The 20 Most Interesting Ideas We've Found in 2026
So far
We are officially halfway through 2026, and ChinaTalk just surpassed 75k subscribers on Substack. To celebrate, here are the twenty most interesting things we learned so far while Making ChinaTalk.
20 Takeaways from H1 2026
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.
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 Most of our legacy exquisite systems were completely out of date, in need of repairs, and basically unusable.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.
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.
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.
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.)
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.
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.
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:
How Much Compute Does China Have?
How much compute does China have? Despite its all-important relevance to American export controls, the AI race between the U.S. and China, and national security, this question remains unanswered.
How Much Compute Does China Have? A Demand-Side Analysis
Yesterday, my colleague Aqib Zakaria published an estimate of China’s supply-side compute capacity. By tallying chip shipments, smuggling reports, domestic production, and estimated Western cloud access, he arrived at ~2.7 million H100-equivalent GPUs
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…
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.
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.
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.
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.
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.”)
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.
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.
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.
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.
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:
Fixing the GaN Problem
In the semiconductor industry, the Trump administration is striving to bring back critical technologies that slipped out of our hands decades ago. The U.S. has attracted billions of dollars in investment to stimulate cutting-edge logic manufacturing, the development of
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 Derek Thompson for the style inspiration!