians demonstrated it can be done. But for China, they wouldn’t even need to shoot missiles into the Taiwan Strait. They have the vessels, they have the means to say, ’We have a new renewable energy mandate. Any energy vessels coming into our territorial waters have to get this permit. If you don’t have a permit, then you’re out of line.’ That’s just one scenario. There are 101 ways they can disrupt the flow that doesn’t even rise to the level of a blockade, which is an act of war.”

Lily Ottinger: We talked to Tsaiying Lu, who runs the Energy Resilience Group at DSET, a Taipei think tank, who told us that Taiwan’s natural gas stockpile is capped at around 11 days worth of consumption. Strategic coal reserves only last about 40 days under current conditions. Once coal reserves are exhausted, wargaming by CSIS found that Taiwan’s total electricity production would fall to about 20% of pre-blockade levels, at which point all manufacturing would cease. So the question is: what would it take to push that 20% number higher?

Aqib Zakaria: To decapitate offshore wind, the PLA navy would have to cut undersea electric cables deep within Taiwan’s air defense identification zone. Bombing solar fields would require air superiority — and even then, it would be a warcrime to bomb all the apartment buildings with solar panels installed on the roof. Renewables are by their nature distributed, and distributed infrastructure is hard to kill.

The Broken Promise 一紙空文

Lily Ottinger: Earlier I mentioned the 2016 reform had two pillars. Pillar one was killing nuclear. Pillar two was breaking up Taipower. Taipower — the Taiwan Power Company — is the state-owned electricity monopoly. It was established under Japanese colonial rule in 1919, nationalized by the Republic of China in 1946. Taipower is highly vertically integrated — today it controls all fossil fuel and nuclear generation, as well as 100% of transmission, distribution, and retail. Every kilowatt-hour of electricity consumed in Taiwan goes through Taipower. And the only thing they don’t own is renewables.

The plan was to separate generation from transmission and create a market where renewable developers could sell freely to industrial consumers. Nine years later, pillar two was strangled in its cradle by the same party that birthed it.

Aqib Zakaria: On May 9th, 2025, Taiwan’s Legislative Yuan formally rolled back the requirement to break up Taipower. The legislation was sponsored by DPP lawmakers and passed with cross-party support. Taipower’s union cheered.

Jason Feng: “From outside, the whole Taipower structure is like a black box. We don’t know where the losses occur, or why the prices are being set this low. That is why now they’re using this administrative decision-making process to decide the electricity price.”

Lily Ottinger: That was “Jason Feng,” a veteran of Taiwan’s offshore wind industry with years of experience dealing with Taiwan’s grid. Jason Feng is not his real name, and we’ve anonymized his voice with AI tools for his privacy.

As he just alluded to, we have to look at Taiwan’s system for pricing electricity to understand why the rollback is so devastating.

Taipower sets artificially low prices for the electricity it sells, which are approved by the Electricity Price Review Committee in the DPP-controlled Executive Yuan. Then, Taipower turns around and begs for government subsidies to make up the difference. Remember, renewable generation is the only thing Taipower doesn’t own. All of the electricity Taipower generates comes from LNG, coal, and nuclear, so these bailouts are overwhelmingly subsidizing fossil fuels. By mid-2025, Taipower had accumulated NT$418 billion in losses — about $13 billion USD. As LNG prices soared in the wake of the Iran war, Taipower froze household electricity prices at NT$3.78 per kilowatt-hour. That’s about 12 cents US.

Jason Feng: “The reason the government wants the electricity price to be kept so is monetary policy. Because electricity price is one of the important factors in CPI. If they control the CPI to keep it below 2%, then they don’t have the pressure to increase interest rates.”

Lily Ottinger: It would be political suicide to suddenly charge consumers the actual price of the energy they use, which is why none of the three parties wants to liberalize Taiwan’s energy market. Meanwhile, the artificially low fossil fuel prices ensure that renewable energy can’t compete.

Jason Feng: “Now we see renewable energy prices like 4.5 to 5 NTD per kilowatt hour compared to the conventional price averaging 3 NTD. Renewables are much more expensive than the conventional electricity. If there’s no such push or encouragement, normally people will not want to use renewables.”

…”They categorize industries in Taiwan, saying we have this high-margin industry — could be semiconductor, could be AI data centers. These are the ones that make more money. They will be directly impacted by this price increase. But then they also categorize another group of industries that are lower margin, make little money, and they say, ’Okay, you can enjoy a delay of that electricity price increase.’ That really discourages these industries from finding innovative approaches to manage their energy costs. That further slows the whole economic development of the country.”

Tsai’s Baby

Lily Ottinger: Taiwan has some of the best offshore wind sites in the world, as the Taiwan Strait is essentially a 200 mile long wind tunnel. International developers lined up to build there. And then it turned into a boondoggle.

Angelica Oung: “Offshore wind was Tsai Ing-wen’s baby. While she was in office, it got every good benefit, and every time it got seriously in trouble, magically somebody would come and help it out. This is no longer the case under Lai Ching-te, and everything has stalled.”

… “The problem with renewable energy in Taiwan — the original sin is overpromising, because when you overpromise, like President Tsai did when she started the renewables rollout in 2016, you create all these conditions to meet milestones that are very aggressive. You put money into it, and when the amount of money is excessive to the amount of pipeline and bandwidth, corruption is just the inevitable product of that scenario.”

Aqib Zakaria: Taiwan’s offshore wind program runs in three phases. Round 1 was small subsidized demo projects. Round 2, allocated in 2018, offered a 20-year feed-in tariff. A “feed-in tariff” is a guaranteed price the government pays renewable developers per kilowatt-hour, locked in for the lifetime of the contract. It’s the standard mechanism for bootstrapping a renewable industry. We’re now in Round 3 — which is aiming for 15 gigawatts for 2026 through 2035. But Rounds 3.1 and 3.2 came with notorious “local content requirements.” To win the auction, developers had to promise to build wind turbine components in Taiwan. The more localization they committed to, the more points they got in the bidding. The problem is that Taiwan doesn’t have a wind turbine supply chain.

Angelica Oung: “Sometimes when you create political consent for a project, you need to appeal to different constituencies. When Tsai got started, in addition to the very ambitious targets, she also promised, ’This will be a new industry for Taiwan. We’re going to make these turbines at home.’ That’s why, even though the electricity might seem more expensive, it’s worth it for the economy. But the problem was it wasn’t worth it for the economy. There are a lot of wind turbine components that are very specialized, and the way you push the price down is by having scale. If you only had the Taiwan market, there isn’t a lot of scale. Made-in-Taiwan components ended up costing sometimes 2-3 times the components made outside of Taiwan, which dragged the whole profitability picture down.”

…”I’m very disappointed because at one point I was a huge fan of President Tsai Ing-wen. She was such a cool girl boss president and she got us so much international goodwill. She was very good at that part of her job. But her energy plan was just such a disaster.”

… “ And to tell you the truth, that’s when I started doubting the DPP. I used to be quite pro-DPP, and I used to be quite pro the idea that Taiwan should stand up to China and retain our, at least de facto independence and our way of life. But I’m just like, “You guys aren’t taking this seriously. You guys are doing political theater.” Because if you’re not taking energy security seriously, how am I gonna trust that you’re taking national defense seriously? If I don’t trust you to actually have the ability to protect Taiwan, then I’m gonna have to take a fundamentally different approach to how to keep my home safe.”

Aqib Zakaria: Then, the government announced that Round 3.3 would drop localization requirements entirely — prompted by an EU challenge at the World Trade Organization filed in July 2024. Taiwan settled the case in November 2024, agreeing to phase out the rules.

Jason Feng: “Now it’s just a draft, if this final version is so attractive than 3.1 and 3.2, then those who got awarded in those two rounds — it’s like asking them to abandon those green fields they acquired and then move toward 3.3.”

Lily Ottinger: When I was interviewing Jason, this part of the story made me a little crazy.

Lily Ottinger: “So, the government has said, ’Sorry guys, turns out these localization requirements are impossible to fulfill. We’ve just now realized that.’ But the government’s going to continue to hold some of these offshore wind projects accountable to satisfy an impossible requirement that the government admits is not possible to satisfy. And the alternative is that they just abandon this zone that they’ve secured — which is a limited resource around Taiwan? Why can’t the government solve this problem?”

Jason Feng: “Because not everyone failed to realize the commitment. Some of them still can. Copenhagen Infrastructure Partners is the one still proceeding with and making a lot of progress on their development in 3.1 and 3.2.”