Becca Wasser, America’s wargaming queen, currently with Bloomberg, joins WarTalk regulars Bryan Clark, Eric Robinson, and .

We discuss…

• Why CENTCOM is using JASSMs to hit targets a glide bomb could handle

• What cosplay costs the Indo-Pacific

• The myth of US air superiority over Iran, and the SEAD legwork no one wants to do

• Who actually benefits from the ceasefire and why Iran has the lower bar for reconstitution

Listen now on your favorite podcast app.

Still No More Ammo

Jordan Schneider: Last week’s theme was no more ammo. Setting that aside, we’re still sending more stuff there. Becca — no one believes us!

Becca Wasser: I think the perennial theme is just going to be no more ammo. And it’s not a matter of the US is running out of missiles to prosecute this war, or Iran is running out of missiles and can’t potentially cause damage if there’s round two that erupts quite soon. It’s really about the longer-term knock-on effects and what it means for some of the choices that are being made now.

My Bloomberg News colleagues yesterday had a great scoop where CENTCOM has requested Dark Eagle, the Army’s long-range hypersonic missile. They’ve asked for whatever exists to come to them. That doesn’t necessarily mean that’s what’s going to happen, but there’s this emphasis on trying to get all of these shiny toys, these next-generation technologies, the ones that haven’t actually been used in combat, and using this largely as a theater of experimentation if we want to use CENTCOM’s terms back at it.

But all of that has knock-on effects for readiness, preparedness for future conflicts, but also regionally. Right now, those would be taken out of INDOPACOM. And the things that China seems to care the most about, it’s things like that. It’s Typhon. It’s having missiles within range, particularly because in all of the war games that I’ve run, that I know Bryan has run, that matters because it becomes very quickly a war of missiles there. I think that’s why we’re just seeing so many choices that are being made now that just get me not only angry, but so nervous for what might happen in the future. And that’s not even talking about the fact that we are probably going to have a carrier gap in the future. That’s kind of bringing us back to the discourse of the early 2000s.

Eric Robinson: Hyperpowers have constraints too. And I don’t feel that advocates on Capitol Hill and the Pentagon and the White House necessarily operate under that understanding. It is a hard reality of contemporary warfare that there are only so many assets they have available, that there are questions of physics, of landing rights, of fuel capacity. The United States, for having $1.5 trillion in aspirational financing, doesn’t get to press the all button every single time. Eventually there are going to be trade-offs.

A theme that we’ve explored for the last 60 days is that we are expending exquisite assets, time, attention. We are accumulating friction, not just in terms of ordnance expended, but in just aircraft engines that are going to have to be refurbished and replaced. American capacity and capability to respond to other crises is necessarily being degraded by virtue of this exchange in Iran.

Bryan Clark: And the operational utility that these systems provide in this context is very limited. We’re using JASSMs to hit targets in Iran that could have been hit very easily with a GBU, or a JSOW if we really want to go fancy. And then Dark Eagle, same thing. What are we going to hit with a Dark Eagle that we couldn’t hit with any other munitions? We’re out there cosplaying so we can show off. They say it’s testing out this stuff in a real environment, but it’s not, because there’s no air defenses there that are gonna be meaningful. So you’re not actually testing it, you’re just showing it off.

Eric Robinson: Yeah, we’re like taking a Lamborghini to Dutch Brothers Coffee. We’re doing really mundane stuff with exquisite tools. There are weapon systems that are designed for Air Force or Navy pilots to get as close to an extremely hazardous situation and have a small likelihood of hitting their target. Those are the JASSMs we’re talking about. But what we’re doing is we are using those weapons systems against a country where their integrated air defense systems aren’t able to function. So this is overmatch and overkill, and it is using tools because it’s fun and exciting, not because it’s strategically apt.

Jordan Schneider: Are we 100% sure about that? Because right before this all ended, you had a few planes get tagged.

Eric Robinson: Yeah, it happens. F-29s are going down.

The Air Superiority Mirage

Becca Wasser: I think it’s really important that we take a grain of salt with a lot of the statistics that have come out of CENTCOM, the White House, the Pentagon. The initial numbers and metrics they use to demonstrate success in various areas — one, they’ve proven not to be true, but also some of it is just a fundamental misunderstanding of where the threats have been.

For example, a focus that emerged probably midway through the initial fighting of trying to sink Iran’s Navy. Iran’s Navy is not the biggest threat in the Strait of Hormuz. It’s the IRGC fast boats. It’s the anti-ship cruise missiles. And despite all of these efforts of going after various targets, the anti-ship cruise missiles were not number one on the list, even though the Strait of Hormuz has been essentially Iran’s biggest tool and biggest leverage in this conflict, because not only is it able to cause pain to its immediate neighbors, it’s able to cause pain to the US and to the global economy more broadly. So there’s a fundamental misunderstanding of what metrics are important, what target sets are most important, but also whether those have actually been degraded or eroded in a really significant sense.

But another thing we need to take into consideration is all of these claims of absolute air superiority. Despite all of that, I’m not sure the US has ever truly gained air superiority in the way that, frankly, Pete Hegseth and Dan Caine have suggested. Oftentimes when they’re talking about it, they’re talking about air supremacy. They’re trying to say that the US can act uncontested in Iran’s skies all across and it’s not a problem. But really what they’re generally talking about is the fact that the US has had more localized air superiority that is geographic and at times time-based, which frankly is something that we’re more likely to see in a future conflict with China where you have these windows of opportunity.

That has fed into some of the use of these higher-end munitions. I agree, some of it is the desire to show off Gucci instead of Tarjay, but really what I think it also is demonstrating is the fact that there hasn’t been this absolute air supremacy or even a higher level of air superiority to go after some of the targets that they wanted to.

Bryan Clark: I agree with Becca that they’re definitely overstating the level of air superiority they have. I think it’s partly also just an unwillingness to do the legwork to make it so that you can use a less expensive, more available weapon. You have to go do some suppression of enemy air defenses. You have to build a package that gets you in there so I can use a JSOW instead of a JASSM. So I can use a glide bomb instead of having to rely on a standoff missile. I can’t believe that you couldn’t do that against Iran, even with its air defenses somewhat intact, by just doing the blocking and tackling that we normally do — you have to include a suppression package in with your strike package. And they just don’t want to do that because they just want to hit as many targets as possible in a given period of time, which means launch a bunch of JASSMs and not have to worry about launching multiple aircraft sorties.

Justin: And that goes back to a combination of they are risk-averse at the command level, and they really have become risk-averse. And they also have to some degree forgotten how to do this. I don’t think the Air Force has forgotten how to do this, but at a command level, the way that they synchronize these — so that they can actually have a suppression of enemy air defense layered with a strike package, have the fast movers come in, do their strike targets, and then come out — they’re not willing to do that because they want to be able to say, well, we can do it whenever, we can hit them all the time. Well, you can only do that under two circumstances. You’ve destroyed all their air defense, à la the invasion of Iraq, or you are going to use exquisite weapons.

And we’ve seen a proclivity to use exquisite weapons against people who they’re not ideal for. If you just look at yeeting JASSMs at the Houthis. The Houthis survived Saleh for decades bombing them in Yemen because they just went into the hills and they stayed in the hills. And then they came out when he got done throwing bombs at them. So CENTCOM’s answer to the Houthis was, let’s throw a bunch of bombs at them. So the Houthis went into the hills and they stayed in the hills until the bombs fell, and then they came back out. Literally no change in their combat power. It’s what they do. Rinse, wash and repeat for the Iranians, who observed what we did in Yemen and went, okay, we’ll make our targets harder to hit. We’ll weather the storm and then we’ll pop back up when we need to.

Who Benefits From the Ceasefire?

Jordan Schneider: What about this whole month they’ve gotten to reconstitute? I don’t know how much Russian stuff has been shipped over the past few weeks —

Becca Wasser: I don’t think there’s even a need for external support. Obviously, Iran would like that. They would like to get some of the upgraded Shaheds that Russia has perfected and have used in the battlefield in Ukraine. They would like to get more sodium perchlorate from China. No problem there. But what they are doing, they’re digging out missile launchers, they’re digging out missiles, they are repositioning. This is just basic stuff and basic reconstitution.

I’ve recently been trying to think who does the ceasefire benefit militarily? Obviously it benefits all sides, and most importantly, it benefits the Iranian people who are no longer at risk of being targeted. Let’s put that one out there. But if you’re looking at the military ledger — the US is flowing in more forces, in part in theory to be able to reinforce the blockade. They’re requesting more forces. There’s more time to re-up some munitions. Cool. Israel’s probably trying to do the same, repositioning, give a little bit of rest to some of probably their pilots who have been doing double duty in places like Iran and to a lesser extent Lebanon.

But there’s no massive reconstitution that they can do of air defenses, of missiles. They’re probably trying to upgrade some of their older air defense interceptors, but they’re not going to be able to pull all that much off the factory line. Same thing with the US. It’s not like all of a sudden we can just, I don’t know, poop out more bombs. That just doesn’t work. No matter how much I think everyone would like that to be the case.

So if you’re looking at that, then probably Iran, which arguably has the lower bar for what it takes to reconstitute, is possibly on the up if you’re trying to look across the ledger. Because they’re doing the right things and they’re doing the smart things and they’re trying to do what they can with whatever it is that they have left. And if you also look at how quick they were able to reconstitute some of their forces after previous bombing campaigns, I think they were able to do it fairly quickly. Mind you, that’s not like six weeks. It was probably closer to six months. But there are some clear lessons learned there.

Justin: There was a quote — I can’t remember which book it’s from — but it was just pre-World War I. The Russians were suing to stop all armament advancement because they liked it exactly where it was. Rifles were good. 1905, that’s a good spot to freeze them. An armistice benefits you when you’re behind. Who was the one that was asking for the ceasefire? For all intents and purposes, it was the United States.

Who does it benefit? At the end, it looked like Iran was coming out on top because they haven’t given up on nukes. There was talk about relief of sanctions, allowing them to sell their oil, and then they were going to have a period of reconstitution.

The forces that the US is choosing to flow into the Middle East right now are interesting because yes, there’s an additional carrier strike group, but there’s also the 82nd. There’s also ground forces going in. Those are only useful if we’re going to use them. Sending the 82nd Airborne Division to places in the Middle East serves as either target, as a warning, or — we’re actually gonna do a ground invasion into a country that’s twice the size of Afghanistan and has a larger population than Iraq did.

Becca Wasser: I’m with you, but I also think that some of the logic has gotten screwy. If there were to be a resumption of US strikes, the time that would make the most sense would be when there’s still a three-carrier posture in the Middle East. You can have one in the Red Sea to hold the Houthis at risk, make sure that if they decide to go after shipping in the Red Sea or even Saudi’s alternate terminal at Yanbu, there’s repercussions. And then you can have one continuing to operate in the Arabian Sea / Gulf of Oman, prosecuting the blockade, and another contributing to broader strikes. You’d also want a three-carrier posture if you are thinking seriously about using your ground forces in any operation to forcibly reopen the strait.

But that’s not necessarily how they think. So there is a part where I’m thinking about these ground forces, and I think some of it is to have them pushed forward, so there is the optionality. But I also think there’s a strain of thought about them potentially being a tripwire, taking almost a page out of the European posture playbook where having forces that are there is supposed to deter further aggression on US partners in the region. That would make the most sense if you would see them in possibly one of the Gulf states. I think Kuwait would make the most sense given the existing ground force bases and infrastructure. So I don’t think it makes a ton of sense. But that is one way of thinking. And that I think also risks continued US ground posture at a bolstered level in the Middle East, which is something previous administrations had tried to push against.

Jordan Schneider: Previous administrations — you mean like this one? Three months ago?

Becca Wasser: Like this one.

Justin: That’s the other big thing. When we look across — the people who are in the policy position, the people who ran, the people who supported all said, we need to be able to pivot. We need to defend our interests in Asia. We can’t do that by continuing our excursions in the Middle East. We have to draw down. CENTCOM has been too big for too long. And now we’re in this twilight zone where it’s like circa 2003, only set to today’s music.

Pivot to Undisclosed Location

Becca Wasser: The pivot to Asia isn’t really working, but instead what we have is the pivot to undisclosed location in Southwest Asia for those of you who are old enough to remember.

Eric Robinson: Something I encounter in my professional life working around the defense industrial base is I often have clients or potential clients I encounter at conferences. They kind of operate under the assumption of, well, everybody’s really serious about PRC. They take the People’s Liberation Army Navy very seriously, don’t they? All of this has some sort of centralized coordination structure. Everybody thinks this is really important, right?

I don’t mean to be a professional cynic, but I will often use a little bit of dead padding to say, actually, no, not particularly. A pivot to Indo-Pacific Command has briefed well across multiple administrations. It was embedded in the 2018 National Defense Strategy and effectively abandoned in the more recent version.

It’s like the old Road Runner / Wile E. Coyote cartoon. There are people who really believed in this moment of United States industrial and military alignment to back Taiwan and to a far lesser extent back Ukraine in its war for independence. And they have left solid earth and they are still running out into space. They’re looking beneath them waiting for some sort of collective policy alignment by and between Republicans and Democrats. It simply doesn’t exist.

We have a series of operational-level spasms. We have random loans going to companies that might make sense. We have military operations against Iran or in Nigeria or in Venezuela that independently might make sense but don’t aggregate into a collective whole. So I think we’re in a moment of profound strategic drift, and I’m waiting for normies or just casual observers to catch up to that.

Becca Wasser: You think that the normies haven’t caught up to that at all?

Eric Robinson: Speaking from my lens in industry, people still think there is a collective vision around reindustrialization to take care of China, to make sure the United States can fight that war.

Jordan Schneider: I don’t think it’s normies, Eric. It’s people who have a financial connection to building for a Taiwan fight. There’s some motivated reasoning there within the China-watching community, as well as the DIB-for-Asia folks that still want to believe that it’s 2018 or even 2023.

Eric Robinson: Re-industrial archetypes. That’s a fair rejoinder — that rather than just normally relative average civilians, people with financial stakes in this did feel like there was going to be a generational commitment to reorienting American domestic spending, defense industrial policy, and the military with it. And I think there are segments of less ideological types, but still intelligent observers, who are recognizing that there’s no there there.

Jordan Schneider: This is a question for Becca and Bryan: how, emotionally, having done war games on both Hormuz and Asia, has it felt seeing these stocks dwindle in ways you guys perhaps more than anyone else appreciate the knock-on effects of?

Drone Wars Over Hormuz

Bryan Clark: We just did a war game looking at this scenario, the Strait of Hormuz scenario, just a month and a half ago, just as the war was starting. And it’s played out kind of like that war game played out — it turns into drone wars over the strait, but the strait closed most of the time. You just have to eventually wait it out until somebody wants to come to a resolution because there is no military solution. The strait kept getting closed by drones and mines. We kept cleaning them up. They kept doing responsive strikes against the guys on the shore on the other side. And it turned into a lot of drone-on-drone action, but nothing that really drove it to some kind of resolution. So it was not very satisfying, but illuminating. In terms of the current war, this is sort of what we found to be the base case.

Becca Wasser: That speaks honestly to why, in all of the fun financial projections that the smart economists I’ve been working with have been doing, our base case has been that this is going to be a protracted conflict, where you have this initial period of intense fighting, and then it becomes a much longer low-intensity conflict with periods of strikes and then rest, reconstitution.

This very much gets into the cyclical dynamics that we see in protracted conflict, both in the literature — for those of you who are nerds like me and think that Cathal Nolan’s Allure of Battle is one of the best books I’ve ever read — but also when we are looking at places like the current conflict that’s been ongoing for years because of Russia’s wanton aggression in Ukraine. Those are the patterns we see. And I think that’s what’s playing out here.

For me, the most emotional reaction I have to seeing how this has been prosecuted is thinking about the future, thinking about an America that’s going to be less secure because it can’t protect against some of the future threats that it and its allies might face. And thinking about a globe that is going to be a lot less secure as well. For the first time in my life, I am thinking about economics, looking at the economy and thinking about the downstream effects of that, not only for me as someone who wants to be able to afford things, but for society, for next generations, and back to Eric’s point about the defense industrial base and the massive amounts of money required to keep that afloat. This is going to be a generational change.

Bryan Clark: In the war gaming we’ve done looking at the Asia-Pacific or China scenarios, what this really highlights is that we need to think about how do you deter China on the cheap. Because we just couldn’t come up with this kind of munition usage and the demands from a traditional approach to the China fight. You just have to think about alternative ways of deterring China that don’t require you to somehow win a firehose competition with the PLA. That’s one thing this has driven our wargaming to look at: a lot of different concepts for how do you deter China without having to have this massive buildup, because you can’t trust that it’s going to actually come to fruition or that we won’t squander those weapons on some other adversary.

Becca Wasser: If I can take myself from being Wednesday Addams and gloom and doom and try and be a little bit more positive — it doesn’t come naturally to me, but I’ll try it anyways — one of the hopeful lessons learned that we’re going to take from this conflict is the need for lower-cost weaponry and effective lower-cost, attritable weapons. Right now, there’s a lot of patting ourselves on the back for LUCAS, which is a reverse-shot Shahed, and that we’ve deployed it in conflict. How? No one really has said. How many? Well, doesn’t matter. We might not even have any LUCAS left for all we know. But we’re patting ourselves on the back and saying that that is our example of low-cost, affordable mass. Yeah, it’s a lot cheaper than a lot of the high-end missiles that we have, but it’s not cheap enough.

I’m hopeful that one of the lessons learned that’ll come out of this conflict is not only this idea of how do you deter on the cheap with smart operational concepts, but how do you actually build to those operational concepts and get the costs down so that you have attritable weapons that can be used and that you can truly lower the cost per shot or even cost per effect.

Justin: To tie both of these points together — Cathal Nolan basically takes a part and looks at: hey, the majority of wars aren’t fought over the single battle. They don’t turn on the decisive fight. They’re generally wars of attrition. Even when Nolan looks at Waterloo, it’s like, yeah, but Waterloo took 14 years to get to. There was a lot of war before that that was attritive before you got to the final decisive battle.

I think the administration thought this was gonna be — they thought they were going to get in, hey, we took out Maduro. It was quick. We killed Soleimani. Nobody did anything. We struck the nuclear reactors. Nobody did anything. We can roll in and we can steamroll this and everything will be fine. Not realizing that this was opening up a different paradigm where it was going to become like, this is now a war. This is no longer discrete operations.

But one of the ways you make things cost less per shot and less per effect is you buy a lot of them and you build a lot of them. That’s one of the things that the administrations — not just this one — have been very reticent to do. We only need a stockpile of like 1,400 JASSMs. We don’t need any more. You make 40 a year, Lockheed? That’s awesome. Great.

And then we start using them and they’re like, oh, we need 10x the production. Well, that only gets you to 400 a year. And you’re using 400 in a month. The delta there is you get things to scale. That’s what drives down the price. You only get things to scale if you’re willing to buy them and fund them and keep refurbishing them. And they haven’t been willing to do that. Even when we talk about $1.5 trillion budget, we’re talking about one-time $1.5 trillion budget. Well, great — over the next 12 months, we’ll scale production, hire all these workers, build all these lines. Wait, no, that’s not what we’re gonna do. Because it takes more than a year to do all of that and to spend that money. Unless we have a much more integrated and forward-looking way that we’re gonna do the acquisitions, it doesn’t matter in the short term how cheap we get an individual shot.

The Stockpile Trap

Becca Wasser: That’s right. But one thing that Bryan and I have actually debated in the past is, yes, you need to be able to have the production capacity, because you need to be careful about what you stockpile and when you stockpile it. Some of it is shelf life. Some of it is just the shift in technology and how quickly that can change. Rather than just going all in on something that’s going to be completely OBE by the time you actually try and field it.

Eric Robinson: Somewhat satisfied that we are not sitting on a quarter-million Excalibur rounds in the United States, because it would have been extraordinarily expensive for the United States to build 155-millimeter artillery shells that are GPS-guided. And we would operate it to the assumption that we would have artillery batteries doing precision strike with wanton abandon in the Taiwan Strait gap, or we could give them to the Taiwanese to help defend the landing beaches. But we now know that these systems in their technological disposition are extraordinarily vulnerable to GPS jamming. They don’t have redundant navigation systems. To Becca’s point, and to build on Justin’s theme of you need to buy a lot of it — that’s absolutely the case, but obsolescence is extraordinarily hard to reconcile.

In May 1940, the French armies’ field artillery and their prime movers and their reserves of propellant, fuses, and high-explosive shells were the finest in the world. The French Army had spent 15, 20 years building that up. They had a better concentration, they had more professional gunners, spotters, and communication systems for their artillery than the next three armies combined. And in six weeks, that artillery was never able to move quickly enough to aim true and to break up the opposition.

So stockpiling weapons is sort of an economic imperative, but can also give you a false sense of security if you anchor your defense on systems that are no longer relevant.

Justin: I had a conversation with one of the consulting firms this week where they were talking about — they want to look at what are the components we need to put in Group 1, 2, and 3 UAS systems. And they were like, come on, tell us what kind of components they need. Well, what’s the threat? Well, that doesn’t matter. No, no, I think that matters a lot. They’re like, no, it doesn’t matter. Just tell us what components we should put in it. And I was like, I think you guys need to call somebody else.

Eric Robinson: Yeah, it’s like asking for a prescription without describing the malady.

Justin: Exactly. I think modularity becomes the key. It’s the ability to slap a cone on the top of the artillery round to make it more precise with whatever the next generation of that precision looks like. But you still need a lot of the artillery rounds. What is the artillery round? We can figure out what the technology is that slaps on top of it. What is the Shahed of tomorrow or the LUCAS of tomorrow? Some of those are gonna cost a lot more and some of those are gonna cost a lot less. If you’re using it on boats in the Caribbean, you probably don’t need the ability to be EW-hardened to the level that it needs to be to fly in Ukraine or off the coast of Taiwan.

The IRGC’s Hardliners

Jordan Schneider: Maybe this is one for Justin: what percentage of the IRGC would be thrilled to have the 82nd Airborne fly on in?

Justin: When you look at the IRGC and their leadership, you’ve kind of got stovepipes. I saw somebody the other day was trying to make this reference that Iran only spends 2% of its GDP on defense. And the implication was that they’ve been able to defend against the US only spending 2% on defense. That misunderstands how the IRGC and the Iranian military are bifurcated and how they actually operate. So yes, only 2% of GDP goes to the actual Iranian military. Then there’s this whole other thing with these hardliners that go out and get to operate and kind of run almost like a criminal cartel where they own construction companies and shipping companies and all kinds of other things that they get to draw money from.

Some of those people will not want anybody to invade or any type of war because they just want to keep making money. They are comfortable owning the concrete company in Lebanon or owning whatever business they’re using to generate wealth and revenue.

But there’s also the Shia martyrs. People struggle with, are they true believers or are they not? But I’ll say this. Imams and ayatollahs and Shia clergymen in the ‘80 to ‘88 Iran-Iraq war, when Iraq was driving tanks into Iran, were walking around handing people plastic keys saying this is the key to heaven, this is the key to the kingdom of heaven, while they strapped on suicide vests to go run at Iraqi tanks and blow them up. There is a portion of the IRGC that are hardliners and they are believers. They are the people who still go and clean off the martyrs’ tombs and they tap on the tombs so that the dead can hear them and know that they’re there. There is a very real undercurrent in parts of the IRGC that would absolutely relish the chance to become martyrs and to take down — like, absolutely. Is that their leadership? Debatable. But you have to really look at what the IRGC is and what they’ve done in the past and where they came from to understand: when people say there’s a group of them that are hardcore true believers, those people absolutely exist.

Eric Robinson: I also don’t know how the Iranians dial in escalation dominance. There’s always going to be a segment of — whether they are religious fanatics, they’re Marxists, they’re Christian nationalists — who will embrace, the worse it is, the better. They think that if you can ratchet up the violence, you can gain a longer-term political objective. I don’t know that that logic holds here.

The Iranians, for all the short-term perhaps excitement of being able to grab an American paratrooper battalion by the belt and start to get at them in a slugfest, recognize that if the United States starts taking serious casualties, this administration has few reservations about committing atrocities against Iranian civilians.

To Justin’s point, there’s sort of a mosaic of reactions, and witnessing a consolidated reaction from inside the Iranian security state that speaks for all elements is unlikely. But I also think they are sufficiently sophisticated to recognize that if you fall back to theoretic escalation dominance, they don’t necessarily have the kind of tools that would wake up the Secretary of Defense, and that they may be subject to extraordinary violence against national-level infrastructure that they cannot account for.

Becca Wasser: That’s why we see constant hedging strategy from Iran. They are willing to engage in diplomacy. They are willing to negotiate. But at the same time, they are willing to fight as hard as they need to, because this is an existential conflict for them. If a deal is offered that is attractive enough, some factions in Iran are more than happy to accept it, and perhaps that is the leadership.

But the one constituent group that we don’t hear from are the Iranian people, in part because there are these internet and technology blackouts. And honestly, they’re the ones who are most at risk from potential threats to wipe out massive infrastructure or civilizations, if we want to quote Trump’s Truth Socials of yore. They’re the ones who have to bear the effects. And frankly, with the blockade, they’re also the ones who are probably going to bear the continued economic hardship. But the Iranian system and the leadership that exists believes that they can ensure that the people will fall in line as needed, in part by brute force.

So I don’t think it’s necessarily a great path forward. But the big thing is Iran’s entrenched. They are dug in and they are willing to see this through. This goes back to all of our discussion about why we think this is likely going to be a long war.

The JCPOA Lesson

Justin: Eric made this point about Libya and the lessons learned from Muammar Gaddafi. I also think there are good lessons that were learned from the JCPOA, because the more reformist-minded kind of got control of the government to some degree in Iran. They were able to wrangle through the JCPOA, which was going to limit the growth of military power towards nuclear ambitions.

To the IRGC — they won, they got some sanctions relief, they got some money. It started to look like things were gonna open up, which ideally would have in turn allowed more opening up and more reform. And then when that got pulled away, the hardliners can look at that and go, see, we told you, you can’t trust them. They gave you something, they got us to agree, they got us to give up all of our highly enriched uranium to Russia, and then they pulled it out from underneath us. Why do we make a deal the next time?

Eric Robinson: If you elevate this to traditional prisoner’s dilemma or game theory, the opponents of the United States can sort of assume with some basis that the United States is always going to defect. Always. And you need to forecast what the defection means, how you prepare to limit the damage.

Becca Wasser: That’s a lesson learned, frankly, for adversaries and allies alike. We talked a little bit about posture and CENTCOM, but we have some potential threats going on in Europe right now when it comes down to US military posture threatening to pull out troops in Germany, Spain, other places in punishment for what they’ve done. So I think the idea of America as being reliable and a reliable ally, or at least a reliable country to negotiate with or strike a deal with — I think those days are long gone.

Jordan Schneider: Say Europe was all in — we’re gonna crash this strait open by whatever means necessary. Does that change the balance of forces at all?

Becca Wasser: I don’t know if it changes the balance of forces in the traditional sense, but I cannot see Europe being willing to commit any significant naval power without the strait being secured. They’ve been so clear about that. The area where I think it would probably make the biggest difference is — there are a number of European countries and frankly Asian countries that have more minesweepers than the United States, because the US divested of them. For the most part, what the US has left is a bunch of littoral combat ships that they couldn’t find an actual role for, so they outfitted them in a minesweeper capacity rather than sending them to get scrapped. Having that minesweeping capability would be really useful. But I just can’t see any European country want to contribute that or any type of offensive naval power and do things like escort missions or contribute to the blockade in a really meaningful way, unless maybe the US put the screws on them further.

If anything, I think maybe we would see a re-up of what they’ve been doing in the Red Sea, maybe plussing up there and saying, we’ll hold this down and you can focus over there. That would be a smart way of playing it, but I find it hard.

Jordan Schneider: I guess my question was — if you wave a magic wand and you get to do whatever you want with all of their assets, does that actually change the fact that Iran can still hit one in 20 tankers that go through and that means that nothing goes through? I don’t think so.

Eric Robinson: Iran has a fleet in being. They have anti-ship cruise missiles. They have an uncertain number of mines that they can employ. They have asymmetric tools. They’re going to keep insurers nervous. They’re going to keep mariners on edge. Just by virtue of their geographic proximity, they will remain dangerous.

Justin: This goes back to that conversation we were having about that terrible FT piece — the idea that the Red Sea was closed for almost all of 2024, which was made without actual evidence by the author. That is not true. But because there was a route around, there were some shipping companies and insurers who were like, just go around. Why would we risk it? You don’t get that here. So why would it change it? Either they make a separate peace with the Iranians, which won’t be able to hold up because they’d have to bring the US along, or they have to join and turn the coast of Iran into rubble, which we’ve talked about why that’s impossible several times. It’s gonna be a long one.

Becca Wasser: To Eric’s point, Iran doesn’t have to do that much to cause havoc. And it’s not just about the missiles. We saw what the Houthis were able to do with just a few drones in the Red Sea. And everyone is so much more locked into the strait because there’s no other viable alternative.

I keep getting asked questions about, what’s the historical precedent for this? What happened when the Suez Canal was closed? Is there something we can learn about Hormuz being closed or what happened during the tanker war? And none of those are applicable. The geography is very different. The time and the technology is very different. With the tanker war, the US was willing to escort ships, which is not something they’re willing to do at this juncture. And that’s not something Europe is willing to do either. So all it takes is just a little bit of being annoying with drones, and that just shuts it all down.

Jordan Schneider: Becca, what a treat. You’re welcome back anytime.

Eric Robinson: Thanks everybody.

Our conversation covers:

• What it takes to build a quantum computer — Inside the cryogenic supply chain, the helium-3 bottleneck, and why mining the moon might actually make sense.

• How export controls backfired — How restrictions on dilution refrigerators helped spur China to go from zero to more cryogenic suppliers than the rest of the world combined in just two years.

• The scaling problem — Simply multiplying dilution refrigerators doesn't get you to a million-qubit machine. Cooling, cabling, and the chips all have to be rethought — and no country owns that yet.

• Why being first isn’t winning — Why long-term victory isn’t cracking Shor’s algorithm first, but locking in supply chains across multiple modalities.

• The public-private fault line — The high-stakes balancing act between the government stepping in to accelerate innovation and letting the market work on its own.

Plus, what China is getting right, where the US still has an edge, whether the US should ban Chinese components, and why quantum supply chains are a national security priority.

If you read ChinaTalk, you care about what’s actually happening in China. The people who understand China most deeply access it directly — reading the news, hearing the debates, talking without a translator in the room. Without Chinese, you’re always one step removed.

That gap is what my old internet friend has spent two decades closing. He built a fantastic substack RealTime Mandarin — a must-open for me each week — for intermediate and advanced learners. Now he’s made something for beginners!

Interactive Mandarin for Beginners is a 12-week program covering characters, vocabulary, grammar, reading, listening, and speaking, all through real Chinese news stories. You’re learning with real Chinese from day one, alongside a community of fellow China watchers. It’s for anyone who’s tried for years and never found something that stuck.

Enrollment closes this Saturday night. Check it out here!

Jordan Schneider: Constanza Vidal Bustamante has done dramatic, impressive work of public service, writing one of truly the best think tank reports I have ever come across: “Quantum’s Industrial Moment: Strengthening US Quantum Supply Chains for Scalable Advantage,” co-authored with John Burke. It goes incredibly deep, and I learned so much about everything that goes into making a quantum computer.

It really reminds me, Chris, of reading those 2018, 2019, 2020 reports where Washington was wrapping its head around the semiconductor supply chain — the work that ended up delivering what became the CHIPS Act and the program office. There is an enormous amount of detail and knowledge here. Every few sentences, I found myself wanting ChatGPT to give me the ten-page version of some three-sentence reference Constanza made. We are excited to give you a taste, but you should all read and dive into the full report.

What Does it Take to Build a Quantum Computer?

Chris Miller: To start, what does it take to build a quantum computer?

Constanza Vidal Bustamante: That’s a big question. In the report, we try to answer this complex question because it seems like it would have a simple answer, but it’s quite complicated — starting from the fact that there isn’t just one kind of quantum computer. Different companies are pursuing different modalities. We have superconducting computers, atomic computers (which could be neutral atoms or trapped ions), photonic computing, and many other modalities cropping up as well.

Each has a different bill of materials, pulling from various layers of the quantum supply chain in different ways. Some of these are partially overlapping, but they’re distinct enough that it gives rise to the idea that there isn’t just one supply chain — we have multiple supply chains we should be taking care of.

In terms of commonalities, they’re all drawing from similar layers of the so-called quantum stack. They draw from specific materials, or they may use distinct atomic sources or isotopes. They place these elements within an environment — a cryogenic environment with ultra-low temperatures or an ultra-high vacuum environment. They use different components to interface with these atomic sources or materials that generate the quantum state, such as lasers or various control electronics. There’s a software layer, and there’s networking, if you think a little further into the future, as we start putting together different chips for these various modalities.

All of these modalities draw from these various layers, but the specific elements that go into each layer vary quite a bit. That’s why it was very important (and why the report runs on the longer side) because there isn’t an easy answer or just one list of elements you can point to. These things are also changing over time, so it gets complicated.

Zachary Yerushalmi: Quantum is at a super early stage as a technology package. We are pre-transistor. Because of that, you have to deal with the inherent uncertainty of supporting all these different supply chains in their current state. But it’s also a wildly fast-moving industry. The next phase will require a step change and a reinvention of that supply chain, even if a lot of these existing modalities are successful.

The example there would be something like photonic integrated circuits, which are the photon equivalent of the integrated circuit of the electron era. Right now, most AMO — atomic, molecular, and optical — approaches, which represent one of the big clusters in quantum computing, are using methods that don’t scale based on current manufacturing techniques. To scale them, you have to move to PICs. To move to PICs, yet again, you need to reinvent the supply chain and do that continuously. It’s a fascinating one to grapple with, with a level of uncertainty that I really don’t think we see in any other technology package at this scale.

Chris Miller: The analogy is sort of like it’s 1945, we’re two years out from the transistor having been invented, and we’re trying to think through what the computing supply chain looks like in 1955. We don’t know what the transistor is going to look like exactly, so we’re going to go through the cabinets at Bell Labs and figure out, on average, what the scientists are using as they run their experiments. Is that a good analogy?

Zachary Yerushalmi: Yeah, almost to the point of pre-vacuum tube. I’d be curious about Constanza’s perspective.

Constanza Vidal Bustamante: When we think about the heterogeneity of supply chains, it’s not just across these modalities horizontally, but also along the time dimension. As we think about the prototypes being built right now, we have a good sense of what those supply chains look like. They’re very globally distributed, and we can point to sources of dependencies — some things we’re importing from China, where it may be the only source, or other things where the best in market comes from Europe or Japan.

But as we look ahead to when we’ll have quantum computers capable of breaking encryption — the version of these machines that will truly be revolutionary — the supply chain is probably going to look quite different from the one we have right now. As I argue in the report with John, when you think about geopolitical stakes and international competition, that’s the place where the United States can still dominate, because nobody has control over a supply chain that doesn’t yet exist.

If we think carefully, we’re not entirely without an idea of what it will look like. As Zach said, we have models. We need to move to photonic integration if we want to manufacture this at scale and at a competitive cost. To actually build these machines at volume, we have a rough idea of the path to get there. It’s just a matter of breaking the chicken-and-egg cycle of waiting for enough market demand before making major investments in the supply chain. Because you don’t have those investments, you never get to a point where the product becomes very attractive to the market.

There’s a path forward. It’s just a matter of gathering enough momentum, political will, and capital. At the end of the day, it’s capital. We need to unlock the next-generation supply chain for these machines, and the United States is definitely still in time to dominate if we move quickly.

Zachary Yerushalmi: Let me give a couple of examples on the stakes of locking in your role as a country in that supply chain, and why you get so much leverage when you do.

Think of the two dominant approaches. One is solid-state or superconducting, which requires cryogenic systems of a wild scale. The other is AMO. Take the cryogenic side. For innovation there right now, you need a dilution refrigerator to operate these systems. It takes 40 hours to go from room temperature down to the level of cold needed to operate a superconducting circuit. That 40 hours means you can only run one test a week. If China invents an ability to take that from 40 hours to 12 hours, you go from one test a week to one test a day. Your iteration cycle changes completely, and they’ll lock that down and grab that supply chain.

On the other side, with PICs for the AMO approaches, nobody has really made a scalable PIC — the architecture transistor part for that computer. It’s really hard. The country that does that has a total lock on the ability to scale whole approaches in quantum computing. That actually reads across to quantum sensing as well, because AMO and quantum sensing approaches are pretty similar.

Subscribe now

The Cryogenic Supply Chain

Chris Miller: We’ve got these different qubit modalities, which are sort of like different transistor structures, if that’s our analogy. We know the supply chain underneath them has some similar parts. We can start with cold temperatures, since they were just mentioned. Constanza, what does the cryogenic supply chain look like today? Also, give us a glimpse as to how cold we’re actually talking about.

Constanza Vidal Bustamante: Very, very cold. Even within cryogenics, it gets more complicated because some modalities require millikelvin temperatures while others operate at cryogenic temperatures over one Kelvin. That doesn’t sound like a huge difference, but it’s actually quite substantial in the energy requirements and the specific components or subsystems that produce those temperatures. Those are almost like another fractioning of the supply chain.

For instance, the superconducting modalities we were talking about — the computers that companies like IBM and Google are building — require those super shiny, chandelier-like dilution refrigerators usually portrayed in the media whenever there’s a quantum piece published. You can dig further into what it takes to make those refrigerators. For photonic computer modalities, some subcomponents also require cryogenic temperatures, but not low enough to require a dilution refrigerator. That leads to other complications, which we can talk about.

In the dilution refrigerator camp, there are a few different issues, starting from the fact that they use — and this is perhaps more well known — helium-3 as part of their cooling approach. Helium-3 is an extremely rare and highly regulated isotope that you can’t simply build or supply on demand. It comes as a sub-product of nuclear processes. That seems to be an area where, if you start developing machines at scale and you need to access a supply of helium-3, you could find a choke point.

Jordan Schneider: But there is some on the moon, right?

Constanza Vidal Bustamante: I don’t think it’s questionable that there is helium-3 on the moon. The question is whether it’s ever going to be feasible to extract it. Zach, you’ve looked more deeply into this. Maybe you can join me on the lunar sourcing option.

Jordan Schneider: Look, if we’re going to put data centers up there, I feel like a little bit of helium...

Chris Miller: Let’s step back and say what a dilution refrigerator is, and how they actually work, before we get to the moon?

Jordan Schneider: Okay, it’s a teaser.

Chris Miller: Before we get to the moon, walk us through how these machines work. We’re getting as cold as outer space. What does it take to make a machine that makes things that cold?

Constanza Vidal Bustamante: At a high level, they take several different stages to get there. You don’t go from ambient temperature directly to extreme, colder-than-outer-space temperatures. If you look inside one of these chandeliers, they have several cooling stages that step down progressively. It goes down to maybe 4 to 10 Kelvin first, and then continues down. The bottom is the coolest stage, where you place the quantum chips for superconducting or semiconducting spin modalities of computers. That’s the actual coolest part — the part colder than outer space.

To achieve this, you require a combination of helium-3 and helium-4. Helium-4 is not really a source of concern. It’s the most common helium isotope, so it’s not a supply bottleneck I’m aware of. But the helium-3 part of the mixture is the absolutely necessary element to get you to the millikelvin temperatures these systems require.

Zachary Yerushalmi: The whole point of building a quantum computer — and why it’s hard — is that these quantum states are incredibly fragile. They get messed with by everything. Heat messes with them. The wider environment messes with them. Cosmic rays mess with them. Looking at them messes with them. That’s the whole point of quantum mechanics.

What you have to do is isolate these quantum states from absolutely everything. The most effective way to do that is to get them wildly small and wildly cold. When you get them wildly small, a different type of physics takes over that enables you to manipulate these systems in such a way that you can do useful calculations.

On the helium-3 side, this is one of those things where I really wish a quantum computer made going to the moon economically viable. The sad thing, particularly for America, is that the major supply of helium-3 is tritium decay from the nuclear stockpile. As long as we don’t go nuclear-free in the US, from most of the calculations I’ve seen, we should be okay.

That said, access to these systems — not just ones today, but ones that actually enable that scale — is critical. There are three credible suppliers in the West that can supply these — Bluefors, Oxford Instruments (which just got bought by another company), and Maybell in the US. It’s that hard — there are only three companies, and really only two of them are credibly there at scale.

China went from having none to, just in the last couple of years, creating more companies building these systems than the rest of the world combined. They went from not publishing in this space at all to dominating over 50% of the publications on new innovation in this area. It takes decades to get good at this. Folks are coming up the chain very quickly in places where the Finns will share their dilution refrigerator IP with us. China is not going to do that.

Constanza Vidal Bustamante: To Zach’s point about China announcing all these different manufacturers of dilution refrigerators, some people point to the export controls that the United States, along with several other international partners, put in place starting in 2020 and into early 2024. Within a year or so, this seems to have backfired. With those export controls — of which dilution refrigerators were importantly a part — you accelerated an ecosystem where China rapidly mobilized to procure their own systems and continue innovating on the computing front.

Chris Miller: I’d love to dig into that as well, but I’d like to go to helium first. Helium is on the moon, and we’re going to mine on the moon, maybe—but it also comes from the nuclear stockpile. Is this from civilian energy production, nuclear weapons, or a mix of both?

Zachary Yerushalmi: Nuclear weapons. Tritium decay. That’s the majority of the source. You can get it from a couple of other places. Evidently, Canada has loads of helium-3 randomly stored away.

The reason there’s lots of helium-3 on the moon is that cosmic radiation strips away helium-4 and converts it to helium-3. Unfortunately, you need to launch a rocket there, harvest it, and bring it back. The only economically viable use of lunar helium-3, from what I understand, is if you need to go to Mars and build quantum computers. You launch off from Cape Canaveral, get to the moon, and if you’re still going with Elon Musk on board, then you have an awesome business. But if we’re focused on the helium-3 supply for the US and keeping it tertiary, I’ve heard a bit of skepticism around lunar mining. I’m bummed, but sadly, we are where we are.

Export Controls Backfiring?

Chris Miller: We know that things like dilution refrigerators are hard to make — which is why there are a small number of companies, and you need to mine helium on the moon or do something comparably difficult. On the other hand, there’s an argument that the export controls the US and Europe put in place on dilution refrigerators a couple of years ago spurred this brand new industry in China. That suggests it wasn’t actually that hard, or at least that the response happened pretty quickly.

Help us understand how we should think about this case study. Does it tell us anything broader about the relevance of export controls in the quantum computing space?

Constanza Vidal Bustamante: It really depends on the specific inputs we’re talking about and the timelines related to the volumes at which you need them. What Zach was perhaps trying to say is that we shouldn’t worry too much about helium-3 in the near term. At the rate we’re building refrigerators and the rate they’re being purchased and acquired, we’ll likely be fine for the next few years. Luckily, the US has a big source for that, so we’re in a privileged position as the provider for much of the world.

But going back to what we were saying earlier about the next-generation supply chain — as we start scaling these systems, you no longer have just one chip to cool. You start building machines that require what becomes almost a side problem of cryogenics — how much dilution refrigerators can scale to support much bigger qubit counts. Once you have a lot of demand for many large systems, I start worrying about whether the sourcing of helium-3 or the refrigerators themselves can keep pace with that demand.

Going back to the China example — I don’t think they’re building these machines at volume yet. It certainly doesn’t seem to be the case that they’re selling these machines beyond procuring them for their own experimentation within their top-level academic research labs — those at the frontier of hardware development for quantum computing. They were able to develop these machines for maybe one or two systems, possibly more, but definitely not in the hundreds yet. They made enough to continue progressing on prototyping and iterating, but I wouldn’t say they’ve reached the level of Bluefors in Finland or Maybell in the US.

That points to a story where the controls accelerated their start in developing these machines in-house. Maybe they weren’t planning to build that domestic capacity quite as quickly, but they were pushed to do so by the controls. They haven’t yet reached the stage where they’ve equaled what Western companies can make, but they seem to be on that trajectory. You can still question whether it was the right time to put controls in place on those.

Zachary Yerushalmi: The China anecdote ultimately boils down to the stakes of this industrial competition — both how high they are and how different they are from other technology packages, because we are so early in that race.

The US actually has an incredible moat around semiconductors. That doesn’t mean we can sleep on it, but we’ve been doing that for decades. We have friends, partners, and allies all across the world. Because we haven’t built a fault-tolerant quantum system, a commercially useful quantum system, we don’t have the same moat. That means China gets the ability to leapfrog and reach near parity with the US on certain manufacturing capacity. As you add in friends and allies, it gets closer, but if you look forward, the stakes are big.

This also hints at something Constanza spoke to in the report. We need to rethink how we do the supply chain to get to real scale. If China is the country that comes up with the intellectual property on the core method to reach that scale — if they invent the kind of transistor of the scaled cryo system you need — then they will have an unfair manufacturing advantage. China is typically quite good at that. They’ll also have an unfair IP and understanding advantage on the key path you need.

We have to think differently here. If we just project forward the existing engineering design of these subscale systems, we will not have enough helium-3. That’s why we have to reinvent the systems that make these computers, the QPUs (quantum processing units), really small and actually get them to that temperature. We have to rethink that process. The country that innovates and locks that down — that holds the manufacturing intellectual property — will have an unfair advantage to win. We just don’t have the decades that we rest on as an advantage in semiconductors.

Constanza Vidal Bustamante: In the report, we elaborate on exactly this point. In the near term, the most advanced dilution refrigerators available on the market can host around a thousand qubits. If you want to get to machines requiring a million qubits or more (though qubit count isn’t the only metric to consider), the path we have right now is essentially to put together dozens of these dilution refrigerators.

But the scaling doesn’t quite work that way. As you add more qubits, at least for the superconducting modalities that would require them, you need cables to connect the qubits together, and that adds to the heat load. That makes the cooling less efficient. It’s not as simple as multiplying the refrigerator by X number. You need to do what Zach was saying — innovate so the cooling approach you’re taking is much more efficient at scale.

That’s where we’re seeing real activity. Maybell just put out a new system at APS this week. I haven’t looked into the details yet, but that’s where we need to focus a lot of attention, as Zach said, so we don’t get out-innovated in this space. Otherwise, it becomes much easier for countries like China to reach that scale before we do.

That’s the challenge. We need to focus both on the near-term supply chain to continue iterating and innovating, while keeping a very strong eye on what comes next. That’s where we will reap the most reward in terms of economic and security benefits from the utility of these large-scale machines.

Policy Recs for Quantum Success

Zachary Yerushalmi: If you could talk to policymakers and give them suggestions on what they can do — the policies, the tools they can adopt to give the US the best shot here — what comes to mind? What’s the strategy to win on supply chain?

Constanza Vidal Bustamante: This goes beyond cryogenics, which is the subject we’ve been discussing. The report tries to be comprehensive in its assessment of the problems, but the solutions we provided are preliminary and need a lot more fleshing out. Maybe I’ll do subsequent reports, putting much more detail into what the solutions could look like.

Broadly, for the cryogenics problem, we’re calling for intentional and targeted multi-year advanced R&D programs on cryogenics. Similar dynamics apply for highly precise laser systems and other optical components, where the systems we have right now work for the prototype machines we’re building, but we know we need to keep innovating to reach utility-scale machines.

This is an R&D tool, but it’s not just fundamental R&D. Given the race dynamic and the time-sensitive nature of this, it needs to be a dedicated, advanced R&D effort. Another big point that cuts across the report is bringing together the enabling technology manufacturers — in this case, the companies building dilution refrigerators — with the end users, the system integrators in the quantum world. We want the computing companies that will use these machines to co-design, where possible, getting down to the specific requirements these machines will have. That accelerates the process rather than just building and hoping the result will be useful.

I’m less worried in the cryogenics sector that this isn’t happening already, because the market for these machines beyond physics research or quantum computing isn’t that diversified. They’re definitely thinking about quantum as their primary sector and paying close attention to the requirements. But for other components with broader markets, you have to be very deliberate from the government perspective when setting up R&D programs to ensure the enabling technology manufacturers are closely aligned with the needs of the quantum end users.

Share

Zachary Yerushalmi: I look at three levers. There’s supply — do I have the widget on the shelf when I need it? That’s the current widget. There’s innovation — do I have the support to skate to where the puck is going in the industry?

The last lever I think of is capabilities that exist in a market failure. The canonical case there is high-mix, low-volume fabs, like what you see in the semiconductor era. At the intermediate volumes you need, there’s an explicit market failure in running those fabs given their cost structure.

To make it specific, take the fab we have at Elevate. It costs about $40 million in capital equipment. Every year, because we focus on a particular level of TRL-ness, we hope and pray that we make about a million dollars a year on it. That sucks — no investor is going to give you $40 million and hope you make a million dollars a year. Governments have to think about supporting that long-term market failure in order to maintain that industrial capacity.

Constanza Vidal Bustamante: I focused on the R&D lever because it was most pertinent to what we were just discussing about next-generation cryogenics. In the report, however, we provide a menu of different policy actions that can be taken to support various elements of the supply chain, depending on the specificity of the issue at hand. Different problems—and different levels of maturity in the components or systems involved—will require different levels of support, and the federal government may be more or less well-suited to take action in each case.

There’s definitely no one-size-fits-all approach. The supply chain is so heterogeneous that it would be very surprising if any single intervention solved the entire problem. Some issues will require less federal activity than others. Helium-3 is a good example of where more intervention is warranted. It’s a highly regulated isotope, the private sector isn’t going to be the right actor here, and the solution can be as cleanly structured as having the isotope program under the Department of Energy take a close look at their inventories, set aside parts of that inventory for quantum needs, and do the right calculations for repurposing some of the helium-3 already in use. They could also think through some out-there ideas for new sources of helium-3, but in a deliberate way.

That’s a very specific example. Others require a completely different scale of investment — for instance, what’s needed to make some of our current foundries quantum-ready or quantum-grade. There you need to call on multiple actors to play a role. There’s a wide range of tools we can deploy depending on the specificity of the issue at hand.

Zachary Yerushalmi: My sense is that governments can either make markets or distort markets. Do you have a North Star or heuristic for when government intervention is needed and when you should let the market do its work? Big question, but it’s so pertinent here.

Constanza Vidal Bustamante: A big piece for me is not whether the private sector could eventually do it. It’s when you put all of this under a geostrategic, geopolitical race dynamic where it’s time-sensitive, you don’t want to wait. If you pressed me, I would say sure, let the market take care of it and figure out which modality is best. Whichever has the most manufacturable supply chain and relies the least on highly vulnerable items should be the one that wins.

But if we believe there’s a higher-priority objective — where we don’t want to be second to anybody, especially China — then we’re under a very different set of circumstances. Every day matters, and we want as many modalities as possible for the US to dominate. It’s not just about supporting whichever is most promising. Let’s say superconducting wins — we don’t have a great definition for what “winning” means, but say a superconducting machine is the first to break Shor’s algorithm, and it’s a US-based company. Even at that point, I wouldn’t call victory. I would still want the other modalities to dominate in their respective categories.

It’s very plausible that a different modality — say, photonic quantum computing in China—will also clear that bar, and they may have figured out a supply chain that’s more nimble, cheaper, and more cost-competitive. That would outshine the superconducting machine that the US got across the finish line first. The finish line is moving, so it’s all hands on deck. When you start thinking about those circumstances, there’s a big role for the government to serve as an accelerator of that market. That’s why I think of all of this in terms of a broad innovation and industrial policy portfolio — because that’s the scenario we’re in.

Zachary Yerushalmi: I love this point. Two things occur to me. First, if we got to vacuum tubes as a nation and said, “This works, this is good enough, down tools,” you’d miss out on the transistor. That was actually pretty important for scaling these systems. It’s a repeat game.

The thing I do worry about is that the stakes of getting policy wrong are wildly big. The example that comes to mind is China itself. There’s a technology area called quantum key distribution. We don’t need to get into the technology of it — folks can look it up online. It’s really cool math. Unfortunately, the math is so cool that if you do your postdoc on it, you just want to do that math all day, and you forget that it’s economically and cryptographically not all that secure.

Because the head of China’s quantum program, Pan Jianwei, is obsessed with this, he puts a wild amount of resources toward it—even though you can literally just look up “NSA QKD” and find intricate detail on why this whole thing is dumb. The upshot is that we need industrial policy because the competition is so intense, but it’s very easy to get wrong. We just hope our competitor gets it wrong more than we do.

China’s Quantum Approach

Chris Miller: On that point, someone recently made the following analogy to me: China has a Manhattan Project for quantum — one plan, one team, one system, and most of the ecosystem oriented around that particular pathway. Whereas in the US and the West, you have these different qubit modalities and different companies competing with each other, and as a result you have somewhat distinct supply chains.

Is that analogy true or false? And if so, who’s got the better strategy?

Constanza Vidal Bustamante: That analogy used to be true, but it’s changing rapidly. The comfortable narrative we had about China for a while was that they’re undoubtedly leading in communications. They’ve deployed large-scale infrastructure, optical fiber, and quantum key distribution systems to exchange keys in a supposedly tamper-free way. In addition to the fibre that they’d deployed over something like 10,000 kilometers in China, they also have some quantum satellite link demonstrations. It sounds very impressive.

But the assessment from the West was that even though they’re leading, at least in deployment of this technology, this isn’t a technology we care about or believe brings a lot of value. It’s a very narrow solution. It’s not a full cybersecurity system in the sense that you still need a lot of classical encryption and authentication systems for other parts of cybersecurity. Even for the piece it does cover, it’s not fully secure. You can hack it in different ways. China can take that piece, and we don’t care about it.

In computing, the narrative used to be that they’re catching up quickly, but, like Chris was saying, they’re really only putting a lot of their chips on superconducting. They’re moving quickly, and they’re impressive, and we should watch them, but they don’t have the diversity that we have. Just in the last year or two, however, we’re seeing a lot of startups appear in China, often led by prominent academics from Pan Jianwei’s group or others who lead quantum research in China across different modalities.

They announced two different neutral atom computing companies last year. They have some photonic ventures — a photonic company has been prominent for a while. They’re growing the number of superconducting ones. Recently, I read about even topological qubit developments. All of these are new companies. With the information we have, they’re probably not very close to matching the capabilities of the various computing modalities we have in the United States, but there’s definitely rapid movement. It’s not just that all of these are state-driven and therefore won’t be effective — these are coming out as private startups from highly talented folks.

We should worry about that. We shouldn’t just rest on our assumption that they’re limited in what they can do.

Zachary Yerushalmi: This is one of the many reasons I think Constanza’s report is literally a national security priority. The reason China can move up the chain so fast is that they’re so thoughtful in their approach to the supply chain. If you have the key components to manufacture all of these different modalities — all these approaches to building quantum computers — then regardless of what you learn about which approach is better, you can react quickly and deliver against it.

We spoke before about photonic integrated circuits, critical tools for scaling these systems. In the US, even for some of the biggest providers, because they don’t have access to the fabs and the supply chain to manufacture those, it can take 12 to 18 months to go from an idea like “I want this new PIC” to actually getting your PIC. In China, because they’ve really invested in this area — it’s used across many different applications in photonics and certain material systems — you can go from “that’s a cool idea” to having your PIC in literally two weeks.

A lock on the supply chain is a gift that keeps on giving, because you can be literally ten times as reactive and adaptive as your adversary. It’s like a supply chain OODA loop of sorts. Nobody has been attuned to this the way Constanza’s report has captured.

Constanza Vidal Bustamante: Two other things came to mind as we were speaking — one for China and one for the US.

For China, in addition to what I said earlier about prominent scientists starting their own companies across different computing modalities, what is true — and what Chris was alluding to in the Manhattan Project analogy — is that China has been deploying moonshot programs to a much greater degree than the United States has. This cuts across different levels of government. A lot of the provinces or local governments frequently launch moonshot programs where they say, “submissions accepted: by 2026, create a dilution refrigerator that can host a thousand qubits with these error rates.” Those targets are usually just matching the top performer of the West. The timelines are typically pretty crazy — within a year, you need to deliver this thing.

That might sound at first like it won’t work, but they do it frequently enough that eventually you get there. Maybe you get a thousand submissions of which 999 are bad, but one isn’t, and that one is successful. Even without highly talented folks running these programs, they have that forcing function of serving as a constant source of demand for these products. There’s some money attached, and even if it’s not substantial, it’s enough to get enough submissions that one of them might be good. I worry about that.

The counterpart in the United States is that even though we haven’t incorporated grand challenge or moonshot–style programs to the same degree — although that seems to be changing with this administration — what we do have are highly talented government folks who are so deep on the specifics that they can craft really thoughtful programs. I’m thinking here of DARPA, DOE, NIST. Programs that aim for the right level of requirements and have enough incentives attached. You don’t need a million programs like in China. You can have a few, but they’re very thoughtful, driven by people who really understand the science and the technology.

That’s an asset we have compared to everyone in the world. China is an easy counterpart, but even in Europe, you don’t have the same level of technical sophistication we have here. I worry a little bit about that changing in the last few years, but in general, we still have incredibly talented folks in government.

Zachary Yerushalmi: To dovetail with this, the reactivity in the Chinese academic sector is incredibly powerful. I was chatting a couple of weeks ago with a prominent quantum physicist. They were telling me about a paper they read on a new type of PIC. What had happened was that a Chinese group had a certain type of material system they were working on, and they had friends over at Columbia working on another type of material system. These were adjacent publications and approaches.

What this Chinese group did was look at the two approaches and ask, “What would you do if you could just put them together?” It turns out you get wildly better results. Their point was that in America, hitherto, you’d never do that, because the bureaucratic system around applying for grants is so intense that you couldn’t just say, “Let’s put these two material systems together.”

What I would call out with the new administration — and for all sorts of reasons, I don’t want to get political, but I’d say as a real positive — is that when you look at Deputy Secretary Dabbar and Undersecretary Gil, they get that, and they’re really driving toward a totally new paradigm. You can say, “That material system is cool, that material system is cool — let’s drive this thing and see what new innovation you can do.” You don’t have to spend new money. You just have to move fast and be creative, and they’re going for it.

Chris Miller: Constanza, lasers — critical for quantum computing. Tell us where they’re made today and how they should be scaled up.

Constanza Vidal Bustamante: In the report, we cover them in this big category of photonics and optics. Lasers are part of that and are the star of the show, but definitely not the only component to watch. They affect different modalities differently. When you think about the main subcategories we cover — solid-state superconducting and semiconducting modalities, atomic ones, and photonic ones — lasers matter most for the photonic and atomic modalities.

It’s not as simple as one laser. You need multiple lasers doing very different things. Take the atomic modalities — neutral atoms or trapped ions. You need different kinds of lasers to cool the atoms. Instead of a cryogenic system, you use a laser system. When you shine the laser beam on the atoms, you bring them down to ultra-cold temperatures so you can manipulate them. Then you have different lasers to elicit different energy transitions, and other lasers to read out the effects. It’s a chain of lasers.

What they share in common is that they’re all highly specified to the specific wavelengths you need to hit, and they need to be extremely stable in those wavelengths to maintain the frequencies that make them usable for these computations. It’s not just one laser, and it’s not just lasers. You need all these optical components to route the light, change directions, change the frequency — maybe double it — various lenses to focus the light, and so on. There are a lot of subsystems involved.

The supply chains are also complicated because they’re all different things. The lasers we cover with the highest priority in the report mostly come from companies in Japan, Europe, and China.

There’s an interesting case study we cover in the report. A provider appeared in China that started manufacturing lasers essentially identical to those of a Danish company — but at a much cheaper price. There has been a lot of behind-the-scenes discussion over whether this was reverse engineered, and this Chinese company is well documented to receive government subsidies. There’s a seemingly clear story of what happened. Nevertheless, they’ve become a very important provider of lasers in the US ecosystem to this day.

What’s especially baffling is that even with the tariffs that have impacted everything, including the quantum industry, you can still call for an R&D exemption for those lasers. Those are still being purchased to this day by companies and universities that can claim an R&D exemption. They have a good product going. It’s price competitive, they apparently deliver reliably, and so it has become the preferred laser system for many organizations.

Zachary Yerushalmi: This point on price is a really big deal. It’s seen both in lasers and on the cryogenic side. There are key components like wiring trees, which you need to operate dilution refrigerators. For most experimental setups, you need a new wiring tree. The cost of a Chinese-produced wiring tree is literally one-tenth of the US equivalent — even after tariffs and all that. I’d imagine similarly so on the photonics.

When you look at the photonics side, like cryogenics, there are only two credible laser providers for quantum systems in the US — Vector Atomic and Vescent. Only two. They’re still medium-sized companies despite their headcount, and they have amazing teams. The criticality is not just that if China underpins them, prevents them from scaling, and undermines their ability to innovate — that’s a real big deal. But even more near-term, these laser systems are used for quantum sensors.

Quantum sensors — going back to the report — covers how the very thing that makes a quantum computer hard to build is what makes these amazing sensors at a unit level that can transform our world. One example is providing navigation without reliance on a GPS uplink, which, as we’ve learned in the conflict with Iran, and as we knew long before, is a really big deal. The same laser systems play in.

Should We Ban Chinese Components?

Chris Miller: China’s producing components for a tenth the cost of Western firms. We’ve seen this far outside of quantum, in many other spheres. We’re at this point where, as you’re saying, we’re going to have a dramatic scale-up over the next half decade or decade in the number of these components, as we build bigger and more capable computers.

Option one would be to subsidize Western producers tenfold so the price equalizes — that seems expensive. Option two, ban Chinese components from our quantum systems, but then you have higher prices. What should we be doing here? And should we be banning Chinese components from our quantum computers?

Constanza Vidal Bustamante: This will not be well taken among the quantum industry, but I do think we should not let this product continue entering the US market. That said, there are lots of considerations. I’m so glad Zach brought up the sensors, because that’s a much nearer-term market that will require these laser systems and various optical components at scale, much sooner than quantum computing. It becomes a real near-term bottleneck.

In terms of options, I think it’s both. We call in the report for some kind of subsidies, but really more like strategic financing or tax breaks. If you think about the supply chains of the lasers, some are dependent on foreign suppliers, including for some of the tooling needed to build them, often from Europe. Part of the solution is to provide some support for domestic suppliers while making it harder for Chinese products to take over the market, given that they may have used illegitimate ways of obtaining or accessing the IP that led to those products and have received substantial subsidies from the CCP.

Chris Miller: Suppose we go down the path of banning Chinese components from quantum computers. You get into a similar set of questions as if you said to a lot of people in Washington, “Let’s ban Chinese components from our AI data centers.” People often, at first glance, say, “Great idea.” Then you say, “Well, wait a minute—what about the screws? What about the light bulbs?” Where do you draw the line? Help us understand how to think about drawing lines in quantum computing.

Constanza Vidal Bustamante: That’s an excellent question. It’s hard to have an easy solution. Even if you stop the import of some of these devices or inputs overnight, that can lead to a lot of problems in our own ecosystem and our ability to keep innovating across broader products.

To your point about all these other sub-components —should we also block those? Is that worth blocking? There are layers of complexity. The inputs that require sophistication and that have the value we care most about are the ones we should bring in-house. In this case, for the lasers, we have some domestic suppliers. We have the talent. We have a path to get there with really good products that are having difficulty because they’re encountering anti-competitive practices.

Those lasers will be useful across different quantum technologies — sensors, computers, and to some degree networking too. They also serve beyond quantum — telecom and various defense needs. That makes them strategic enough as an enabling technology that I’d want to preserve our domestic capabilities. Compared to much simpler inputs to the inputs, that’s where I draw a distinction. I don’t have a super clear line in the sand, but that’s broadly how I think about it.

Zachary Yerushalmi: Take the wiring tree. If we say no Chinese wiring trees, that means for some groups, they can buy ten times fewer wiring trees, which means they do ten times fewer experimental runs. It’s not exactly linear, but it would act immediately as a hindrance on our ability to innovate. There are real trade-offs.

What I’m more clear on is the end state we aim for, which is some mix of three things. First, access — is the widget on the shelf? Second, security — particularly for end-stage products, do we know where that supply chain is, so that China isn’t putting a little microchip into the thing to listen to our experiments, or some Stuxnet-style attack? Third, can we continue to out-innovate?

Out-innovating is a lot more reliant — maybe even more than on price — on the speed at which you can get the widget. For a lot of these systems, you don’t require new fundamental physics — you just require being able to run through ideas quickly. There’s a separate learning that comes with that, around how you get to scale.

If you can balance these different priorities — using iteration speed as a proxy for staying innovative at every stage of a technology cycle — I think we’ll be in a good place. There are lots of different ways to skin a cat. We just have to be mindful of those trade-offs.

Constanza Vidal Bustamante: Another key aspect, related to what we were just saying about how many layers down you go. A big point we make in the report is the category of specialized materials. These are the ultimate substrate you need to actually build many of the components — for instance, the photonic integrated circuits we were talking about. Even some of the bulkier lasers rely on highly specialized photonic materials, including wafers that you process to make into devices. Some of those are sourced single-source from China right now.

That’s another concern. It’s not the laser itself, or the optical or photonic component itself, but the raw material you need to build it. If you don’t have access to that, you can go upstream in the innovation chain.

Zachary Yerushalmi: There’s a law that the second you create a metric, it ceases to be useful. The thing that comes to mind is — if you focus on whole product systems and ask, “How long does it take to go from initial design to inception of the product?” and you try to reduce that as much as possible, then you identify the requisite bottlenecks that you need to prioritize for investment.

You can do that for non-national security tech and just allow the component tree from anywhere. But then you have to apply a separate lens of national security, where you probably don’t want a certain chip coming from a certain place that’s not the US. Look at the lead time for that. If you compare these two lead times and try to ruthlessly bring them down — in a general sense, but also compared to your adversaries — you have a bit of a North Star: how are we doing, where do we prioritize, what do we do next? I hold that pretty lightly. You’ve been thinking about this more, but that’s where my silly bad supply-chain brain goes.

Comparing Stress Levels

Jordan Schneider: Constanza, you did your PhD thesis on managing people’s stress over time. As you talk to all these people in the quantum supply ecosystem, what’s their stress level? Are they like, “I’ve got exams in a week”? Or are they feeling good?

Constanza Vidal Bustamante: Oh my gosh, what an honor that you went back into my history. I guess it hasn’t been that long.

In talking to the quantum folks, there’s a lot of excitement, but also a lot of uncertainty. Obviously, I approach this from a policy perspective. There’s a lot of enthusiasm from the administration and Congress to do something big on quantum and to build on the foundations of the National Quantum Initiative, which came out during the first Trump administration, and the National Quantum Initiative Act, which solidified that and provided funding mechanisms for specific programs at different agencies. There’s a lot of expectation, but also a little bit of fear about what will actually happen. The stress levels are real.

All of these companies have a lot of pressure to deliver on these machines by the roadmaps they swore by. They’ve all been claiming they’ll start delivering utility-scale machines by the end of the decade, and the clock is ticking. Some have been more aggressive than others about what they’ll deliver, so there’s a lot of expectation about whether they’ll come through. If they don’t, there’s worry about what will happen to the field. Even if a competitor firm fails, will that lead to a generalized lack of confidence that brings down private capital writ large? There’s a lot of fear about what will happen, and a lot of pressure they’re feeling right now.

Jordan Schneider: Can you compare that to the semiconductor community? You’ve also done research interacting with them. I don’t think the chips folks are worried that chips won’t be a thing in five years. What anthropological differences have you picked up on?

Constanza Vidal Bustamante: That’s such a great question. It’s a very different environment. At the same time, there’s an incentive to present quantum as being as close to the semiconductor industry as possible — to give the idea that we have a path to manufacturability, or to build on top of the CHIPS Act or the CHIPS and Science Act energy and come up with this big industrial moment, as I called it in the report.

What’s interesting is that compared to the semiconductor industry, you have all these different modalities, with apparently close to 90 companies now building quantum hardware across various modalities. That’s so different from the semiconductor industry. It leads to all sorts of competition among them over who has the best qubit and why the others are inferior. It’s funny to hear — everyone will tell you endlessly why you should support their qubit modality and why they have the right one going.

Jordan Schneider: That was really my big takeaway from my little quantum journey over the past few weeks. In the semiconductor industry, things are consolidated. You have two EDA players, a handful of people making photomasks, and one company making EUV machines. It’s been that way for a pretty long time, and it’s probably going to stay that way. Maybe you’ll have an entrant here or there on the design side, but the entire industry is pulling in basically one direction, with everyone trying to capture an extra 10 or 20 percent of where they sit in the supply chain.

When you walk through the quantum stack, everyone’s using more or less the same ingredients to varying extents, but what the computer is going to look like is totally up for grabs. It’s not like Game of Thrones with six or seven royal houses — there are 90 different little empires competing for the prize.

Constanza Vidal Bustamante: That’s right. We’ll see how many survive, and what diversity survives. We didn’t even get to this, but even within a modality, there are different ways to build your architecture. There’s a lot there.

Zachary Yerushalmi: My mental model for quantum is biotech. When you’re trying to cure cancer, you have small molecules, CAR-T, antibodies, immunotherapies — all these different approaches trying to address something out there, which is a kind of unified target.

In other domains, we’ve figured out how to take hardcore fundamental science and mature it to impact our lives, even when there are many different approaches. That’s a slightly different mental model from semiconductors, but it doesn’t mean it can’t exist.

One of the 50 reasons I’m so excited that Jordan is covering this, that Chris is super attuned to it, and that Constanza writes these seminal reports, is that folks have spent decades from different perspectives asking, “How do we get biotech right?” Public policy folks have a frame of reference around biotech. Public finance folks understand it. Doctors understand it. Physicists understand it. In quantum, that hasn’t happened yet. We haven’t had proper academic rigor across the disciplines.

I really don’t think we’ll get this right unless we bring that interdisciplinary best practice now, at this stage. I’m super stoked for more.

Jordan Schneider: All right, kids, shout out to Zach and Constanza. I’m sure they’ll find some work for you. This was a pleasure.

• Jake Sullivan, former NSA now at the Harvard Kennedy School

• Chris Miller, Chip Wars author and belt-holder for most ChinaTalk appearances

• Dan Kim, former Chief Economist for the Chips Program Office

• Dan Wang, author of Breakneck

We had two prompts:

• What are the most important high level KPIs that policy should aim for? What is the analogy of the Fed’s ‘2% inflation and full employment’ target for economic security?

• Where today would you put $10-50bn to get the most for your investment in economic security? Feel free to propose both defensive and offensive ideas, and either a portfolio of ideas or the one large idea you think will deliver the most value.

We had a literal four-way tie for first place as each judge gave one of these essays their first-place designation. We will be running the contributions from our winners in the next two weeks.

The first essay comes from , a Lieutenant Colonel in the U.S. Air Force, a Fellow at the U.S. Naval War College, and a Senior Fellow at the Payne Institute for Public Policy.
Disclaimer: The views of Lt Col Matisek are his own, and not those of the U.S. Air Force, Department of War, or U.S. Government.

Economic security is frequently invoked but is the least disciplined policy concept. It now justifies subsidies, export controls, stockpiles, reshoring, and strategic deals. What has emerged is a kind of “China-light” strategy in Washington, where activity is everywhere but coherence is not. Government money is committed and facilities are built, yet the U.S. defense industrial base remains a “Black Box” to policymakers and Pentagon officials. Key senior leaders still struggle to answer a basic question: which parts of the American industrial ecosystem can actually withstand disruption and which will fail first under pressure? That uncertainty extends well beyond the defense industrial base. It touches semiconductors, energy systems, critical minerals, logistics networks, machine tools, and the enabling infrastructure behind Artificial Intelligence (AI) and advanced manufacturing. The deeper question is not if the United States can keep peace with China in peacetime. It is whether the American economy can surge fast enough, recover fast enough, and keep producing when coercion sets in.

This question matters because strategic competition is reshaping the global economy. Deglobalization and selective decoupling are unfolding through policies meant to reduce dependence on China and minimize the risks of economic coercion. Yet coercion is not an abstract threat. It is a practical strategy aimed at supply chains, production systems, and the nodes that support them; something Washington let atrophy after the Cold War ended. Beijing’s advantage does not rest only on scale or innovation. It lies in its position across time-intensive, tooling-intensive, and capital-intensive parts of the industrial ecosystem that are difficult to replace once disrupted. Pressure at those nodes usually does not cause a collapse. It produces delays, shortages, and missed output targets that give strategic leverage to an adversary.

Recent events illustrate how different types of shocks reveal critical vulnerabilities. Direct coercion is the most obvious threat. Since 2023, China has imposed export controls on antimony, gallium, germanium, graphite, and heavy rare-earth elements, creating upstream bottlenecks, leading to over 300 F-35s being delivered without its new AN/APG-85 radar due to a lack of gallium. But vulnerabilities are also exposed by indirect shocks and internal failures. The 2022 neon gas shortage, a consequence of Russia’s invasion of Ukraine, showed that requalifying new sources for key semiconductor inputs can take over a year. Likewise, the U.S. Army’s struggle to scale 155 mm artillery production highlighted domestic bottlenecks with energetics and tooling. Similarly, instability in Mozambique led to suspension of graphite mining, disrupting supply chains for batteries. The 2026 Iran War only further reinforces the point because precision guided munitions are being expended at an unsustainable rate, while materials like helium and sulfur are unable to be transported out of the Gulf causing economic shocks and undermining the defense industrial base. Different shocks reveal the same problem: what matters most is not the source of disruption, but how long it takes to recover from it.

The popularization of the “small yard, high fence” approach was an important shift, correctly moving the focus from broad decoupling to specific chokepoints. However, its theoretical elegance has been challenged by its messy reality, leading some to claim the need to move beyond the concept. This is because a fence is useless if the factory behind it can’t operate, especially when a peer competitor like China has cultivated an “engineering state” capable of building its own capacity with unusual speed. The hard truth is that even a perfect fence cannot solve for domestic industrial weakness, as capital investment alone cannot erase long lead times. Real chokepoints extend well beyond chips to include energetics, chemicals, and tooling. Therefore, speed in reconstitution is the enabler, and capital must target the real constraints on output under stress.

This logic aligns with the theory of weaponized interdependence, which explains how asymmetric control over centralized nodes in global economic networks allows states to turn interdependence itself into a coercive instrument. In that world, diversification is sometimes useful, but it is not a cure-all. If upstream chokepoints remain concentrated and slow to replace, vulnerability persists beneath the appearance of redundancy. The task of economic security policy is therefore not to eliminate interdependence altogether, but to identify which nodes are most susceptible to weaponization and to harden them accordingly.

Seen this way, economic security is best understood as endurance under coercion.

It is the ability to sustain national power over time when disruption is deliberate and recovery is contested. The defense industrial base provides the sharpest stress test of this problem because wartime demand exposes bottlenecks quickly and brutally, but the underlying logic extends across the broader industrial ecosystem. Markets alone will not solve it. They reward efficiency, discount tail risk, and rarely invest in idle capacity, redundant tooling, or costly resilience without strong incentives to do so.

Washington has begun to recognize this reality. Industrial policy has returned as a tool of statecraft. But without a clear standard for success, these efforts risk becoming fragmented, episodic, and politically fragile. Economic security needs the equivalent of a dual mandate: a disciplined way to distinguish between activity and capability, between announced investments and real industrial endurance.

I contend that economic security should be organized around a coercion-endurance mandate centered on sustained production under pressure. From that mandate flows a small set of high-level indicators designed to capture where industrial systems break first and how they can be strengthened. The point is not to outbuild China across every sector. It is to create an American industrial ecosystem that is difficult to interrupt, slow to degrade, and costly to coerce.

That is the standard that matters. And it is one that can be measured.

Why Existing Economic Security Frameworks Fail Measurement

The turn to economic security has yielded frameworks and diagnostic tools that map supply chains and dependencies. While these improve situational awareness, awareness is not measurement; description does not equal control under pressure. Knowing where inputs originate does not reveal whether production can be sustained once disruption begins. Mapping dependencies does not show how systems behave when time, cash flow, bottlenecks, and recovery capacity become binding. Economic coercion exploits dynamics, while most current metrics remain static.

A central weakness is that many existing frameworks measure peacetime structure, not performance under stress. Indicators like import reliance describe a world without crisis, but are blind to how a system actually degrades or recovers when a critical node fails. The goal of economic coercion is rarely total destruction. History has shown, from the bombing of the Schweinfurt ball-bearing plants onward, that completely destroying an industrial node is nearly impossible. Instead, the modern objective is attrition: pushing output below a threshold long enough that it undermines operational abilities that shape of desired strategic outcome. Frameworks that fixate on static exposure while ignoring recovery speed are just fighting the last war. It means misunderstanding the mechanism through which externally imposed pressure works.

Another common error is the assumption that diversification automatically yields security. Spreading production across more locations can reduce exposure to a single supplier, but it does little to address constraints that are global in nature. Tooling, specialized labor, certification timelines, and precursor chemicals often remain concentrated even as final assembly disperses. Under stress, these upstream constraints reassert themselves quickly. Diversification without fixing reconstitution can create the appearance of resilience while leaving industrial endurance largely unchanged.

Current approaches also mistake announcements for achievement, overemphasizing paper capacity rather than usable output. While celebrating CHIPS Act investments is politically useful industrial reality is far harsher. The lesson becomes not so much that major industrial investments are misguided, but that physical construction is only one part of capability. TSMC’s Arizona project is best understood in these terms. Its early delays and operational frictions do not invalidate this major domestic effort. They reveal the difficulty of transplanting advanced production into a new ecosystem where skilled labor, supplier networks, water, power, and tacit organizational know-how all matter. The larger point is that industrial capability is not created by capital expenditure alone. It must be built, staffed, supplied, and sustained. That lesson extends beyond semiconductors to AI infrastructure, advanced manufacturing, and other sectors whose growth depends on fragile enabling systems.

An equally important issue lies below the prime contractor level. Industrial systems fail from the bottom up. Tier-2 and Tier-3 suppliers operate on thin margins, depend on steady cash flow, and often lack access to emergency credit. When shocks occur, these firms fail first. Payment delays or input disruptions cascade upward, halting production regardless of demand or funding at the prime level. Frameworks that do not measure sub-tier financial resilience are missing one of the most common ways in which economic pressure becomes systemic failure.

The cumulative result is that we end up having a policy environment rich in inputs but poor in outcomes. Economic security initiatives remain politically fragile and strategically ambiguous when they cannot distinguish between visibility and control, diversification and resilience, or announced investment and actual performance. What is needed is a shift from descriptive risk mapping to the measurement of industrial behavior under stress. Economic security requires indicators that capture how quickly systems recover, how far they can surge, where pressure concentrates, and how long production can be sustained once disruption begins. Only then can policymakers distinguish between industrial activity and industrial power.

The Economic Security Dual Mandate

If economic security is endurance under coercion, then it requires a governing logic that privileges performance over posture and outcomes over activity. Without such logic, policy fragments into disconnected programs that are difficult to evaluate and easy to politicize. Economic security needs an organizing principle that disciplines decision-making across institutions and administrations. Monetary policy offers a useful template. The Federal Reserve’s dual mandate translates abstract goals into durable targets that anchor expectations and guide action. Economic security requires a similar level of clarity. It needs a mandate that defines what success looks like under pressure.

I propose an Economic Security Dual Mandate built around two complementary objectives: (1) Minimum Viable Capacity and (2) Maximum Credible Coercion Cost.

1. Minimum Viable Capacity: The ability to sustain production of essential military, industrial, and technological outputs at a defined level for a defined period under adverse conditions. It is not about peak performance or global dominance; it is about the floor of output that must be maintained when disruption occurs. This logic reflects how practitioners have already begun to approach the problem. At the CHIPS Program Office, for instance, economic security was framed through interrelated dimensions of capacity, capability, competition, and criticality. In practice, however, the binding constraint repeatedly surfaced as recovery time. The central issue is not paper capacity, but how quickly production resumes after disruption. Minimum Viable Capacity formalizes time to recovery as a strategic variable.
   

2. Maximum Credible Coercion Cost: Capturing the flip side of the mandate, this reflects how expensive, slow, and uncertain it is for an adversary to disrupt U.S. production through targeted pressure. The higher the cost and the longer the timeline, the less effective coercion becomes as a strategic tool.
   

Together, these two objectives define economic security as a contest over endurance. Capacity without a coercion cost just invites pressure from an adversary. Coercion cost without capacity yields hollow resilience. Only when both are present does an industrial system become strategically resilient. The mandate also clarifies the role of the state. Minimum capacity and coercion cost are public goods. They require coordinated investment, long time horizons, and a tolerance for redundancy that markets alone rarely provide. The point of the mandate is not to prescribe a single industrial policy. It is to create a standard against which policies can be judged.

The challenge, of course, is measurement. Abstract mandates only matter if they can be translated into indicators that track real performance under stress. That does not require perfect precision at the outset, but it does require repeatable methods. Some indicators can be estimated through supplier mapping, sector-level concentration data, and confidential firm reporting. Others would require stress tests, red-team exercises, trial production runs, or disruption simulations conducted jointly by government and industry. A flawless dashboard cannot be created on day one. However, time and intentional policies are needed to build the institutional machinery needed to measure recoverability, surge potential, bottleneck concentration, and financial resilience in a consistent way over time.

Five KPIs for Determining Economic Security under Coercion

A mandate without measurement is rhetoric. American economic security can only be achieved through endurance under coercion. This means having indicators that capture how industrial systems perform when pressure is applied. Static measures of exposure or announced capacity won’t work. Time matters, as does throughput, concentration, and financial resilience.

The five indicators below translate the Economic Security Dual Mandate into a usable scoreboard. They focus on where systems break, how quickly they recover, and where coercion delivers leverage at lowest cost.

1. Time to Reconstitute

This measures how long it takes to restore meaningful production after a critical disruption. It is the most important indicator of endurance because time is the currency of coercion. In practice, reconstitution timelines are governed less by capital availability than by industrial physics. Semiconductor process-node requalification after a supplier loss often take 6 to 18 months. Rare-earth separation and magnet manufacturing lines have historically taken 3 to 7 years from permitting to full output. Machine-tool rebuilds after a chokepoint failure can exceed 18 months due to specialized castings and skilled labor. Systems with long reconstitution timelines remain vulnerable even when diversified on paper. Measuring this KPI forces policymakers to distinguish between theoretical substitutability and operational reality.

Target: About 6 to 12 months for any Tier-1 critical input, validated by red teaming supply chains.

2. Surge Ratio

This captures the maximum sustainable increase in output relative to peacetime baseline production over a defined period. It answers a simple question: how much more can be produced, and for how long, before the system breaks? Before the 2022 Russo-Ukraine War, U.S. production of 155 mm artillery shells averaged roughly 14,000 rounds per month. Despite ambitious targets to produce 100,000 rounds per month, production has stalled due to issues of sourcing energetics, fuzes, tooling, and skilled labor. Even a wartime demand — with a $6 billion infusion from the Pentagon — did not generate industrial willpower to meet surge goals, as the U.S. Army is only able to produce 56,000 shells a month as of February 2026. Achieving high surge ratios requires pre-positioned idle lines, redundant tooling, and cross-trained labor.

Target: Sustaining 3 to 5 times peacetime output for 12 to 18 months without cascading failures.

3. Chokepoint Concentration Index

We also need to measure how much control a supplier actually has over a non-substitutable input resides with the top one or three suppliers. Unlike traditional concentration metrics, this KPI focuses only on nodes that cannot be bypassed. The leverage embedded in such nodes is substantial. China controls basically 90% of global rare-earth refining and 90% of permanent-magnet production. Disruption at a single node, can undermine the U.S. economy and military, because both are so heavily reliant on these inputs to produce precision motors, guidance systems, and actuators. This KPI aligns directly with weaponized interdependence logic. It identifies where network topology creates coercive leverage and where investment most directly raises the cost of disruption.

Method: A Herfindahl-style index applied only to non-substitutable nodes, weighting supplier share by the degree to which inputs lack viable alternatives, so that concentration reflects true coercive leverage rather than nominal market diversity.

4. Sub Tier Supplier Liquidity Coverage

This measures how long critical Tier-2 and Tier-3 suppliers can continue operating under stress. It captures how financial shocks propagate through the industrial base. Recent assessments repeatedly show sub-tier suppliers operating on thin margins with limited access to emergency credit. When disruptions occur and payments slow, these firms fail first, triggering production stoppages that cascade upward.

Target: About 90 to 180 days of assured liquidity for priority suppliers, supported through guaranteed credit lines or accelerated payments.

5. Assured Inputs Stockpile Days

This measures how long production can continue using secure inventories of irreplaceable inputs. These are materials and precursors that cannot be substituted or sourced at scale under duress. Current stockpiles of critical energetics precursors and magnet alloys often cover a few weeks or months of wartime consumption. Operationally meaningful reserves must be sized to production rates rather than abstract tonnage. For example, stockpiles should be calibrated to sustain guidance-system production for key munitions.

Target: Approximately 12 to 18 months of sustained production for critical systems.

KPI Summary

Taken together, these five KPIs operationalize the Economic Security Dual Mandate. They shift attention from exposure to performance, from announcements to outcomes. They also explain why many well-intentioned policies fail to improve endurance. If investment does not move these indicators, economic security remains aspirational. Measurement, however, is only half the problem. Progress will only be made by allocating useful capital to shift these KPIs in meaningful ways. That is where economic security requires intentional investments to ensure a resilient industrial base.

Why Capital Allocation Determines Economic Security Outcomes

Measurement identifies where industrial systems fail. Capital allocation determines whether those failures persist. Without disciplined investment to overcome chokepoints, even the best KPI framework is just more policy pontification. Economic security is achieved by where the money goes, when it goes there, and what it is allowed to buy.

This distinction matters because much of today’s economic security spending still treats capital as a signaling device rather than a constraint-solving tool. Funds are often dispersed to demonstrate commitment, attract private investment, or spread benefits geographically. Those goals may be politically useful, but they do little to improve endurance under coercion unless they target the factors that actually govern output when the system is disrupted. The relevant question is not whether investment is large, but whether it measurably improves recoverability, surge potential, and resilience at critical nodes.

Industrial systems are shaped by irreversibility. Tooling, facilities, workforce pipelines, and qualification processes lock in production patterns for years or decades. Once these structures are set, they are slow to change regardless of demand signals downstream. Capital that arrives after a constraint is revealed cannot be repurposed quickly when conditions deteriorate. By the time disruption exposes where the system is weakest, it is already too late to build around those weaknesses.

This is why economic security investment must be evaluated differently from growth or innovation spending. The objective is not to maximize returns or accelerate diffusion. It is to raise the floor of output and steepen the recovery curve after disruption. That requires a bias toward assets with long lead times, high fixed costs, and limited substitutability. These are precisely the areas where private capital is least willing to invest without help from The Entrepreneurial State.

Capital discipline also matters because economic security spending competes with itself. When resources are spread thinly across too many initiatives, no single constraint is meaningfully relaxed. The result is a portfolio that looks comprehensive but delivers marginal gains everywhere. Endurance improves only when investment is concentrated at anticipated points of failure. Public capital is not meant to replace markets or permanently subsidize production. State-involvement is just meant to shape industrial structure in ways that markets alone will not, such as the Pentagon deal with MP Materials to source domestically produced magnets with guaranteed price floors. Once endurance is built into the system, private actors can compete within those bounds; meaning this framework makes investment options clearer.

A $50 Billion Endurance Build for a China-Proofed Industrial Base

Capital improves economic security only when it is used to solve binding constraints. The purpose of a $50 billion investment is not to chase technological primacy or replicate China’s scale across every sector. It is to illustrate how public and private capital might be concentrated against the highest-leverage vulnerabilities in the American industrial ecosystem. The allocation below is therefore best understood as a stylized portfolio, not a full national industrial strategy. Its logic is simple: prioritize sectors where lead times are long, substitutability is low, spillovers are high, and coercion vulnerability is acute.

The four pillars below reflect those criteria. Together, they target bottlenecks that would matter not only to the defense industrial base, but also to semiconductors, energy systems, logistics networks, and advanced manufacturing more broadly.

Pillar One: Energetics and Munitions Throughput ($15 Billion)

Sustained mass and fires, not exquisite platforms, are vital for warfighting during a prolonged conflict. Energetics, propellants, explosives, and their upstream chemical precursors govern what weapon systems can be employed at scale. Unfortunately, these production lines are capital intensive, environmentally constrained, and slow to expand. Despite billions being committed to ramp up munition and missile production, industry has struggled to match demand.

A $15 billion investment focused upstream would fund redundant precursor plants, idle surge lines maintained for crisis activation, and workforce pipelines for energetics chemists and technicians. These investments directly impact KPIs for: increasing Surge Ratio, shorten Time to Reconstitute, and extend Assured Inputs Stockpile Days.

Pillar Two: Midstream Processing and Magnet Production ($15 Billion)

Economic coercion is applied most efficiently in the middle of supply chains, where raw materials are converted into usable industrial inputs. Rare-earth magnets illustrate the problem, such as Chinese sourced magnets ending up in American-made F-35s. American mining doesn’t help because it takes 29 years to get a mine up and running — and it still takes 16 years elsewhere for a new mine. Even worse, separation, refining, alloying, and magnet manufacturing are what determine usable output across the industrial base, which is also dominated by China. U.S. and allied efforts are meaningful but will not reach a useful industrial scale until the early 2030s.

A $15 billion commitment would accelerate heavy-rare-earth separation and magnet facilities enough to get China out of the Western supply chain. Moreover, such investment would support qualification and offtake agreements and enable stockpiling of inventories in industrially usable forms. For the KPI, it directly reduces the Chokepoint Concentration Index and extends Assured Inputs Stockpile Days.

Pillar Three: Sub Tier Industrial Finance as a Security Instrument ($10 Billion)

Industrial systems fail from the bottom up. Tier-2 and Tier-3 suppliers absorb shocks first and recover last, yet remain largely invisible in industrial policy debates. That invisibility is dangerous because these firms often sit at the exact points where localized disruption becomes systemic failure. Besides small suppliers being fragile, the structure of defense and advanced manufacturing supply chains often leaves them exposed to cash-flow shocks even when prime contractors remain insulated. For instance, sub-tier firms usually contract through major integrators rather than directly with government, which means they often do not benefit from the favorable financing terms available to primes. The same reporting also shows a shrinking supplier base, with more than 17,000 firms leaving the defense sector in recent years and small-business participation down sharply.

A $10 billion sub-tier industrial finance facility would therefore function as a standing shock absorber rather than a subsidy program. Revolving credit, rapid-payment guarantees, government-backed liquidity lines, and resilience-linked contracting for priority suppliers would stabilize the firms most likely to fail first in crisis. This directly improves Sub-Tier Liquidity Coverage and helps prevent payment delays or input disruptions from cascading into system-wide failure. Because many of these firms also support aerospace, automotive, electronics, and energy systems, stronger sub-tier finance improves resilience across the wider economy, not just defense production.

Pillar Four: Machine Tools and the Industrial Commons ($10 Billion)

The ability to make the tools that make everything else is foundational to endurance. Machine tools, advanced manufacturing equipment, and specialized components underpin every industrial sector, yet domestic capacity has eroded. China commands around 33% of global machine tool production. Specialized components, precision castings, and skilled labor pipelines take years to rebuild.

A $10 billion investment in the industrial commons would combine tax credits, purchase commitments, and targeted R&D for next-generation automated machining tools, additive manufacturing equipment, and the supplier networks needed to sustain them. These investments shorten Time to Reconstitute across sectors and reduce chokepoint concentration in the tooling base itself. Machine tools underpin semiconductor fabrication, aerospace production, automotive manufacturing, and energy infrastructure. Rebuilding “industrial commons” improves recoverability across the entire economy.

How the Endurance Build Moves the KPIs

The value of this portfolio lies in its ability to move the indicators that actually define economic security performance. Investments in energetics, midstream processing, sub-tier finance, and the industrial commons collectively raise surge capacity, shorten recovery timelines, reduce chokepoint concentration, stabilize supplier liquidity, and extend assured inputs coverage.

These effects reinforce one another. Surge is hollow if sub-tier firms fail first. Stockpiles buy little time if tooling cannot be replaced. Reduced chokepoint concentration matters only if alternative capacity can be staffed, financed, and brought online quickly. By explicitly linking capital allocation to KPI movement, the endurance build turns economic security from a collection of programs into a coherent, measurable strategy.

Designing an American Economy that cannot be stopped

Economic security debates often gravitate toward scale, speed, or technological edge. Those attributes matter, but they are not decisive on their own. The more fundamental question is: Can the United States sustain production when pressure is applied deliberately and continuously? If the answer is unclear, economic security is a slogan.

Endurance under coercion must be a governing strategy that offers a clearer organizing principle. It shifts attention away from peacetime efficiency and toward performance under stress. It also clarifies why geography alone is an insufficient proxy for security and why market forces, left to themselves, rarely preserve the redundancy, recoverability, and slack that continuity under disruption requires.

My proposed Economic Security Dual Mandate strives to provide that discipline. Minimum Viable Capacity defines the floor of output that must be sustained under adverse conditions. Maximum Credible Coercion Cost defines how difficult it is for an adversary to interrupt that output. Together, they turn economic security into a testable proposition.

The five KPIs operationalize that mandate. Time to Reconstitute, Surge Ratio, Chokepoint Concentration, Sub-Tier Liquidity Coverage, and Assured Inputs Stockpile Days capture the mechanisms through which coercion actually works. They reveal where industrial systems break first, where resilience is real rather than assumed, and where investment can generate the greatest strategic return. Just as importantly, they offer a common language for distinguishing between industrial activity and industrial power.

Capital allocation is the bridge between diagnosis and capability. Economic security is strengthened by concentrating investment where lead times are long, substitutability is low, spillovers are high, and failure would be strategically costly. The endurance build outlined here is not a complete industrial strategy. It is an illustration of how public and private capital can be aligned to strengthen the industrial ecosystem where coercion would otherwise bite hardest.

That broader ecosystem matters. The defense industrial base is the most visible and unforgiving stress test, because wartime demand exposes bottlenecks quickly. But the same logic applies to semiconductors, energy systems, logistics networks, machine tools, and the enabling infrastructure behind advanced manufacturing and AI. A serious economic security framework must therefore extend beyond any single sector while still recognizing that some sectors reveal the problem more clearly than others.

This approach also offers political durability. A KPI-driven framework anchors debate around shared outcomes rather than changing rhetoric, reducing the temptation to relitigate economic security with every change in administration. That continuity is itself a strategic advantage.

The goal of economic security is not autarky. It is minimizing time to recovery while preserving competition, capability, and continuity across critical sectors. Endurance is what converts economic capacity into strategic power. If economic security is to move beyond slogans and into strategy, it must be judged by a simple test: how quickly production recovers, how long it sustains, and how costly it is for an adversary to interfere. That is the standard that matters. And it is one that can be measured.

We’ve recently tried to pin down how much compute China actually has, approaching the question from both the supply and demand sides. We converged on roughly 2.5 to 2.8 million H100-equivalents. But a single aggregate figure only captures part of the picture.

Jensen on China

On Dwarkesh’s podcast last week, Jensen Huang argued that China already has enough compute to build frontier AI.

“They manufacture 60% of the world’s mainstream chips, maybe more.”

When Dwarkesh raised the gap in advanced chips, Jensen responded,

“AI is a parallel computing problem, isn’t it? Why can’t they just put 4x, 10x, as many chips together because energy’s free?”

Jensen is wrong, but that doesn’t mean people aren’t compelled by this line of reasoning. John Moolenaar, who chairs the House Select Committee on China, sent a letter to Lutnick in December proposing a rolling technical threshold that would cap Chinese aggregate AI compute at 10% of US compute capacity. It’s much more nuanced — accounting for memory and network bandwidth as part of this calculus — but ultimately seems motivated by preventing, as the letter calls it, “death by a thousand sub‐threshold chips.”

Export restrictions are a difficult line to walk, and total computing power does matter. But not all compute is created equal. The compute that can train a frontier model, serve inference on an existing one, and power your laptop are different things, and a “death by a thousand sub-threshold chips” is less concerning for the trajectory of AI than a concentration of the most important chips.

Legacy Chips Don’t Matter for AI

It’s hard to know where Jensen is getting his claim that “China manufactures 60% of the world’s mainstream chips.” Perhaps originally from a 2024 projection from previous Commerce Secretary Gina Raimondo about new legacy chip capacity coming online in China. But this is not a measure of AI compute. It includes the chips running your car’s engine management system, your washing machine’s control board, and the power electronics in an industrial motor, typically manufactured at 28nm or larger. They matter, but they are not the chips that train frontier AI. A chip in your microwave cannot do matrix multiplication for a transformer, and a 40nm microcontroller in a Chinese EV does not help run DeepSeek-V4.

The sliver of Chinese chip output that is actually AI-relevant, primarily Huawei’s Ascend line, is roughly a million chips. But even the flagship Ascend 910C (with yields of about 300-600k chips this year) delivers slightly worse than Nvidia’s H20 for training, nowhere close to a Blackwell, and much of current production still depends on a stockpile of TSMC dies acquired before controls tightened. The remainder of China’s frontier-relevant compute comes from smuggled Nvidia chips and legally imported lower-tier chips like the H20. In short, China produces lower-quality chips and still cannot manufacture as many of them as the U.S. does; for them to reach anything close to a “death by a thousand sub-threshold chips” scenario, Chinese companies would have to concentrate what compute they do have to a degree greater than any American lab — a difficult task given the vigorous competition taking place between them.

This is why FLOPs is a more honest metric than total chip count. FLOPs, or floating-point operations per second, measure how many arithmetic calculations a chip can perform in a given second, and they are the fundamental currency of AI training and inference, since every command an AI executes is ultimately a sequence of multiply-and-add operations. And the FLOPs gap between frontier and legacy chips on this metric is staggering. A single Nvidia Blackwell B200 delivers roughly 10 petaFLOPs of dense FP8 performance, while a typical 28nm automotive microcontroller delivers around 0.12 teraFLOPs of FP32, roughly twenty thousand times less.1 To put that in concrete terms, if a country had 100,000 Blackwells, its rival would need more than the absurd number of two billion legacy chips to match the same FLOPs output.

But putting mainstream legacy chips aside, if China somehow did stack up many weak AI-focused chips (like Ascends), its problems would not end at matching FLOPs.

A Tale of Two Hypothetical Countries

Nvidiana and Huaweiopolis each have 2 million H100-equivalents. On paper, they are peers.

Nvidiana’s stock is top-heavy and lean. Roughly 300,000 frontier chips, the Blackwells and soon-to-arrive Vera Rubins, sit at the core, tightly interconnected in a handful of purpose-built data centers that can host training runs of tens of thousands of chips in lockstep. Another 600,000 chips, the H100s and H800s, handle large-scale training and serious inference. The remainder is padded out by around 650,000 older accelerators and general-purpose silicon for lighter workloads. Total physical chip count, roughly 1.55 million.

Huaweiopolis got to the same total a different way, by stacking weaker chips in enormous volume. Its top tier is thin, perhaps 50,000 frontier chips acquired before the latest round of export controls, and even those are scattered across several clusters rather than concentrated. A middle tier of around 450,000 chips, a mix of older Hopper variants and Chinese accelerators like Huawei’s Ascend 910B, is capable but constrained by weaker interconnect and memory bandwidth. The remaining mass of Huaweiopolis’s stack, close to 6.5 million chips, is older, inference-oriented chips like the H20, and repurposed general-purpose hardware. Total physical chip count, roughly 7 million — more than four times Nvidiana’s.

Nvidiana can train and serve the next generation of frontier models. Huaweiopolis cannot, and more chips will not close the gap. The difference in their AI trajectories will be substantial, even with identical FLOP counts.

Why Fewer Powerful Chips Beat Lots of Weak Chips

Huaweiopolis’s performance will lag behind for three main reasons: numerical precision, memory bandwidth, and network bandwidth.2

Numerical Precision

Older chips are not designed for the latest trends in numerical precision — that is, how finely or coarsely a chip represents numbers when doing calculations, which directly affects how much data needs to be moved and processed. Older chips, like the Hopper series, are designed to handle INT8 operations at best, meaning numbers are calculated to eight digits. Meanwhile, newer chips like the Blackwell series are designed to handle both INT8 and FP4 calculations, a jump that essentially doubles the speed of a chip. These chips can instead calculate numbers to only four digits while minimally compromising performance. By calculating half the digits, these chips have double the speed. If you are comparing chips across a standard of INT8 operations, which most studies do, then you are obfuscating the extra capability that newer chips get from being able to perform at FP4. Newer models are being trained at FP4, and inference also does not really care about less precision, meaning the capability to perform at lower numerical precision is a boon.

Memory Bandwidth

Measuring FLOPs alone also overlooks the critical importance of memory bandwidth. For most inference workloads, chip performance is not constrained by FLOPs but rather memory, since running a model means searching for and pulling billions of its stored values just to do a handful of simple calculations on each one before moving to the next. Instead of waiting for the logic to crunch numbers, the logic is waiting for the memory to fetch it numbers to crunch. A chip with ample FLOPs but insufficient memory bandwidth is like a chef with incredible knife skills but a single narrow hallway between the pantry and the kitchen, where she often has to waste time waiting in line behind the other chefs to get her ingredients. No matter how fast her hands move, the ingredients accumulate too slowly for the speed to really matter.

Frontier AI chips typically rely on high-bandwidth memory (HBM) to maximize memory bandwidth so that this downtime is minimized. Older chips use older HBM, which has worse memory bandwidth. The Hopper series uses HBM3e with a bandwidth of 4.8TB/s, whereas the Blackwell series uses newer HBM3e with a bandwidth of 8TB/s. (TB/s stands for terabytes per second, the rate at which the memory can deliver stored values to the compute units.) The newest Vera Rubin chips use HBM4 with over 22TB/s of memory bandwidth. Meanwhile, domestic Chinese chips have yet to crack HBM3; Huawei’s Ascend 910C uses (foreign-made) HBM2E with only 3.2TB/s of memory bandwidth. This means that despite Huaweiopolis’s superficial equivalence in FLOPs to Nvidiana, a large proportion of those FLOPs are unusable for inference workloads, since the logic units end up twiddling their thumbs waiting for memory, making query response times far too long.

Network Bandwidth

Lastly, network bandwidth — the speed at which data moves between separate chips or racks of chips — would severely limit the performance of Huaweiopolis’s cluster. Memory bandwidth is a limiting factor for within-chip communication because it determines how quickly data can move between a chip’s memory and its logic, effectively setting how fast the chip can stay fed with work. Network bandwidth — how quickly different chips can exchange data across the rack — is the limiting factor for between-chip communication, and network bandwidth is significantly slower than memory bandwidth. For an eight-chip cluster of B200s, memory bandwidth is an aggregate of 64TB/s, whereas network bandwidth is only 14.4TB/s. For training and serving inference on models, you don’t want to use network communication if you can help it because every time chips need to exchange data, they must stop and wait on one another; at scale, this turns communication into the dominant cost, meaning that adding more chips yields diminishing returns and eventually no additional performance at all.

Unfortunately for Huaweiopolis, if their strategy is to connect a massive blob of lower-quality chips to compete with a tiny cluster of higher-quality chips, they cannot succeed; network communication is unavoidable, and it will hurt. A Nvidiana cluster, with more power and memory storage per chip, can do a lot more within-chip before needing to resort to between-chip communication. A Huaweiopolis cluster will be running into this bottleneck a lot more frequently, and it will slow down operations. Particularly for training large models, where using multiple clusters of chips is necessary, the network bandwidth limitations will be crippling.

Jensen likes to dismiss this issue by arguing that “Huawei is a networking company” and dismissing the importance of HBM, but this is simply not the case. Networking will always be worse than memory bandwidth because data inside a chip moves over much shorter, more direct connections, while networking requires sending data across longer links with added coordination delays. Even God’s best NVL72 or Huawei optical fibre could not beat HBM in this battle because “beating HBM” would mean feeding the chip inputs as fast as its own memory can, which no external network can match.

FLOPs matter, but they are not the only metric. They are perhaps our best metric of comparison for now, but a proper comparison requires consideration of multiple factors. A naive equivalence on FLOPs of a Huaweiopolis cluster with a Nvidiana cluster hides the fact that the Huaweiopolis cluster will suffer in performance for both training and inference. This is not just a question of efficiency or speed. In extreme cases, the system can simply fail to train properly. Modern training requires tightly synchronized gradient updates across many chips, so if communication is too slow or inconsistent, those updates arrive late or out of step. The result is that the model is no longer being updated in a coherent direction — gradients do not reliably descend — and training can become unstable or fail to converge altogether, not just take longer or require more energy.

Conclusion

Aggregate compute matters, especially for the broad diffusion of AI across an economy. But when the question is whether a country will have the most powerful AI model, the quality and concentration of its best chips matter far more than its total headcount, and even more than total FLOPs.

There are signs that policymakers are beginning to internalize this logic. Moolenaar’s SCALE Act, introduced this week, still uses the rolling technical threshold framework but has shifted away from his earlier proposal to cap China’s aggregate compute at 10% of US capacity, which was the more aggregate-focused approach. Instead, it would permit exports only up to 110% of the performance of the best chips China can already manufacture domestically at scale, pegging the threshold to Chinese domestic capability rather than total compute. It is a narrower, more observable target, and it takes the quality-over-quantity insight more seriously than the aggregate headcount approach did.

No chip policy is going to be perfect, but the underlying logic is to focus the policy on the specific chips that matter most. We should be building enforcement around these crown jewels rather than solely around an aggregate FLOP count, and definitely not based on dubious chip counts!

Mood Music (Jordan)

Bucking a recent trend of Chinese AI labs moving away from open source, V4 was released under the highly permissive MIT license. It performs admirably on various benchmarks and leads the pack of Chinese open models, but did not close the gap with closed models from the US, with the authors themselves admitting in the paper that V4 is “3 to 6 months behind” state-of-the-art frontier models (though we think it feels further). And as we will discuss later, while its architecture shows progress towards indigenizing the Chinese stack, the model probably still relied on Nvidia GPUs.

Is V4 a letdown? Today on ChinaTalk, we bring you our takes alongside those from Chinese observers on:

• Troubles at the lab prior to V4’s arrival;

• Why DeepSeek’s idealism may not hold;

• What V4 did — and did not — achieve with domestic hardware;

• And why DeepSeek’s symbolism persists inside China, even after it lost the frontier

Translations were drafted with the assistance of Claude Opus 4.7, and then edited for accuracy and fluency. Bold markings added by the editor.

How V4 Got Here

Chinese tech journalists have doggedly followed the DeepSeek story. Zhou Xinyu 周鑫雨 of 36Kr, a prominent Beijing-based tech news outlet, has some behind-the-scenes scoops.

The reasons behind [V4’s] belated arrival are related to migrating its training framework from NVIDIA to Huawei Ascend, as well as to internal decision-making changes at DeepSeek. We learned that in mid-2025, DeepSeek ran into a relatively serious case of training failure.

“At the time, DeepSeek was facing the problem of re-adapting to chips,” one insider mentioned. “Internally, opinions on the direction of training were not entirely unified. Liang Wenfeng put forward some of his own demands, but it was difficult to find compromises at the execution level.”

However, contrary to outside speculation that the new model might support multimodal generation and understanding, V4 remains a language model. The decision to postpone multimodal generation training stems mainly from constraints on computing power and cash.

Multiple insiders told AI Emergence [a 36Kr sub-brand focusing on AI] that DeepSeek’s external financing window opened in mid-April 2026. Internally, the trigger was that DeepSeek needed more funding to train models with larger parameter scales, while also retaining and recruiting more top-tier talent.

Shanghai-based news site The Paper 澎湃新闻’s Fan Jialai 范佳来 compiled a comprehensive roundup of DeepSeek’s talent losses, losing core contributors to Tencent, ByteDance, Xiaomi, and DeepRoute.ai. “Across multiple areas — foundation large language models (LLM), agents, text recognition (OCR), multimodality, and more — DeepSeek has suffered losses of core talent.”

DeepSeek operates with the ethos of a frontier lab. Back in November 2024, we translated an interview with CEO Liang Wenfeng, done by Lili Yu 于丽丽 of Chinese media outlet Waves 暗涌. In it, Liang explained that DeepSeek was uninterested in product development, and that their goal has always been “AGI.” It was why, instead of adopting Llama architecture, they poured resources into the new model architectures behind R1. On why research, rather than products, was their raison d’être, Liang remarked:

For many years, Chinese companies are used to others doing technological innovation, while we focused on application monetization — but this isn’t inevitable. …

We believe that as the economy develops, China should gradually become a contributor instead of freeriding. In the past 30+ years of the IT wave, we basically didn’t participate in real technological innovation. We’re used to Moore’s Law falling out of the sky, lying at home waiting 18 months for better hardware and software to emerge. That’s how the Scaling Law is being treated.

But in fact, this is something that has been created through the tireless efforts of generations of Western-led tech communities. It’s just because we weren’t previously involved in this process that we’ve ignored its existence.

In this way, its closest American approximation might be OpenAI in its pre-ChatGPT Microsoft days: mission-driven, amply funded, and committed to nonprofit development of the AI frontier. If early OpenAI’s animating force was safe superintelligence, DeepSeek’s was a combination of AGI ambitions, open-source idealism, and national pride.

In the latest sense, it succeeded: DeepSeek became China’s national champion for LLMs. But that designation, and its founder’s high-minded aspirations, bogged down its research potential. Liang did not ride the DeepSeek wave in early 2025 — like Sam Altman did for ChatGPT — to build a scaled consumer product. Instead, he focused his team’s energy exclusively on the “hardcore research” he made his name on. By not building a revenue-generating business over the past twelve months or partnering with a Chinese hyperscaler, Liang bled talent and lost the lead he had over his domestic competitors.

More than any other lab, DeepSeek shouldered expectations to produce the proof-of-concept for Chinese-made chips, rather than follow other labs by relying on smuggled chips and Nvidia cloud compute abroad. This cost it financial runway and talent, and probably led to a failed training run that delayed V4 by months. The aforementioned 36Kr story reports that over the past year, DeepSeek recruiters were seen lurking the dorms of Peking University in search of Chinese majors to staff a new marketing unit.

After R1 came out in 2025, Jordan and Kevin Xu of Interconnected speculated on a podcast episode that DeepSeek, in the near future, could be lured by deals with hyperscalers or some other deep-pocketed entity. They were prescient. Per 36Kr:

As for the external trigger for pivoting toward open financing, several industry insiders speculate that it is related to the investment stance of a certain major company. Before opening DeepSeek up for financing, Liang Wenfeng and the top leader of that company had held several rounds of discussions regarding exclusive investment. But according to two sources connected to the matter, Liang Wenfeng did not agree to that leader’s condition of giving away a 20% stake.

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.

When we examined DeepSeek’s lack of a path to profitability and the enormous political pressure it had begun to shoulder, we thought the lab’s tragedy might have been foretold. Fast forward to now, and the 36Kr story just declared the “post-DeepSeek era”. A Qwen employee told 36Kr that “the golden age of nonprofit AI development is over.” But the article also acknowledges that DeepSeek, in just one year, shaped China’s AI landscape. Beyond its model architecture innovations and the open source ethos, its flat internal hierarchy, focus on emerging talent, and AGI-inflected open research culture have all influenced management decisions at other labs hoping to replicate its success.

American Training, Chinese Inference?

V4, ultimately, was still trained on Nvidia chips. However, Huawei on April 24 confirmed that its own Ascend supernode cluster will be able to support V4. Earlier this month, DeepSeek did not give Nvidia and AMD early access to V4, perhaps superficially signalling distance from Western chipmakers. Popular tech blogger Digital Life Kha’Zix 数字生命卡兹克 examined V4’s technical report, and returned with four observations regarding how the model was optimized for Chinese-made hardware.

1. V4 has introduced MXFP4 into its post-training and inference systems.

Although training still uses the NVIDIA ecosystem, using MXFP4 in post-training and inference essentially means that DeepSeek is moving toward open low-precision formats and multi-hardware adaptation. It can adapt to domestic chips such as Huawei Ascend, Cambricon, Biren, and others, reducing its reliance on NVIDIA’s FP8 ecosystem — especially during inference. That would make it a genuine domestically-produced model running on a domestic ecosystem. …

2. V4’s underlying kernels are no longer written entirely in CUDA, but instead in a domain-specific language (DSL) called TileLang. DeepSeek hopes that low-level operator development won’t be completely locked into CUDA, but will instead use a higher-level language to describe computations and then compile them to different hardware as much as possible. This is seriously impressive and can greatly reduce migration costs.

3. V4 has specifically developed a fused kernel called MegaMoE, designed to reduce communication waiting in expert parallelism. It has already been successfully run on Huawei Ascend.

Putting these three points together, the direction is crystal clear: V4 is, from top to bottom, a model designed for domestic chips.

This really isn’t some patriotic story. Everyone knows how scarce computing power will be in the future, how slow computing power production is, and—under the acceleration of Agents—how terrifying the token consumption will become.

With computing power being choked off, no one has any good options. Just look at how a model as excellent as GLM-5.1 has been limited by inference compute.

The computing power game is, in many ways, a top-level geopolitical game.

DeepSeek V4 is the reality forced into being by this computing power struggle.

There was a curious footnote attached to DeepSeek’s official announcement of the V4 models:

Due to constraints on high-end compute, V4-Pro’s service throughput is currently limited. Once Huawei’s Ascend 950 supernodes ship in volume in the second half of the year, Pro’s pricing is expected to drop significantly.

The compute story probably demonstrates that Chinese models like DeepSeek will fall further and further behind Western counterparts. Western models are increasingly being trained and run on Blackwells and eventually Rubins, which can support FP4 numerical precision, effectively double the compute from previous generations that can only go to FP/INT8. DeepSeek has been stuck using old Hoppers, which only go to INT8; to have any chance of catching up, they will have to pray Huawei’s Ascend 950, which supports FP4, will be produced in sufficient numbers. According to Reuters, Huawei plans to ship 750,000 of their Ascend 950PR this year; for reference, that is just one week of quality-adjusted American chip production.

“The People Long for DeepSeek”

When the “DeepSeek moment” arrived in 2025, it didn’t only represent indigenous technical capabilities for China. For some developers and average people, it also meant having genuinely affordable access to frontier AI for the first time. American frontier labs have always restricted chat and API access in mainland China, and while many Chinese users found ways around the firewall anyway, DeepSeek was a model they could use with no fuss and, for a brief window, nearly comparable performance.

But after a year, there are now far more domestic models for Chinese users to choose from, embedded into many real-life applications. Meanwhile, OpenAI and Anthropic seem to have cemented their lead. With soaring demand and mounting financial losses, AI companies have no choice but to offload more costs onto paying customers. Fewer and fewer people can afford to extensively utilize frontier models. China’s OpenClaw craze earlier this year showed many people the true costs of AI, as their home-cooked agents guzzled tokens and left them with expensive bills.

In 2017, blogger Fang Hao 方浩 published a viral article titled “The People Long for Zhou Hongyi”. Zhou was the founder of security software firm Qihoo 360 and a famously pugnacious figure in China’s tech industry. Written at a time when Alibaba and Tencent were consolidating their monopolistic positions in e-commerce and social media, Fang couched pessimistic future predictions in irreverent humor: as Chinese Big Tech cannibalized opportunities in the private sector indiscriminately, it would leave average consumers worse off.

Last month, Su Yang 苏扬 of Tencent’s tech media blog wrote a sequel: “The People Long for DeepSeek”. He pushes back on Jensen Huang’s “tokenomics” rhetoric:

When token usage costs can’t be brought down, and when the effective return on investment remains unclear, aggressively pushing token consumption — even tying it to performance reviews — amounts to manufacturing token anxiety. Calling it manufacturing AI anxiety wouldn’t be an overstatement either.

Looking back a bit further, Jensen Huang also called on tech industry leaders to speak prudently and avoid stoking irrational public fear of AI technology. That’s essentially telling the whole industry: stop suppressing AI by manufacturing panic — you all need to keep the tokens burning.

But the question is, who’s going to solve the price problem? Will it be the long-delayed DeepSeek V4?

Su expands on the price issue in a follow-up post. While he is ultimately optimistic about the future of competitive innovation in China’s AI industry, he thinks DeepSeek will no longer be a singular flagbearer:

Broadly speaking, in 2025, China’s open-source forces reshaped the global AI landscape. By 2026, China’s AI development has entered a stage of exporting capabilities.

From the perspective of the global AI industry, the diversification of technical pathways has invigorated talent mobility and strengthened supply-chain resilience. For downstream application developers, having multiple suppliers to choose from means greater bargaining power and lower lock-in risk.

Another encouraging feature of China’s AI narrative is that the market has yet to be monopolized by a handful of oligopolies — a positive sign for competitive innovation and talent-ecosystem building, and one that also helps build cluster-level advantages in the U.S.–China AI competition.

…

In the landscape of full-ecosystem competition, DeepSeek — whose principles generate its force, with breakthroughs at the foundational layer — still holds advantages, but its weaknesses are equally clear: it lacks the industrial ecosystem support of an IT giant, its product application features are relatively thin, and its multimodal and agent ecosystem still need strengthening.

Is Coding the Way Forward?

V4’s coding capabilities have grown significantly, potentially signalling that DeepSeek, after the success of products like Claude Code, also sees promise in coding agents. Programming blogger Large Model Observer 大模型观测员 tested V4 on software engineering projects, finding two pros and two cons:

First, broad programming knowledge. Across the four engineering projects [that the author tested V4 with] extensive niche-domain knowledge is essential. Without it, you can end up unable to fix even simple bugs, such as a macOS application failing to display its window properly because the storyboard wasn’t correctly configured. V4’s knowledge base essentially covers these less mainstream areas, and when faced with various edge cases, V4 Pro can pinpoint the root cause of a bug directly rather than guessing — much like GPT and Opus. … V4 Flash isn’t far behind Pro on broad-strokes knowledge; Flash mainly falls short in edge-case knowledge and tends to be stumped by non-obvious bugs.

Second, low hallucination over long context. Because the engineering tests use a mode in which features are layered on round by round, the later rounds often require the model to re-read the entire project and locate every related detail when a global modification is requested. This is no problem for the likes of GPT/Opus, but it’s a real hurdle for domestic Chinese models. V4 Pro and Flash, at the high and max tiers, can essentially maintain a quite low hallucination level, with bug rates in downstream flows over long codebases still kept low.

Third, occasional lapses in attention. When projects are large and requirements are many, V4 Pro at the high tier — constrained by its thinking-budget allocation — has some probability of randomly dropping certain implementation details. The saving grace is that with a reminder and one or two rounds of self-testing, the issues can almost always be fixed. …

Fourth, an unfussy approach to architecture and UI. V4 largely inherits DeepSeek V3’s thinking on architectural design — not particularly tasteful, not refined, but not slapdash either: the layering and decoupling that ought to be there will be there. It can’t deliver the kind of polished, clearly master-crafted architecture you see from Opus at a glance. UI is the same story — direct output isn’t outstanding, with the occasional touch of refined expression, but most of the time it’s just at the basically-usable level. The high tier can occasionally have an even lower floor, with insufficient consideration. If the development workflow includes a design spec to follow, this is not a big issue. But for pure vibe coding, getting a satisfactory result requires a lot of rerolling.

Could V4 do for AI coding what V1 and R1 did for LLMs — democratize access to the frontier, especially for the Chinese user base? It’s not impossible, but the model faces ample competition among open-source peers. A quick comparison of leading Chinese open models’ token prices, in RMB:

DeepSeek’s prices are competitive, if not an obvious standout. BusinessAlert 知危 summarized it as such:

By now, users are no longer impressed by chain-of-thought. At most, it’s an engineering technique that boosts accuracy by throwing more compute at the problem, and in coding-agent scenarios it’s probably ignored most of the time.

The ceiling of [V4]’s capability makes it unlikely to play a leading role in real-world programming tasks, and as an executor it’s too slow. … All in all, from the perspective of the cases we tested, DeepSeek V4’s performance wasn’t as good as expected, and its capability seems not particularly stable either. But then again, the official technical report itself openly states that there’s still a gap between it and top closed-source models, and that this update merely narrows that gap — so the result isn’t surprising.

Still, as the saying goes: take another look at the price. It’s this cheap — you can put up with it.

While the Chinese-open-models price war looks fierce from the outset, it belies fundamental challenges: the business model is not yet clear, and the ecosystem is starved for funding at a much more severe level. We’ll leave you with Nick and Jordan’s recent analysis of why some Chinese labs are going closed-source, and why DeepSeek does not change the core political equation:

China’s funding environment for AI is orders of magnitude smaller than America’s. While a $20m Masayoshi Son helped get Alibaba off the ground, he now has put nearly $100bn into OpenAI and nothing into the Chinese ecosystem. Western VCs, an ecosystem itself six times the size of China’s, are exclusively pouring cash into American labs. Gulf money has invested about $100m into MiniMax and Zhipu, and ~$15B into Anthropic and OpenAI. …

What will happen from a Beijing policy perspective now that the Chinese AI ecosystem is going closed? Probably not much. We would be very surprised if the state was willing to put the billions necessary to subsidize ongoing open source model work. Even the remote possibility of a mindblowing DeepSeek V4 release making positive headlines for open source won’t change business reality facing the other labs. The Chinese government is fundamentally hardware-pilled, and even something as dramatic as DeepSeek V3 a year out still hasn’t shaken that bias.

DeepSeek Waxes Auto-Poetic

Jordan: I gave DeepSeek V4 this article and asked it to write a poem of how it made it feel.

And a Chinese one:

Joining us: Bryan Clark of the Hudson Institute and former submariner; , former Green Beret now in defense tech; Eric Robinson, former OSC/NCTC analyst and 101st Airborne officer, now a lawyer; and .

We discuss…

• Why on earth we fired L-RASMs at the Iranian Navy — and what that means for the Pacific

• The case for modular weapons over exquisite Cold-War-era munitions

• Why this admin is telling the press “We can’t win a war over Taiwan” — what leaking this that actually means

• The Special Forces Romeo who made $400K on Polymarket betting against Maduro

• The Phelan firing, the waffle bar, and Driscoll’s survival odds

This is a transcript of a podcast!

Listen now on your favorite podcast app.

Out of Ammo

Eric Robinson: Why is the Pentagon starting to scream to the press about war stocks? Something that countries should probably keep in their back pocket is how many precision weapons you still have available. And in an open society like the United States attempts to be, some of these secrets are difficult to conceal. The number of specific munitions gets published by Congress, the Pentagon speaks about it openly. You can assess the capability of the United States Navy to carry certain warheads into certain locations. You can look at the Air Force and determine what kind of ordnance can go over a target. So there is a baked-in ability to do informed speculation, and people are doing that.

When General Caine and the Secretary of Defense go on TV and talk about the number of targets that have been struck in the Iran war, you can start to take an X and do a bunch of minuses beneath it and reach some conclusions. But we’re starting to see incremental precision coming out informally from the Pentagon that’s indicating the American ability to fight a sophisticated war is substantially degraded because of this war with Iran.

Bryan Clark: I think this is a case where people in the Pentagon are trying to get the attention of the president via the press. It’s a time-honored tradition. Not honored, but — it works. So yeah, definitely people trying to get the attention of the president by leaking to the press how we’re low on munitions and maybe we need to wrap up this Iran war post-haste.

Justin: I wouldn’t say it’s honored, but, anyways. This is one of those natural tensions we’re starting to see bubble up. You have these OPLANs and CONPLANs that are supposed to be in place for contingency operations. They are predicated on certain availability of weapons systems in the first hour, first 12 hours, first 24 hours. To make it to 72 hours and beyond, we need to have certain things in place. And you’re probably starting to see people who have very vested interest looking at INDOPACOM and making sure that their CONPLANs are sufficiently funded and have a robust capability. Looking at what’s happening in CENTCOM — to no fault of CENTCOM, but because they were directed to do this — and saying, that thing over there seems to be taking away my ability to do the thing that you’ve told me to always be ready to do, the most dangerous. I think that has a lot to do with it as well.

Eric Robinson: Yeah, hyperpowers have constraints.

Tony Stark: And to lean into that — because I think where some people try to get around this is, but we’ve already ramped up production, we’ve ramped up production in the last few years after Ukraine. The problem is we were on minimal sustaining rates for way too long. It’s not like we just had a max capacity magazine and we decided to empty it out and we can get back to it in two years. We were already low. We’ve burned way too much, and now just to get back to that previous low standard, it’s going to take years.

Justin: We have over-indexed on exquisite technologies because we have the ability to produce them. The problem with exquisite technologies is they take a very long time, they have very tenuous supply chains, and you can’t really do exquisite in high capacity.

Wrong War, Wrong Weapons

Jordan Schneider: OK, so with Phelan gone now we’re not doing the Trump ship. Is this an opportunity to kind of reset in a potentially better direction?

Bryan Clark: What Justin just brought up — maybe this is a good chance to rethink the munitions portfolio and say, do we want to refill our munitions stocks with the exact same thing we just spent on Iran? Because maybe JASSM-ER is not the only weapon we want to have in the inventory for doing air-to-ground attack. Maybe we need some of these low-cost weapons. There’s a bunch of options out there that the Pentagon has been funding development of, but just never funded procurement of. This is a great chance to rethink what the portfolio should look like and rebalance it towards these more modular weapons — not even really lower-end, but more modular weapons that maybe don’t have quite the performance of the preferred munitions, but you can buy them at much higher volumes and the production is much easier because they’re modular and some even use commercial components, like this ERAM missile that the Air Force developed.

The other thing it makes me think of is why are we using JASSM-ERs against Iran? I get that you wanted to use PRSM to test it out and the Army guys like to show off their toys, fine — and we didn’t have that many so it wasn’t a big loss. But to go out there and burn through a bunch of our JASSM-ER stocks and our Tomahawk stocks — why are we launching Tomahawks into Iran? Supposedly they have no air defenses.

Eric Robinson: You’re referring to L-RASMs against the vaunted Iranian Navy. Bryan, you’re at the heart of the question. Was there an American strategic assumption — we would imagine this is an Obama issue, this is a Trump I, a Biden issue, there are numerous parents of this failure — but had the Pentagon ever baked into its war planning that we were going to conduct a military campaign of this style, where you’re going to just go after targets without a political objective and the expectation being that if you employ a sufficient amount of violence, if you do gunfight properly, all of a sudden strategy emerges? That there is this baked-in collective understanding that if the United States is going to go to war against one of the big four threats, there’s going to be a Clausewitzian strategic approach to it? We are not there.

Instead we’ve got people doing wheelies in Lamborghinis, and it looks really cool and it gives you your sizzle reels, but we are still at an impasse with an Iranian state that refuses to fundamentally break down — and using your entire inventory of L-RASMs, which for viewers that haven’t engaged with this before, is an advanced anti-ship missile almost expressly designed for the United States Air Force to employ against the People’s Liberation Army Navy in Pacific contingencies. Instead, we went after the Iranian Navy, which is sort of like expending L-RASMs on the Austrian Navy. It doesn’t matter. But here we are. We are disaggregated from strategy. And now we’ve got empty war stocks, so we’ve got a six-year pipeline to try and restore this, assuming Congress gets behind it and funds it.

Justin: Yeah, I mean, we sank the SS Minnow with an L-RASM, so we’ve done well.

Eric Robinson: Yep, we got the good ship Lollipop and the yellow submarine.

Justin: Even for the PRSMs — while I get that we wanted to showcase what PRSM could do, there’s a good argument that PRSM is most useful in the land war in Europe. It’s not super useful in an over-the-sea battle in the Pacific because 900 kilometers is a —

Tony Stark: The other thing about PRSM is it solved one problem. It still didn’t solve the mass problem, which was the actual Russian artillery problem. It is now for-purpose in the Pacific, which actually makes it effective. It still wouldn’t have been effective in a Europe fight because it doesn’t solve the close-in conventional artillery battle.

Justin: It solved the wrong problem in its development. Then they kind of came up with the potential to have a solution within the Pacific. But still, the place where it’s the most useful, it’s not useful. It’s not a great answer.

Eric Robinson: Ultimately, the LUCAS system is probably just as good and less expensive in certain environments for long-range strike.

Justin: I imagine PRSM still has the same anti-personnel capabilities that some of the marks of HIMARS have, which are very nice, which LUCAS won’t have, which do make it more formidable against mass formations.

Eric Robinson: For the audience, LUCAS is — I’ll give credit — the Department of Defense innovated and stole a model from the Iranians. They didn’t go and buy a turdcopter from Silicon Valley. They got something that they knew worked on the battlefields and copied it relentlessly. And maybe the IRGC will go to court in the Southern District and go after the United States for stealing its IP, but in this war, the United States tested what we effectively stole from the Iranians and repurposed.

Bryan Clark: My concern now going forward is we’ve got this big defense budget that we’re looking at. And you’ve got this lobbying campaign by people inside the Pentagon, over at INDOPACOM, to basically restock all these weapons that they’ve built their plans around — not new weapons, but existing weapons. That’s going to take up a huge chunk of the budget. And we want to make these long-term commitments on weapons production to try to get production capacity increased, because a company like Raytheon is not going to expand production capacity unless they have a long-term commitment from the government. So if you do a five- or seven-year multi-year procurement, they’re going to be willing to support that level of production.

But that means you’re locking in a huge chunk of investment over the next seven years that’s going to be devoted to weapons designed in the Cold War or the immediate aftermath, and designed primarily to go after the highest-capability threats posed by China. It just seems like we’re going to lock ourselves into a portfolio that’s going to have the same challenges as our existing portfolio. It’s never going to get big enough, and it’s never going to have the ability to surge in the way that maybe a new portfolio of weapons more modular or more focused on one-way attack drones might.

Justin: The other thing worth thinking about is the consternation we had even in the Biden administration over giving HIMARS to the Ukrainians. If we’re really talking about enabling partners in the Pacific, the next question is: do we build these high-end systems that we’re going to instantly have concerns about giving to partners — the Philippines, the Japanese, the South Koreans? Because if that’s the case, then it’s not really a capability if it isn’t in theater and controlled by people who can use it. The department really has to come to grips with this and go to Congress and say, we need weapon stocks that we can actually give to our allies that allow them to present a credible deterrence or response capability.

Bryan Clark: Another aspect of this is adaptability. In Ukraine, they found that the Excalibur rounds we sent were quickly obviated by Russian electronic warfare against GPS, and same with GMLRS.

Eric Robinson: Excalibur is a 155-millimeter round that can be guided by GPS. In the global war on terrorism, Excalibur came out because the wars in Iraq and Afghanistan were uncontested electronic warfare environments. It enabled individual battalion commanders to effectively choose a 10-digit grid — a one-meter spot on a battlefield — and say, I’m going to blow that up. So Excalibur revolutionized precision strike. To Bryan’s point, when it got to Ukraine and the Russians — a far more sophisticated adversary — jammed global positioning satellites, those rounds were dumb rounds and were not particularly helpful.

Tony Stark: GPS is this wonderful invention that we’ve had around for over 40 years. The problem is, because of the way it functions, it’s incredibly easy to jam. If your entire targeting package is built around being able to find that 10-digit grid, that’s a massive complication, and the US still hasn’t quite gotten around it.

Bryan Clark: All these weapons we’re talking about — L-RASM, JASSM, SM-6 — they all use GPS to some degree as part of the guidance solution. Even if it has a seeker, even if it’s getting a guidance update via the radio, it’s still using GPS to orient itself in the world. If it loses that, it’s very difficult to orient itself. You now have to either adapt the inertial navigation units onboard to make them more capable, or provide a constant radio signal to tell it where it is, or have other sensors onboard that allow it to predict its location based on star shots — there are companies doing that — or detecting emissions from cell towers or TV and radio antennas, which you can use to geolocate yourself.

The problem we have with these legacy weapons — these kind of high-end weapons that are highly integrated, like Excalibur — is they’re too hard to modify. We still haven’t really fixed Excalibur to address this GPS jamming issue. And they’re desperately trying to fix GMLRS to make it able to use other sources of navigation, like radio emissions. So we’re going to invest a bunch of money and make long-term commitments in weapons that are difficult to adapt, because we don’t know what the next countermeasure from the opponent is going to be. It’s GPS jamming today, but it could be something else tomorrow that goes after their seeker mechanism or their ability to orient because we’re going after something else in the electromagnetic spectrum. There’s all these opportunities for move–countermove competitions that these weapons don’t give you the ability to respond to.

Leaking for $$ From Congress and OMB

Eric Robinson: One point I want to build on that Bryan really helpfully said earlier: these stories are coming out in the press because somebody is trying to signal to the White House. It is arguably stakeholders inside the Pentagon. It is probably also the primes for these dastardly-six companies who are trying to communicate — one to Congress, two to the White House — hey, pay us money, let’s get reconciliation through, let’s get the president’s budget through so we can start ramping this process up.

It reflects an interesting communication mechanism in Washington right now. Because there’s really not a whole lot of value in going to Pete Hegseth and talking about munitions stocks — it’s just not something he cares about. You can go to the deputy, and the deputy absolutely cares about this, but he’s not necessarily going up to glad-hand on Capitol Hill. So you do the scattershot communication strategy to raise the profile of vital issues. Then you can go to OMB, to Senator Wicker at SASC, to Representative Rogers, and say, OK, we do have a crisis, let’s start to solve it. We’re witnessing the creation of an information environment where once you get past the official bluster from the Secretary of Defense, there is an authentic problem now that has to be resolved.

Jordan Schneider: Here’s the problem. You know who else uses remove-paywalls on the Washington Post? Our allies in the Pacific and the Chinese government. So this is not ideal the way you’re going about this, if we’re trying to preserve deterrence capacity. One of the quotes from a senior administration official was, we can’t win a war today over Taiwan. If you’re a Taiwanese politician and you read that, what do you feel like? It is a signal of the breakdown of this administration and how they’re trying to rebuild.

Eric Robinson: It is not new. In the late 1950s, when you had giants of the Senate like Jack Kennedy and Arthur Vandenberg, they would go to the floor and talk about weapon systems. You had carrier generals and missile generals and bomber generals — or senators. They all had their pet members of Congress, and they would basically dump the national secrets into the Congressional Record. And the Soviets were monitoring that.

So Congress, through its position of supervising the executive — which is the core function of the American Republic — has this habit and maybe even a responsibility of conducting aggressive due diligence on what the executive is doing. And we’re witnessing that now. It’s ugly, and I don’t mean to minimize it, but it’s not new.

Can We Actually Lose?

Justin: I will push back a little bit, Jordan. I don’t see this as demonstratively different than the purges that occurred with the PLA Rocket Force when Xi found out there was a lot of corruption within the Rocket Force. The difference is that this comes from multi-facets, and it’s not like people like Bill Bishop pulling out reporting and figuring out what’s happened. So it’s slightly different in the volume and the tenor, but it’s not indifferent. I don’t think we think China is in the best spot either. That being said, saying we’re not in a position to win a war is different than saying we’re in a position to lose a war.

Bryan Clark: Obviously this guy tried to grab attention with what he’s saying, but I feel like what he means is we’re not able to win a war on the terms we want to win it on, using the things we want to use to win it. There’s lots of ways that a war over Taiwan could play out. We’ve wargamed a lot of them. And usually it results in the Chinese losing. The main differentiator is how much do we lose in the process? So when they say we’re not in a position to win a war, normally it means the losses we’re going to incur don’t seem attractive. And it may be enough to cause a president to be reticent about intervening on Taiwan’s behalf. So it’s a lot more about how well does it go rather than are we able to stop an invasion of Taiwan by China.

Tony Stark: It’s a pretty good bet that the best description for that fight is a drunken bar fight. As nice as you want to make it, it’s going to be ugly just by the geography, by the munitions burned. There was reporting this week that some of the most recent purges in the PRC are due to skepticism over whether the weapon systems actually work — going beyond the Rocket Force. Xi Jinping has to have some concerns if he’s watching all these Russian SAM systems, which the Chinese cloned, just getting burned around the world. We might be out of ammo; their ammo might not work. It’s a great time all around.

Bryan Clark: At least our stuff works…

The Special Forces Gambler

Eric Robinson: Speaking of opponents — this is a good opportunity to talk about a perhaps lesser opponent in the Maduro regime and their Cuban personal security detachment, and Polymarket being a vehicle for not informed speculation but direct knowledge of events being used to generate financial rewards. Yesterday, the Department of Justice revealed an indictment of a Special Operations soldier at Fort Bragg with knowledge of pending action against the Maduro regime, who elected to speculate on Polymarket and apparently got, according to DOJ, $400,000.

We are in a world that the president has described as a casino, and he’s not particularly concerned. But this incident — a master sergeant in Army Special Operations — is now being held to account for insider trading.

Tony Stark: I want to play this through. You’re going to place this bet, you win all this money. You’ve got to file that on your taxes, right? Especially when you have a security clearance — people notice the sudden windfall, which is a literal insider threat mechanism: $400,000 to an enlisted man.

Jordan Schneider: But it’s in crypto in some account. That’s the thing — the taxable earning — presumably if you want to repatriate it into a US account, someone’s going to start asking some questions.

Jordan Schneider: The weird thing about Polymarket is it’s just this crypto setup. A month ago they started saying they were going to deal with insider trading — that it wasn’t cool anymore. And presumably the Department of Justice gave them a call at some point between February and now.

Eric Robinson: It’s not like an American sports book.

Jordan Schneider: It is wild that they thought this would be a sustainable thing. And it’s also so predictable. How many people knew this was going to happen? Probably thousands, right? And you only need one, and they’re not going to bet 50 bucks. Though Kalshi just did catch a handful of people in their primaries betting on whether they were going to run or not, which —

Justin: I bet you that’s how it started. This is that slippery ethical dilemma. We had been already talking about it with other actions — Midnight Hammer, we had been talking about Boots on the Ground, we had been talking about the strikes on the drug boats.

Eric Robinson: There’s a substantial American presence. There was obviously something coiling in the Caribbean. It was not lightning.

Justin: That’s the pernicious effect, and where it’s going to get super regulated. To caveat what I’ve seen about the guy — he was a Special Forces guy, but he was a Romeo, a radio operator. He was a communications specialist supporting Tier 1 units. He was not actually one of the operators in the Tier 1 unit. So there is a distinction just there, because they go through a slightly different selection process. But if that’s all correct — before operations, before Afghanistan, before Fifth Group went into Afghanistan, they took all the teams and they isolated them. It was 2001, so it wasn’t a huge deal. They put the teams in a tent, and the only way they could ask questions or get information was to write it down and hand it to someone who’d go out, get it from the real world, and bring it back. Maybe that’s what we’re going to go back to — we’re going to start isolating these guys well in advance, and then they’re literally going to be cut off from anything that is not classified systems until the missions are over. But what are the effects of that?

Eric Robinson: Modern information markets — whether Predictit, Kalshi, or Polymarket — encourage individual action to spike these markets. And it’s not unique to American special operations.

This is not new behavior, but it’s being dramatically exacerbated. It used to be you had to be a sports hero in order to throw a game, or you had to have sufficient financial backing to witness something coming and then place a bet in commodities markets — you could look at the price of oil, there were ways to do this. But sleaze is now super democratized. There are such a great variety of markets, and there are ways for individuals to push events one way or another.

Jordan Schneider: We’ve had congressional insider trading for decades now. Which is the justification one congress member made yesterday saying this person should get a pardon and just have to give their money back. But it starts with that — there’s some level of permissivity coming from the legislators themselves. Which is not to say that should be legal or this stuff should be legal. But there’s a wholesale reckoning with the whole graft ecosystem that really should happen sooner rather than later.

Tony Stark: There’s a bipartisan bill in Congress right now on this topic. But of course, I think it only covers the troops and not members of Congress or any other senior federal official. I’ve got an article in draft somewhere about all these reforms that need to happen. One of them is just — if you are a US government official in any capacity, you do not get to bet on anything. You want to bet? Go to Vegas. But you do not get to place Polymarket bets or anything else. You don’t get to use the stock market, sorry. It’s become absurd to the point of ridicule at this point.

Graft Rot

Justin: If you want to have responsibility, you have to be beyond reproach. Congress also needs to get their shit together. But the fact that they commit a wrong doesn’t mean that other people should be allowed to commit wrongs. I get that people will use that as a justification, but it is what it is.

Bryan Clark: It creates perverse incentives inside the department. On the operational side, weird behaviors could emerge if you get the chance to bet on an op you’re going to be a part of. And even more important — if you’re making decisions on this new big defense budget, those decisions obviously could be useful in the equities market. That information might be something you could use for your own insider trading. At the SES level, and at the level of people who make decisions about money, you’re supposed to be reporting all those potential exposures. But I wonder if we’re really starting to throw all that by the wayside because at the top levels we’re sort of ignoring it. The senior bureaucrats in the Pentagon and elsewhere are now not really worried about being transparent about their holdings, because their bosses and their bosses’ bosses aren’t.

Tony Stark: One of the ways CIA allegedly made inroads into the Chinese Communist Party in the 2010s was because of the way the graft system worked. You don’t want to create that in your national security apparatus — oh, here, I’ll give you an insider tip for a Polymarket bet if you give me information. That is the type of stuff that intelligence services will attempt to use in order to gain intelligence.

Justin: If they haven’t already at least attempted it, I would be shocked.

Bryan Clark: It undermines the efficacy of the organization. Which is what you’re seeing in the PLA — these purges are in part because people are not effective in their jobs because they’ve been corruptly operating their fiefdoms. You end up with things like weapons that don’t have warheads, because it’s cheaper — you can pocket the difference if you buy the weapon without the gas or without the warhead. Are we going to start to see that kind of behavior because there’s all these opportunities for malfeasance that we’re ignoring and allowing to fester inside the department?

Tony Stark: There was a tolerance even among the voting base of, well, if they’re corrupt but they’re effective, then it’s OK. That was the CCP’s model for a long time — graft and corruption are allowed because that’s how you build your patronage network. Don’t get caught. That’s the stakes of the game. But it’s quite clear from the PLA purges that corruption corrupts absolutely, and it will eventually, no matter how effective they were in the first place, it will degrade your readiness.

The Phelan French Kiss

Jordan Schneider: Should we do a Driscoll death-watch check-in now that Phelan’s gone?

Eric Robinson: Let’s do some Kremlinology. We lost the Navy Secretary this week. Long may he rest in his fantastically lucrative art collection and nice collection of poems. He will be fine.

Jordan Schneider: Can we read that just for posterity? I think it’s worth it. Was it a Politico article?

Justin: I think it was the Wall Street Journal.

Eric Robinson: The former Secretary of the Navy had a long-standing personal relationship with the president. He wasn’t a random selection. He is a West Palm Beach resident. He’s a Mar-a-Lago diner. He probably hits the waffle bar on the weekends. He was a bundler. He raised tens of millions for the president’s reelection campaign. He didn’t really have a national security background — he’s an independently wealthy financier.

Jordan Schneider: All right, let me read it from the WSJ.

John Phelan sat in the lobby of the West Wing for more than an hour Wednesday night, waiting to see if his longtime friend and neighbor, President Trump, would save his job. He would leave disappointed. That afternoon, Phelan, the Navy secretary, had received a phone call from his boss, Pete Hegseth, asking for his resignation. Phelan had spent most of Wednesday on Capitol Hill meeting with lawmakers about Navy shipbuilding.

A few miles away at the White House, another gathering was taking place that would decide his fate, according to US officials. Hegseth and his deputy Feinberg had made the argument to Trump that Phelan wasn’t moving quickly enough on Trump’s shipbuilding priorities, especially the Golden Fleet and increasing reliance on US use of steam. The Navy, they determined, needed new leadership.

Phelan made a round of calls, including to the president’s executive assistant, saying he needed to speak with Trump. Phelan then headed to the White House. Once the president had a spare minute Wednesday evening, Phelan asked to keep his job, but the commander in chief backed Hegseth’s decision, according to a senior official.

Eric Robinson: There are some other interesting reveals. The people commenting on the entire saga are saying it was a combined decision between the secretary and Deputy Secretary Feinberg — that we often talk about — and how he’s disciplining things throughout the Pentagon. There are some organizational decisions Deputy Secretary Feinberg made that really undercut the Secretary of the Navy. He captured the submarine program office and put it under his direct supervision. He wasn’t inviting the Navy Secretary to meetings. That strikes me as the death rattle of his tenure.

Ultimately, Stephen Feinberg is sufficiently sophisticated to not promise the Trump battleship — the Defiant class, whatever you’re going to call it. He knows it’s fantasy. He might as well promise an Imperial Star Destroyer; it’s just never going to happen. The level of personal animosity that existed — with Secretary Hegseth firing the chief of staff to the Secretary of the Navy, the personal relationship between the Undersecretary, Hung Cao, and the Secretary — there’s this thicket of interpersonal hostility that boiled over. John Phelan went to Washington without much of a constituency, and I think he found himself without a friend. But he’ll be back in the president’s good graces at a personal level. He didn’t break from the phalanx. He just didn’t really deliver.

Bryan Clark: I’d like to commend Secretary Hegseth for his apt bureaucratic wrangling. Because neither he nor Feinberg like the idea of the Trump battleship. They don’t like the idea of the Golden Fleet. They have certain priorities that they want the Navy to pursue — unmanned systems, submarines, electronic warfare, some other stuff. Those are the things they want to focus on, new technologies that help us better address a more contested environment. And you’ve got the Secretary of the Navy freelancing, pursuing the battleship with the president, and then this new frigate, which is just an effort to make the fleet bigger and spend a bunch of money on shipbuilding. Which is not a bad thing in general.

Driscoll Lives…For Now?

Justin: They promoted him to constituent again — it was good. This is a demonstratively different situation than Driscoll finds himself in. Driscoll and Hegseth seem to have personal animosity for whatever reason — we can speculate.

Eric Robinson: Ranger tab, Ranger tab…

Justin: When you’re an infantry officer. But I don’t know that Secretary Hegseth would be able to create the groundswell for Driscoll, because from everything you can see, Secretary Driscoll and the Deputy Secretary get along and seem to be in lockstep on most things. If I had to guess at who was better at bureaucratic machinations — Secretary Hegseth or Secretary Feinberg — I’d imagine it’s Secretary Feinberg who’s like, no, no, don’t do the public pronouncement thing. Do it this way. I don’t see him offering that same level of support for Secretary Driscoll’s removal.

Tony Stark: Let’s step back and look at the broader political landscape, because that’s part of the differentiation between Driscoll’s position and Phelan’s. Driscoll has an ally in the Vice President — they’re close friends. Driscoll is also known to be effective within the building. He is liked by his own service. On top of the fact that he’s liked on the Hill, he’s liked in the White House. That is substantially different from what SecNav’s position was.

And that matters at a time when you have an already schisming Republican Party, where Driscoll could be one of the fault lines. From a Washington, DC, politics standpoint — the grassroots does not care — but in terms of the various factions in Washington, that is a schism you don’t want to happen, particularly before midterms, especially when most of the party at this point has a 30% approval rating and is looking to what comes next. And it’s very clear that with Vance — I think at 42% — is the most likely candidate for 2028 right now. You don’t want to piss off an ally of Vance if you want to have a future in the Republican Party, at least as it stands right now.

I’m not saying Driscoll is untouchable, because I could say that and in five hours a phone call could happen and he could be gone. But I think he is a lot more protected, both through his own actions and the broader politics of the party, than Phelan was.

Jordan Schneider: How do you square that with all these generals and his chief of staff and random advisors that work for him getting —

Tony Stark: Generals are the domain of the Secretary of Defense. That is how Trump sees it. They’re not his appointees. And you can get away with well, they were Biden generals. I think that is fundamentally it. I also think those firings have caused part of this schism — it’s made it worse. And I think Driscoll is the fault line for something you can’t walk back from.

Justin: When you have Republicans coming out and openly supporting — like Representative Cole coming out and openly supporting Driscoll while excoriating Secretary Hegseth on the decision to fire General George, saying you’re exactly the right person at this time — that’s a strongly worded letter said to the media.

Jordan Schneider: If Hegseth is out in three months, what happens to all these people? Do they just go hang out at Raytheon? Or can we reel them back into the fold? What is the mechanism here? Probably not, right?

Tony Stark: You’re running into two issues. One, people get tired after two years in politics even normally. The Biden administration was a bit of an aberration in that most people stuck around for four years — and some might argue that was part of the problem. Most people look at changeover anyway with a completely open primary on both sides in 2028. You’re going to see a lot of people scrambling to look for a congressional seat, to look at which campaigns they’re going to work on.

The last two years of the admin are going to be people either trying to secure their legacy — because they know they won’t be welcome in Washington for the next 20 years until we have another round of this — or they are actively going to look for their next job. A lot of these people are going to say, eh, I don’t want to go back in. The people that will stick around or bump around — Driscoll might be one of them. He might become SecDef, he might become National Security Advisor. I actually think that might be a better space given his age. At the ASD and DASD level, most of these people are going to try to hold on because there’s nothing out there for them.

Justin: I can’t imagine either of the next administrations are going to be like, we really need a former officer to lead the Department of Defense. I have to think it’s going to be a non — they’re going to go back to the Gates model, where it’s somebody who understands policy.

Eric Robinson: He was a captain once upon a time.

Tony Stark: Junior officers are fine, which is what Secretary Driscoll is. I don’t think you’re going to see another GO. I’ve seen some other names floated.

Jordan Schneider: DeSantis, baby. Let’s go down the list. We got DeSantis, Joni Ernst —

Tony Stark: That’s how you know Pete’s days are numbered, because Trump only floats those names when he is thinking about making a change. He’s not going to pick the Senate — that’s not passing the Senate. This one or the next one.

Eric Robinson: Why don’t you think Governor DeSantis can get through the Senate? Is he personally unpopular?

Tony Stark: He’s not well-liked. There’s probably a world in which if the House flips — and it probably will — that maybe Chairman Rogers gets the tap, which I think most people in DoD would welcome. But it’s more likely to be a personal ally of somebody else, or somebody who wants to be a caretaker for the next two years. Or you might see what happened in the first admin, which is people change out every six months. That’s a pretty likely possibility too. And you’re going to see a lot of people in performing the duties of.

Jordan Schneider: Tom Cotton — another name. But honestly, would you take that for two years?

Tony Stark: Tom Cotton wants to run for president, and that’s the problem.

Justin: The only thing I see with him is, he thinks his time in the halls of Congress is probably over, but he wants to move up to something else.

Tony Stark: I’d also say that if you’re thinking there still might be a GOP admin in 2029 — sure. You don’t want to taint yourself by saying, well, I already held a cabinet position from ‘27 to ‘28, and it was, you know, the great oopsie of 2027 that I was responsible for. A lot of these people who want to work for the DoD — their politics are not necessarily the people that a Vance would have in his administration, given his proclivities toward isolationism.

Bryan Clark: One thing to think about too: as we move into the last two years of this administration, if they lose the House, if they lose the Senate for sure, the president’s wiggle room in terms of things he can do with his time is going to really be constrained domestically. So he’ll pursue more foreign adventures, which means he needs a SecDef that’s going to be willing to go along with those — which means not somebody who’s going to have an independent base of support or an independent perspective on what the DoD or DoW should be doing. Somebody who’s more compliant in the mold of Hegseth, who depends on the president for his position and future.

So it’s likely we’ll get somebody in there who’s willing to go along with most of what the president wants to do for those last two years, which could be on the model of last time, where it was people who were there every six months — because if you know you could be fired that easily, then you’re always going to either go along or bounce, get bounced, and the next guy comes in. I don’t think the president is going to want to continue the foreign adventurism in the face of domestic resistance.

Tony Stark: When Congress flips, it’s going to get so weird. I think that’s the best description for the last years. It’s going to get really weird.

Jordan Schneider: Well, the question is, how much can a Democratic Congress physically restrain the president from invading new countries?

Tony Stark: Nominees are the one control they have. So if he wants to make the swap on Hegseth and he wants somebody, he has to do it now. He can’t do it when the Senate flips even by two seats.

Jordan Schneider: Or you can have a boring person.

Justin: Conceivably. Heritage probably has somebody that they would pick.

Jordan Schneider: What are the odds on the last two years of Trump laying down to just being like normal, boring conservatives? I guess probably zero. So yeah, we’ll just be in an acting world for the end. Anything else to close out on?

Eric Robinson: They might fleet up like Earl Matthews or somebody like that. That’s not going to be normal.

Jordan Schneider: See you next week on WarTalk.

ChinaTalk songs are now up on Spotify!! Follow ‘ChinaTalk Records’ here or have a listen to ‘Phelan on the Couch When It Happens’ below

At its inception in July 2025, China Fusion Energy Co. (CFEC, 中国聚变能源有限公司) was the biggest nuclear fusion company in the world by registered capital.1 Major state-owned enterprises pledged a total of US$2.1 billion in funding, reflecting a serious commitment on the part of these companies — and by extension, the CCP — to making nuclear fusion happen.

This is one of a series of connected announcements and breakthroughs coming out of China in the nuclear fusion space. In January 2025, China’s “Artificial Sun” set a record for maintaining steady-state plasma for over 17 minutes. In March 2025, Shanghai-based startup Energy Singularity announced that their new high-temperature superconducting magnet (necessary for confining the fusion reaction) had generated a magnetic field of 21.7 teslas, breaking a previous record from the US. In May of 2025, researchers at the Chinese Academy of Science Institute of Plasma Physics published the results of their successful 12-year project to develop a new type of steel to use in the reactor core, which can handle magnetic fields almost twice as strong as the steel to be used in the International Thermonuclear Experimental Reactor (ITER) currently under construction. These announcements certainly create the feeling that nuclear fusion is progressing rapidly in China.

The History of Fusion in China

In December 2022, the National Ignition Facility (NIF) at the US Lawrence-Livermore National Laboratory made history by achieving net energy gain on a nuclear fusion reaction. This was a significant step in demonstrating the feasibility of fusion power generation, and led to significant investment in fusion startups in the US. However, for both the US and China, nuclear fusion research began much earlier.

Fusion was first achieved in bombs, marking the shift from “atomic bombs” that relied on fission, to “thermonuclear bombs” that used fission to drive a fusion reaction, releasing significantly more energy. However, controlled fusion has been much more elusive.

The two main “general approaches” (pdf, page 10) currently employed by most labs and companies are magnetic confinement (MCF) and inertial confinement (ICF). MCF uses magnetic fields to contain continuously burning plasma for long periods, while ICF uses intense lasers and small fuel targets to create short fusion bursts. The ICF approach is what the NIF used to achieve their net-energy breakthrough.

Historically, most fusion experiments have focused on tokamaks, stellarators, and laser-driven inertial confinement. Tokamaks have been particularly significant — the International Thermonuclear Experimental Reactor currently under construction in France utilizes this design, as do China’s main research reactors.

The Soviets were the first to operate a tokamak starting in 1958. The US started research with stellarators in 1953, and didn’t operate a tokamak until 1970, after Soviet scientists made promising breakthroughs, and the tokamak came to be viewed as the most likely route to commercial fusion. China’s first large-scale tokamak, HL-1, started operating in the early 1980s at the Southwestern Institute of Physics (SWIP) in Chengdu. (SWIP is affiliated with CNNC, China’s main nuclear energy SOE).

SWIP now operates the HL-2A(M), a more advanced tokamak, one of the three major domestic tokamaks currently operating in China. The two others are the Experimental Advanced Superconducting Tokamak (EAST — aka the “artificial sun”) and J-TEXT. Operational since 2006, EAST is located at the Institute for Plasma Physics, Chinese Academy of Sciences (ASIPP) in Hefei. ASIPP and SWIP are the two main research institutions driving China’s fusion progress. J-TEXT is affiliated with the Huazhong University of Science and Technology (HUST). A number of other universities and institutes also contribute, albeit in a less substantial way.

Despite all the press that these reactors have generated in China, they are not generally considered to be the most advanced within the industry. The “triple product” is a metric that gives a single value for how close a fusion experiment is to net power, found by multiplying the density of ions in the plasma by their temperature and the energy confinement time in seconds. As the annotated graph below illustrates, China’s current tokamaks fall behind global leaders.

The Master Plan

Route 1: Research Institutions

Chinese research institutions have a clear plan and conservative timeline that will get them to a commercial fusion reactor. The foundation of this plan is the three aforementioned tokamaks. With the results of their experiments, China is contributing to developing ITER, and simultaneously planning the China Fusion Engineering Test Reactor (CFETR). CFETR will serve as a bridge between ITER and a full-scale commercial reactor. According to this 2022 timeline, they will have an operational commercial plant by 2060.

However, they are not in a holding pattern while they wait for ITER to come online (which likely won’t be until 2035 or later). The Chinese government has a host of projects planned or currently underway that will continue to fill in fusion knowledge gaps. The following is an overview of some of the key projects:

• Comprehensive Research Facility for Fusion Technology (CRAFT).

  • A 40-hectare, 20-facility, US$570 million research center intended to solve additional obstacles on the way to building CFETR. It does not include a new large-scale tokamak. Construction started in 2019 and should finish this year. It is located in Hefei, near ASIPP.

• Burning Plasma Experimental Superconducting Tokamak (BEST).

  • BEST is an intermediate-step tokamak between EAST and CFETR, designed to achieve real-world energy production. Construction began in 2023 and is expected to conclude in 2027. It has been described as a copy of one designed by US-based Commonwealth Fusion Systems. It is also located in Hefei. I was unable to find an official cost estimate, and unofficial sources varied. One user on Zhihu (Chinese equivalent to Quora) had the cost at $8.5B RMB, ~US$1.2 billion.

• Shenguang-IV (神光-IV (literally “God Light-IV”) or SG4）

  • China is building a massive, mysterious X-shaped facility in Mianyang 绵阳, Sichuan province. Western news outlets don’t even have a name for it. However, China-focused analysts and Chinese media have identified it as Shenguang-IV (SG4), the fourth iteration of laser facilities operated by the China Academy of Engineering Physics (CAEP). CAEP is also China’s principal weapons design lab, and hence, there has been little said about the facility (the NIF plays a similar role in the US). Analysts estimate that SG4 will be similar to the NIF, but 50% larger. Official budget figures are not available, but as a reference point, the NIF cost the US $3.5 billion to construct.

  • Some Chinese sources state that the energy output of SG4’s lasers will be 2 MJ, which is similar to NIF, which has done experiments with 2.2 MJ bursts. It will have 288 lasers, in contrast to NIF’s 192 lasers. According to Chinese forums, construction for SG4 began in 2017, and one article states that it was supposed to be completed in 2020 or shortly thereafter. However, none of this information could be verified.

• Xinghuo 1 (星火一号)

  • World’s first fusion-fission power plant, with Z-FFR design (Z-Pinch Driven Fusion-Fission Reactor). It has the aim of generating 100 MW of continuous electricity for the national grid by 2030. It is being built in Nanchang, and is expected to cost $2.76 billion. The environmental impact assessment began in March 2025, with initial orders of superconducting material for the plant being made in December of 2024.

• China Fusion Engineering Test Reactor (CFETR)

  • A demonstration power plant (DEMO)-scale fusion reactor expected to enter construction by the late 2020s. It is seen as a bridge between ITER and a commercial plant. Preliminary conceptual design for CFETR was finished in 2015, and the engineering design was completed in 2020.

Results from all these projects will be used to continue refining the design of CFETR, before finally being rolled out into wide-scale energy production a few decades from now.

Route 2: Private Sector

Within the global fusion startup space, there are a host of conventional and unconventional methods being tried to realize fusion much sooner than SWIP and ASIPP’s 2060 timeline. That being said, Chinese companies still have a high degree of alignment with state research institutions. While there are 24 different approaches listed in the Fusion Industry Association’s report, the main Chinese players are sticking to the tokamak and the spherical tokamak, a more compact variant which has lower engineering costs.

Different fusion approaches pursued by 45 global fusion companies, based on reporting by the Fusion Industry Association in 2024. Source (pdf)

These players include NeoFusion (聚变新能), Startorus Fusion (星环聚能), Energy Singularity (能量奇点), and ENN (新奥).

• NeoFusion

  • Founded in 2023, Neo Fusion is a private enterprise backed by the Anhui provincial government. It has over $2 billion dollars in funding, just short of SOE China Fusion Energy Co.’s capital.

• Startorus Fusion

  • Startorus is another state-backed private firm, this time with Shaanxi and Xi’an city as sponsors. Founded in 2021 by Tsinghua University grads, it has $207 million in funding. They are pursuing a conventional tokamak design.

• Energy Singularity

  • Founded in 2021, with $120 million in funding. They are operating HH70, the world’s first successful fully high-temperature superconducting (HTS) spherical tokamak. The company overall is pursuing an approach similar to Commonwealth Fusion Systems. They aim to build the next iteration of their HTS design, HH170, by 2027, targeting a 10-fold energy gain.

• ENN

  • ENN is an established gas company that is also pursuing fusion projects. They have raised $400 million thus far, and also intend to use a spherical tokamak.

China Fusion Energy Co.: The Bridge?

The creation of China Fusion Energy Co. this year is intended to coordinate the various parts of the nuclear fusion endeavor, and help it make the important jump from experiment to commercial reality.

In Chinese media, CFEC is referred to as the “national team.” Although it may look like a cash-strapped investment vehicle, its significance goes beyond that. Wang Zhigang (王志刚), a professor at Tsinghua University’s Institute of Nuclear and New Energy Technology, described its significance this way:

“This is not a simple financial investment, but rather part of the national energy strategy layout. The seven major shareholders cover the entire chain of technology R&D, engineering construction, capital operations, and industrial applications, forming an ecosystem of deep integration among ‘industry, academia, research, application, and finance.’”2

Once one of China’s private companies or research institutes makes the final breakthrough, CFEC will be ready to take the baton and sprint with it.

Race Outlook

So, in this race between the US and China, who is in the lead now, and who is likely to win long term?

It’s hard to determine who has the momentary lead. Especially when insiders and experts seem to disagree. The “Artificial Sun’s” record certainly seems impressive. A report from the MIT Technology Review suggests that China commands in 3/6 of the key industries and technologies that will go into fusion reactors (assuming the conventional tokamak is the eventual victor). After leading annual patent submissions on fusion technology for years, China has now surpassed the U.S.

But others suggest that the Artificial Sun’s records are “unremarkable,” and the real indicator of progress is net positive reactions, which China has yet to achieve in the nearly three years since the US first crossed that milestone (and crossed seven more times since). IAEA’s annual Nuclear Fusion Award, given to the most impactful paper published in the Nuclear Fusion journal, has never been given to a Chinese scientist.

Most seem to think that the U.S. and China are roughly tied at the moment. The US is leading China in investment, but only slightly, and the nature of the investment varies substantially. Nimble private funding is dominant in the US, which lacks the kind of national modern fusion facilities that China has, while China’s investment is almost entirely public. Annual public funding between China and the US is roughly 2:1, US$1.5 billion to US$800 million. Which investment model is more effective remains to be seen.

Fusion in the “Engineering State”

Breakneck by Dan Wang has sparked a great deal of discussion and scholarly disagreement about what has made China a building and manufacturing powerhouse, and what holds the US back. Whether it is an “engineering state” vs. a “lawyerly society” (Dan Wang’s theory), a “Leninist developmental state with Chinese characteristics” vs. “lawyerly society” (Jonathan Sine), or “developmental state” vs. “regulatory state” (JS Tan), the fact remains that, at least at the present, China is much better at building stuff.

EVs, solar panels, and high speed rail are often held as examples of America (or Japan, in the case of HSR) winning “0-1” innovation and China winning “1-2” innovation. While it isn’t as widely discussed, another striking example is conventional nuclear fission energy. The US was the world leader in fission technologies, and has the largest fleet of nuclear fission reactors in the world. But China has been on a prolific building spree, and analysts now say that “China likely stands 10 to 15 years ahead of the United States in its ability to deploy fourth-generation nuclear reactors at scale.”

In the case of nuclear fission, perhaps the most succinct explanation of this was offered by Kenneth Luongo, who said that China doesn’t “have any secret sauce other than state financing, state supported supply chain, and a state commitment to build the technology.” More broadly, another author described how private companies in state-supported industries gain access to the “standard triple package of cheap financing, cheap land, and cheap regulatory cost.” Fusion will have all these benefits.

China also has significant “process knowledge” for large infrastructure projects (with their nuclear reactor building spree particularly relevant). They are also developing a deep bench of scientists who will be able to work on fusion projects. Experts estimate that China has thousands of PhD students in fusion, compared with hundreds in the US. Even if the US makes the breakthrough first, China is likely to imitate quickly and roll it out much faster than the US can, gaining additional insights along the way to then pull ahead.

The US Moonshot

Simultaneously comforting and concerning is the knowledge that this isn’t news to US officials and lawmakers, and… little is being done. A Feb 2025 congressional commission report called for a one-time, $10 billion investment to build critical research infrastructure. They argue, and I agree, that “American ingenuity has proven time and again that, particularly when catalyzed by a long-term strategy and public-private partnerships, it can solve seemingly insurmountable problems.” But whether or not the US can really unite the full force of public and private efforts behind anything in these polarized times remains to be seen.

Are AI chip exports a red herring?

The Trump administration has agency over the two variables most relevant to whether China will have enough compute to really compete with the US: how many chips they can make and how many chips they can buy. But for all the drama we’ve had this administration around whether Trump will allow Jensen to sell chips to China, we’ve had basically zero movement on the tooling side. Without access to foreign tools the US could control, Chinese chip and memory makers would not be in a position to even produce the meager amounts they can today.

This administration teased controls on sub-systems in Trump’s July 2025 AI Action Plan, but absent the headfake around the Affiliates Rule that was wound down after Beijing escalated on rare earths, we’ve had zero movement to close loopholes.

Congress is looking to take the matter into its own hands. The MATCH Act would make controls country-wide as opposed to entity-specific, address servicing of already installed equipment, and squeeze allies to comply by putting a timer on the application of the Foreign Direct Product Rule.

Which layers in Jensen’s layer cake benefit from SME and AI chip exports to China?

Only Nvidia and AMD win from AI chip exports, and nobody in the western ecosystem wins from SME exports except China and SME companies. The table walks through the stack:

The table above illustrates commercial interests only. Widen the frame to include national security, where the Chinese fab industry that can’t exist without US tools is the single most important lever we have, and the case for holding the line on SME gets stronger still.

Beyond fear of retaliation from Jensen when it comes to chip allocation (which Jensen promised didn’t exist on the Dwarkesh podcast!) I am surprised that more of the industry hasn’t been more vocal in their support of Congressional limits to how much Trump can loosen chip controls. Congress acting also makes it less likely for SME escalation to trigger a tit-for-tat on rare earths, as legislation can tie president to the mast and give him the ability to tell Xi, “sorry but there’s nothing I can do on this.”

Real ones ask real questions

Since the October 2022 export controls began, Jensen Huang has been on over fifteen podcasts of over an hour or more. Almost all of those didn’t really raise chips or China.

Most unforgivable was John Hamre of CSIS, a former Deputy Secretary of Defense who runs what is ostensibly a serious national security think tank. He did not do any real homework or ask a direct export controls question, and instead took the time to joke about how dumb he is.

Dr. Hamre: I went in for an MRI recently and my wife said, make sure you take a picture; I don’t think there’s a brain up there, but I’d like to see it – (laughter) – to prove there’s something.

Mr. Huang: And what did you find out? (Laughter.)

Dr. Hamre: We were – there was – there was nothing, I mean. (Laughter.) She was right and I was wrong. (Laughter.)

It took Dwarkesh, who at 25 has not yet served as DepSecDef, to ask America’s most prominent CEO about his most controversial national security policy.1

How did this interview even happen in the first place? My guess is that Dwarkesh cold-emailed Jensen, who said yes, leaving his PR team to watch through their fingers as he gave his first interview to someone who really did their homework and had the guts to bang for twelve rounds. Kudos to Jensen for taking the interview, and after Dwarkesh gave him an off-ramp to say, “You don’t have to move on! I’m enjoying it!” Kobe energy.2

National Security as Jensen’s Willful Blind Spot

Jensen spent decades building a company with zero dual-use implications and practically no reason to interact with Washington. He relied on the world’s most international supply chain which would not exist without the peace that East Asia has been blessed with the past fifty years thanks to unquestioned American military preeminence. While selling chips to gamers and bitcoin miners, he had a lodestar of one day unlocking scientific advancements.3 And now he’s doing that, while also rapidly upgrading the technologies that provide national security without having truly grappled with their implications.

Watching a national security community get in the way of that vision of global empowerment must be infuriating. But wishing away the reality of AI’s dual use implications on cyber by saying that “the way to solve that problem is to have dialogues with the researchers and dialogues with China, and dialogues with all the countries to make sure that people don’t use technology in that way” is willfully naïve. Obama tried to negotiate some cyber boundaries with Xi at Sunnylands, and that truce lasted maybe three months. In recent years, Chinese hackers have been caught inside US power grids, water utilities, ports and pipelines. Dwarkesh is correct in saying that “If you had a cyber hacker, it’s much more dangerous if they have a million of them versus a thousand of them. So that inference compute really matters a lot.”

Jensen’s response to Dwarkesh’s repeated pressing on PLA cyber use ran as follows: “They have plenty of compute already. The amount of threshold they need for the concern you’re worried about, they’ve already reached that threshold and beyond.”4 But Jevons’ paradox applies for the military industrial complex too: demand for compute is skyrocketing across industries because more of it means more productivity.

Jensen also waves off the idea that compute constraints meaningfully slow Chinese labs, but algorithmic innovation itself requires compute. There is nothing special about military organizations or other dual-use technology where past a certain point more compute isn’t useful. And if we’re, as Jensen argues, five years away from “understanding the biological machine,” we’re also five years from some mind-blowing new weaponry.

Is cyber a shiny object?

For all the excitement over the past few weeks around Claude Mythos, there’s a real limit to just how pointy cyber can be. Claude Mythos 3.0 won’t be able to keep the Strait of Hormuz open.

New military technologies and doctrinal innovations are most impactful when first introduced and as adversaries adapt to them over time. As we discussed in last week’s WarTalk, the initial shock of an AI cyber capability like Mythos is real, but the playbook for a response is straightforward: air-gapped and local mesh networks, partitioned internets, and hardwired secure comms. The half-life of a first-mover edge, particularly in software, is short.

None of which is to say AI doesn't matter for warfighting. Beyond cyber, we’ve already seen dramatic impact of AI around targeting and logistics that allowed the US to conduct an unprecedented air campaign over Iran. We’ll soon see similar leaps around command and control. But as Ukraine has reminded the world, you still need lots of stuff that goes boom to feed “the greatest of consumers.”5

We will remain in an era of mass precision, where you still very much do need mass, for a long time to come. Until AI has robots building robot armies, the US will still need to do the foundational work of scaling up its defense industrial base to produce enough attritable mass to deter high end conflict. We should not expect AI capabilities on the next few years to get the Pentagon and Congress off the hot seat to reform and build.

Jensen’s Inner Fire and Lawyers vs Engineers

I claude coded a website that diagrams out their arguments, with different modes including LD-style high school debate and a rap battle (Dwarkesh as Kendrick, Jensen as Jay Z). Dwarkesh would have won on substance and speaker points, with Jensen’s biggest truth-stretching coming around talking about Chinese fab capacity (see this podcast I did with Chris McGuire on why Huawei can’t catch Nvidia). Even though Jensen was playing with the handicap of making his case while dancing around investors, China, and Trump, you should still take him seriously.

’s take (endorsed by Jensen on his Lex interview) is that American society biases too much in favor of lawyerly Ivy League polish and against China’s engineering bias.6 Setting aside the shade I’ve thrown at Jensen for his export control policy, Nvidia is an American company, Jensen has lived the American dream while swimming culturally upstream for decades, and the U.S. is much better off for it.

I want to close with an extended excerpt from twitter account teortaxes:

Jensen is the gangsta poster boy for American Dream. He is REALLY is Not a Loser. He’s also not a Car, but indeed is the driver. Moreover, there are almost no people alive with a greater dynamic range of lived experience, who have gone from positions many would die to escape and into a position entire institutions fight to death over, and only tightened their grip since. Xi Jinping would qualify as a peer, maybe? (Musk has less range, even though he ended up in a similar place.) These individuals are fascinating outliers, and I believe that when they deign to explain their ways, however awkwardly, us mortals should sit our asses down, listen and learn.

Jensen has basically ascended from a toilet-scrubbing immigrant runt to a demigod, from a random NPC to a Singularity Kingmaker, a whole vertebra of the Universe’s backbone; and that journey informs his views, just like Dwarkesh’s “be really good at Reasonably Conversing, insure your middle class stake” informs his. Jensen’s journey is not about luck, he is definitely not “1 SD IQ lower”. He hasn’t trained himself in our exact mode of coffee salon intelligence that allows for casually cooking up consistent, defensible, lawyerly arguments about, basically, the structure of written information. So he’s worse than us at it. Not because his epistemology is inferior, as in «less predictive»; it is just different, and insistence on Not Being a Loser is its functional part. He is supremely motivated to Not Lose, so he’ll not make self-defeating moves. How he sorts moves into self-strengthening and self-defeating is, therefore, very important, more than verbally persuasive arguments.

Our conversation covers…

• What Quantum Computing Actually Is — A primer on qubits, superposition, and why quantum computers aren’t “faster classical machines” but fundamentally different systems designed to simulate nature and solve specific classes of problems.

• Why Quantum Matters Now — Breakthroughs in error correction and hardware have shifted quantum from theory to an engineering race, with major implications for drug discovery, materials science, artificial intelligence, and cybersecurity.

• The Economic and National Security Stakes — Quantum’s potential impact on cryptography, advanced manufacturing, biotech, and defense makes it a strategic technology with an extremely small margin for error in global competition.

• From Science Project to Industrial Policy Challenge — The bottleneck is no longer just physics but scaling. Talent pipelines, fabrication capacity, supply chains, and the kinds of public-private partnerships needed to move from lab prototypes to deployable systems.

• What Winning Looks Like — Leadership isn’t just building the first powerful machine. It’s shaping standards, securing supply chains, protecting encryption, diffusing capabilities across industry, and sustaining innovation in a tight U.S.–China technological race.

Plus, the encryption stakes, the engineering bottlenecks, the race with China — and a reading list and job resources for those interested in the field.

Thanks to the Hudson Institute for sponsoring this episode.

Why Quantum Matters Now

Jordan Schneider: All right, Quantum 101. Why should ChinaTalk listeners turn their attention to this topic?

Zachary Yerushalmi: Listeners should care about quantum because, with AI on our doorstep, quantum represents the single biggest lever we have to pull as a society for the next couple of decades.

For the ChinaTalk audience specifically, this isn’t just a big economic and national security opportunity. For such a policy-oriented group, the margin of error is incredibly thin. We have more at stake here than any industrial program since the atomic bomb. It’s multi-layered — maybe that’s quantum for you. But that’s why I think folks should care.

Jordan Schneider: What do you mean by margin of error?

Zachary Yerushalmi: You’ve all spent a lot of time thinking about semiconductors. My observation is that in semiconductors, we have literally decades of moat over China and other competitors that we care about. That doesn’t mean we sleep on semiconductors or forget industrial policy there.

But quantum is fundamentally new for many of the capabilities we’re trying to bring forward. By definition, that means our moat is pretty small. Whereas in semiconductors we can get some things right and some things wrong, in quantum, our margin of error is freakishly small. We have to get it right from the beginning.

Jordan Schneider: All right. Make the case for why it matters. Why is this the thing that could be the next door for humanity to open over the next half-century?

Zachary Yerushalmi: It’s the single biggest lever we have left now that AI is on the doorstep and/or here. Why it’s so important goes back to the inception of the idea of a quantum computer. It came in 1981 from this guy Richard Feynman. He has this famous quote of effectively “nature’s quantum, dammit.”

It’s the realization that if we want to solve the problems that quantum mechanics governs, which are really the world of the atomic realm — not Marvel, but the very small, the very cold — this includes drug discovery, catalysts, material science. This is all the things that govern the building blocks of the universe. We can’t use our classical approach and classical computers to solve that.

This gets a little bit into the math and could get dangerous, but the quantum world behaves in ways where even very small systems explode in complexity. A standard two-particle system — these systems explode at 2 to the n. If you have a two-atom system, that’s 2 to the n states that you need to understand. If you add a third atom to that, it’s 2 to the third. You suddenly need to understand 8 states, not 1 additional. By the time you get to 20-atom systems, you need to understand a million states. A 20-atom system is freakishly small.

This gets back to why this matters. Let’s use the specific example of penicillin. Penicillin is 42 atoms. That is not a big molecular system, but penicillin is obviously pretty important. If we want to understand penicillin, much less where it falls short, then in order to do that classically using our current computing paradigm, even with the world’s best AI, we’d have to use something like 10^86 transistors to do it.

Just to double-click on that — I said 10^86 quickly — it’s a shockingly large number. 10^86 would need more transistors than there are atoms in the observable universe. In simple terms, we could literally use the energy of the entire universe, and we couldn’t quite make a basic physics-based model for penicillin.

If we ever talk about living in the Jetsons age, rationally designing all these things, our current paradigm, even with the world’s best AI, is just never going to get there.

This gets back to that quote from Feynman in ’81. He just went on this rant, and it was really a thought experiment. He said nature is quantum mechanical, damn it. What if we’re just approaching this on the wrong terms? What if we built a computer that operated on the same principles as penicillin operates itself? It’d probably be more efficient.

It wasn’t just a little more efficient. It was a reinvention of what a computer could be. Instead of needing more transistors than there are atoms in the observable universe, you need something like 186 of these quantum bits, or qubits. We’re not there yet, but we’re more or less on the cusp of it.

If you can do that — again, because these systems are exponential in nature — when you go from 186 qubits to understand penicillin to 187, just one additional system, it’s not a little bit better computer. It is thinking about penicillin interacting with its neighbor. When you get to 1,000-qubit systems, you’re talking about rational design of much more complex systems.

If we ever want to get to what folks talk about from the AI world of curing cancer, solving climate change, addressing the material science of the world all around us, actually, really the only cowbell that we can really hit on for some of those problems is quantum.

Chris Miller: Could I ask maybe the same question from an economics perspective, which is thinking about the market for quantum computing capabilities? It’s an unfair question to ask because if you’d asked people at OpenAI in 2019 what the market for AI was, they would have given you a very large number without much justification — because who knew exactly how it would play out. But how do you think about where we’d like to be deploying quantum computing capabilities in 2035? Drug discovery is obviously the first answer everyone gives, and that’s obviously a potentially huge market. But beyond that, how do we think about the economic impact of quantum computing?

Zachary Yerushalmi: There are two applications that folks go to. The analogy I think about is — Chris, you cover this so well in your book — one of the really first killer apps of the semiconductor era was the hearing aid or transistor. All of this is like forecasting the impact of this technology with a transistor set of examples for what is possible.

But the interesting thing about quantum is that the transistor capabilities are worth hundreds of billions to trillions of dollars.

A couple of applications — the first would actually maybe not even be life sciences. It would be something around corrosion and accurate corrosion modeling, which is worth tens of billions of dollars to the global shipping and national security sector. That’s because it’s a simpler problem and is tractable to get around. There are some interesting things around nuclear chemistry. As these systems get bigger, you start to look at drug discovery and material science.

Folks talk a lot about room-temperature superconductors, which is a whole other YouTube wormhole to get into. But if we want cell phone batteries that never lose power, the ability to rationally design these systems at the molecular level, at the atomic level, opens up that possibility.

The second class of problems from an economic impact perspective is, candidly, in many ways, maybe more near-term and more scary. It gets back to the transistor analogy. Most of the classical algorithms we developed occurred after we had the computer. The only other big class of problems that folks know quantum computers are useful for, aside from the molecular modeling piece or the physics modeling piece, is that they’re really good at the hidden subgroup problem, which sounds jargony because it is jargony.

The big one there is factoring large primes. For anybody aware of cybersecurity concerns, if you can factor large primes, you basically crack the code that underpins all of our existing cybersecurity infrastructure. This is why global governments, even putting aside the Jetsons nature of what we can unlock, are really worried — because all of their codes and all of our financial infrastructure and things like Bitcoin are underpinned by this problem that a classical computer literally takes the age of the universe to solve, and a quantum computer looks at, laughs at, and steals your Bitcoin wallet.

Understanding Quantum

Jordan Schneider: Before we go too deep into applications, Zach, what’s your favorite analogy to give folks to start wrapping their head around?

Zachary Yerushalmi: Two analogies that come to mind. The analogy I often think about is this — if classical computers are like a car, quantum computers are like a rocket ship. We’ll still use classical computers, just as we still use cars for certain applications. But for certain problems, a faster car isn’t going to get you to space more efficiently. You need to completely rethink your mode of transportation.

With quantum computing, because of the nature of the problem set, we need to invent the equivalent of a spaceship. The idea is to create a computer that operates on the same principles as the systems we want to solve.

How do these computers actually work? The best analogy I’ve seen is the maze analogy by Matt Langione, who’s the quantum partner at BCG.

Picture a maze in your mind. The way humans approach a maze is actually quite similar to how classical computers do it. You walk in, face a decision to go left or right, choose right, hit a wall, and then backtrack. The time it takes to solve the maze is the cumulative time of making each decision and working through the maze sequentially.

A quantum computer approaches a maze fundamentally differently. When a quantum computer enters the maze and faces that first decision to go left or right, it leverages principles of superposition, entanglement, and interference to say “yes” to both paths. It explores both simultaneously. Not just at that single junction — it examines every single junction in the maze and explores all possible paths in parallel.

While a classical computer takes time to make each decision sequentially, accumulating time with each choice, a quantum computer evaluates the first junction and every other junction simultaneously.

The real-world applications that resemble this maze structure are found in the molecular world, worth hundreds of billions to trillions of dollars. Whether in chemistry or molecular science design, these fields share many similar characteristics with the maze problem I just described.

Jordan Schneider: I want to make a pitch based on the background reading syllabus that Zach sent Chris and me, which we’ll put in the show notes. There’s a YouTube video by 3Blue1Brown called “But what is quantum computing?” which I confess I had to watch probably three and a half times before it started to settle in.

One of the interesting things about this topic is how quickly you try to make analogies of left/right or two-dimensional space, and then they give you three-dimensional space. But all the exemplars are actually 16 dimensions, 20 dimensions, and when they show the equations, it actually makes more sense than when they try to give you the analogies.

Whenever these folks try to make it simpler by giving you some spatial analogy for what’s happening, I found myself going back to the parts of the YouTube video and the Wikipedia pages that just had equations in them. You don’t have to expand your mind like you’re some enlightened Tibetan Buddha or something — you can just take it for granted that these equations are what all of the particles are doing or not doing.

It was fun to stretch my mind in a way I haven’t in a while. I contrast that with earlier today when I was listening to some World War II fighter history where the physics was straightforward — the plane’s going this fast, the other plane’s going this fast. They invented a little computer that did some straightforward calculations to shoot 50 yards ahead so that it would hit your Messerschmitt or what have you. And here I am with Zach going down the deepest rabbit holes of the universe.

But Zach, make the case for people spending that long weekend actually trying to wrap their heads around some of the physics fundamentals of this, as opposed to just jumping to thinking about all the cool applications.

Zachary Yerushalmi: First, it’s just cool. If you’re into listening to Neil deGrasse Tyson and StarTalk, why not make a little bit of time for quantum?

But the second, I think, is separating hype from reality. If the case for quantum from an economic and national security perspective is that important, having a base intuitive understanding of what a quantum computer does, what it doesn’t do, and where it’s useful is essential for ChinaTalk listeners because a lot of them are engaged on policy — prioritizing where we make investments to actually have that lead.

This is true of any discipline — there can be a lot of smoke and mirrors and hype and reality. But in quantum, almost because it is such a hard thing to grapple with, I find more of that. So I think that investment is super worthwhile.

Chris, I welcome your take. Where was that sea change? What was the ROI calculation there for yourself?

Chris Miller: To me, every next step in computing capabilities seems magical or impossible 15 years out, then it becomes possible and normal, and then we forget that it’s happening. My analogy for where we are today vis-à-vis quantum is that it’s like 2015 in AI. All the researchers were saying progress is coming very rapidly, and everyone outside of AI said, “I don’t know what this means, and it’s probably not real. Even if it’s real, it’s a long way away, so I’m going to ignore it.”

Then the world was surprised in 2022 when ChatGPT dropped in a big way. It seems to me that a roughly comparable time horizon is where most quantum researchers think we’re going to be — in half a decade or so.

Zachary Yerushalmi: Taking the AI analogy, if you got attuned to AI as a government, as an investor, or as a policymaker when ChatGPT hit, it was too late. The time to really be attuned to it was probably around 2017 — “Attention Is All You Need,” the birth of the LLM.

What was wild about quantum last year in 2025, is that the ChatGPT moment isn’t there. It was not there last year. What I’d argue, though, is it was the birth of the LLM for the industry. That’s because with the Google Willow paper that came out and a couple of other breakthroughs from others, it went from a technology domain where, as I alluded to before, for these things to be useful, you have to add a certain number of these qubits.

It turned out that up until last year, every time you added a qubit, the entire system got less stable, which is probably bad news from a “these things are going to be useful” perspective. Where that sea change happened — this came out with Google’s paper — was when they found out a way through error correction, where you added a quantum bit, and the entire system got more stable.

In my head, that’s a shift. With that base architecture, it becomes more of a “when do these capabilities come online in a way that changes the world around us” instead of an “if.” It just reinforces why it would be attuned now, because we can’t afford to wait until ChatGPT hits.

Jordan Schneider: That’s another argument for actually spending the time to understand the fundamentals here. Reading the results of all those AI papers, even for someone who isn’t a computer science PhD, has been relatively straightforward. Certain benchmarks are legible to human beings, like labeling images, or you can talk to the models and feel how good they are.

That is something that even a layperson has been able to follow without really understanding the transformer architecture or what have you. In order to separate hype from reality, when Microsoft or IBM or Google comes out with a paper, it requires being able to digest more secondary technical commentary than just, “Oh yeah, try the model out for yourselves.”

It’s not like we all have quantum computers in our backyard that we’re trying to model penicillin with, and all of a sudden it’ll work. But maybe we’ll get there one day.

How Policy Should Change

Jordan Schneider: Let’s stay on the industry history piece, Zach. If 2025 is our turning point, where were we before? Where was 2014 to 2024?

Zachary Yerushalmi: Quantum computers started as a thought experiment in 1981. The entire industry was born in that moment when Feynman said, “Let’s build something on nature’s own terms.” The first 2-qubit gate operation was an important breakthrough that actually happened in Colorado as a kind of unitized thing for how you would create a quantum processing unit.

We’ve had not just steady but exponential progress in the capability of these systems since that time. The quantum supremacy paper came through with Google in 2019. Then last year, we had a sea change where the focus shifted from the more fundamental side of R&D to engineering these systems to a sufficient scale.

These problems that keep the NSA up at night and the rest of us dreaming from a new capability perspective are really on the cusp. If we look at industry timelines, even a couple of years ago, folks would have said that a useful quantum computer is about 10 years out. Last year, it went from 10+ years out to 3 to 5 years. With these breakthroughs, folks think we’ll get systems capable of cryptographic capabilities that we’re candidly worried about, or material science capabilities that open up new economic opportunities.

Chris Miller: It’s happening fast. I agree that we’ve shifted from the realm of science experiment to the realm of engineering. The question that brings up is: how should policy change?

For fundamental research, you support academics, and they do their studies and push the frontier of knowledge in physics and other fields. But for engineering, you need different tools for different problems. Scaling up is as much an economic problem as it is an engineering problem. Can you walk us through how you think the types of policies that we should think about in the quantum sphere ought to be changing, given the shift from solving the science, which we’ve made a lot of progress on, to addressing the engineering, which is where we are right now?

Zachary Yerushalmi: First, if we use computing history as a lesson, you can’t just put down the fundamental R&D. If we went from the computing age to vacuum tubes and put our tools down and said “job’s done,” we would have missed out on the transistor. That engine of growth and innovation — thinking about what the next generation is and maybe even a reinvention of what people are thinking from a basic architecture — is essential.

Second, how do we look at the industry now that there’s this sea change? The mental model I always use for quantum is biotech. In biotech, you have drugs — maybe small molecules, or you have CAR-T or these different modalities. You have that for quantum itself. Those are the things that will cure cancer. You also have the tools for addressing that.

As we cross into this chasm right now, what I would think about from a policy lens is that we’ve entered phase 1 clinical trials. What we need to do from a policymaker perspective actually looks like a similar toolset to how we foster the right environment for biotech.

The important distinction though, is that in biotech, we have a huge moat. The cluster, both commercial and scientific, around Boston is so globally dominant that we can screw up a bunch of stuff. Now it’s about what tools we need, given the same architecture — material technical risk, need for commercial payout, and so on. But now we need to do that with a much earlier stage industry.

Jordan Schneider: Let’s do a little bit of industry analysis, because there are some very shiny, polished releases from some of America’s largest publicly listed companies. You’ve got startups doing things, government labs, and academia. Let’s start with the giant companies. Is this just like everyone wants to be Bell Labs? Why are Microsoft and Google spending time on this sort of thing? They’ve got data centers to build with their CapEx, right?

Zachary Yerushalmi: There are weird R&D tax incentives in the state of California that we won’t get into. Why do they have to focus on this? It gets back to the fact that they can’t afford strategic surprise, right?

Your comment about this being such an important thing — them spending a couple of billion dollars on it a year to make sure they have their finger on the pulse so there’s not a Sputnik moment for them as a company — is just worth the ROI. Ultimately, they have big quantum programs, arguably some of the largest, but relative to everything else they’re doing, it’s rounding errors on their balance sheet.

I’d argue the same about the US government. DARPA, indicative of how important this is, has made quantum its largest public program in the agency’s history. It’s called the Quantum Benchmarking Initiative. This almost gets to your question, Chris, about now that we’re in this new era, what are the policy levers that we need to undertake to get this right?

That’s canonical. It all comes back to what are the right commercial structures — because commercial structuring is what makes biotech motor on — and then what are the right technical levers that you need?

Chris Miller: Do you want to explain what DARPA’s Quantum Benchmarking Initiative actually is? Because it actually seems to me exactly like the clinical trial analogy that you just mentioned. Can you dig into that?

Zachary Yerushalmi: The Quantum Benchmarking Initiative, as I alluded to, is the largest public program in the agency’s history. This year alone, they’re looking to spend $600 million on it.

The frame there is — they don’t say this explicitly — but the US government can’t afford to be surprised that China can break its codes. They stood up this program that is effectively about learning about the capabilities of the leading players in the world, sans China, because China’s not going to talk to the US government about its quantum capabilities.

The model that it took to do that — and huge respect for the team there — actually reconstructed an advanced market commitment like what you see in drug discovery, but applied that to quantum. It’s effectively like a grand challenge prize structure.

The first phase, you get a little bit of money — $1 million here or there. It’s really just a table stakes thing. The next phase is $15 to $20 million if you pass a certain scientific threshold. But things really get going by the third phase of the program where you can get paid $300 million if your system is deemed credible to that stage and you build a demonstration-scale capability for it.

The implied thing is after that — once you get through that third phase of DARPA — it’s kind of your FDA approval. Some government agency is going to buy one of these things because it’s good. It’s going to investigate parts of the cryptographic world that they would be very interested in.

If we get back to that frame, those payouts map really well to the different sorts of commercial staging you’d see in biotech. The $300 million prize is like some Phase 1 or 2 stage, and then a billion-dollar prize if you get past that. That’s super important, not just because these companies want to earn money, but because if you want a virtuous cycle going, you need to get the venture investors and the private markets excited. Those payouts are big enough to create the incentive.

Jordan Schneider: Let’s go to startups. Why do they exist? Are they real yet? What are they doing when it comes to funding?

Zachary Yerushalmi: Startups exist here just like startups exist in biotech, right? Big companies — IBM is putting billions of dollars into this thing. IBM has the flagship program in quantum. But big companies face limitations. They can do a lot of innovating and have important programs and distribution. But just like Novo Nordisk doesn’t do all of the innovation in the world around discovering and developing drugs, you also need early-stage players disproportionately coming from academia that are bringing new paradigms that could disrupt what’s happening in this technology space.

The innovation in this market is like what you see in biotech. You get the big tech players driving programs with amazing access to supercomputers, but some newer-stage programs and approaches could be incredibly disruptive. Those are typically pioneered by startups. If those really take off, then typically the big tech players swoop in and either make an investment or buy them out. We’ve seen that in quantum.

There are a couple of approaches in quantum computing — call them modalities. The historic one is called superconducting qubits. This is the solid-state approach for which John Martinis won the Nobel Prize recently. That’s gotten 60 to 80% of the investment to date for the industry.

But there are newer approaches. The most prominent are probably neutral atoms. Instead of using an almost synthetic quantum bit, they use atoms themselves as the qubit. This was considered science fiction literally three, four, or five years ago. People thought this idea was a total joke.

A bunch of startups, as you see everywhere else in the world, grabbed the mantle and said, “No, I did my postdoc on this. This is not a total joke,” and ran with it. There was a sea change probably two or three years ago, and it’s now one of the leading approaches. Nobody would have bet on that so short a time back. Now it’s one of the things that could really have a shot.

A marker of this is Google. They pioneered and continue to push forward with superconducting as their core approach. But they’re so worried about and interested in neutral atoms, they just gave QuEra $250 million because they think that approach could work as well.

That’s the role that startups can play. It’s the classic disruptive innovation — driving forward what folks thought couldn’t be possible because the risk-reward wouldn’t make sense for an existing company.

What Does “Winning” Mean?

Jordan Schneider: Now it’s time for a little Elevate detour. Why don’t you tell the folks out there what you do all day, Zach?

Zachary Yerushalmi: Elevate is the US government’s quantum tech hub. We’re the first and only major place-based investment that the US government has made in the quantum industry. They did that ultimately because the Mountain West cluster is the largest quantum cluster on the planet. It represents almost half of the US quantum jobs and half of the deployed capital. It’s massive by quantum standards — small industry, but still massive by quantum standards.

What I do, and this relates to the policy measures I’d be keen on discussing, is work toward our mission to dramatically accelerate the commercialization of quantum. We do that as a public-private partnership. Our work focuses on specific, typically technical bottlenecks that are market failures that other players aren’t well-placed to solve.

We look at things like fabs, packaging, and certain shared-use equipment that, for various reasons, national labs and universities aren’t well-placed to address. Startups might not have the capital, expertise, or time to solve these issues either. That’s ultimately what Elevate addresses. We have 140 to 150 members in our consortium, including all the usual suspects — national labs and universities in the Mountain West — but also every big tech player with a quantum program. That’s what we dive in to solve with lots of partners.

Chris Miller: Zach, we’re in this quantum race, and China is a major competitor. What does winning actually look like?

Zachary Yerushalmi: In my view, it’s getting there first and maintaining the best capability as a nation long into the future. Getting there first means building the first — folks will throw out the word “fault-tolerant,” but think of it as a commercially useful system. Something that drives commercial value, whether through cryptographic use cases, material science use cases, or other applications. This is like building the first useful computer with vacuum tubes.

The second part is continuing to have the best capability in the world, or really, access to that capability. That’s what winning means to me. The tricky part is identifying the lead indicator. We can’t look at the price for that. The challenge I’m trying to figure out is — in the absence of price, what’s the lead indicator for us winning?

Chris Miller: When we get to our first commercially useful quantum computer at scale.

Do you think there will be one company that dominates the market like NVIDIA, or should we expect multiple different paradigms to be relevant, perhaps for different applications where they’re better or worse suited?

Zachary Yerushalmi: People smarter than I am suspect that, at least for the foreseeable future, we’re going to have pretty purpose-built machines. This comes from the previous computing era, where you had purpose-built machines based on the application. My instinct is that’s what this is going to look like.

That may at some point converge on a transistor-like architecture for quantum — something that everybody converges on and uses. But we’re not there yet. As a system, and this relates to what I was talking about earlier with the car versus rocket ship analogy, most experts suspect quantum will play a specific role in computing.

Chris, you talk about this as the three-paradigm model for computing. You have classical CPUs — those will stay useful. You have GPUs for AI-accelerated compute — those will stay useful for a certain set of problems. But then you’re also going to have quantum processing units. These will work in tandem to solve some of the biggest problems we care about from a science and cryptography perspective.

Chris Miller: That’s something that a lot of people who are new to the field don’t understand. There’s a common assumption that quantum will replace classical, which is obviously not the right way to look at it at all.

Zachary Yerushalmi: Just like GPUs didn’t replace CPUs, these are Turing-complete machines. They technically can do all the computation; they just won’t be efficient at many of the problems you’d want to computationally solve.

Jordan Schneider: One of the remarkable things about AI is how quickly the learnings at the frontier diffuse to firms trying to catch up. We’re recording this on February 23rd, and we just had an interesting story come out today. Anthropic reported that DeepSeek, Dripu, and Minimax were all making millions of queries to try to get data they could then feed into their models.

There’s this whole narrative about how if you go to enough parties in San Francisco, you’ll hear about the cool new training techniques that you can bring back to your own lab. To what extent do you see frontier breakthroughs leaking out to other firms trying to do the same thing? And spilling across borders as well?

Zachary Yerushalmi: That aspect of quantum is still driven by academic researchers in a big way, so publication remains important in quantum. Just like you see publications thrown on arXiv and then diffuse, that very much happens in this industry.

There’s an interesting caveat, though — and this is mainly received wisdom — that the Chinese government actually keeps publications on lockdown. They typically wait for a breakthrough from one of the firms in the West, and then they’ll allow their researchers to publish something similar. This isn’t the kind of hackneyed stereotype about Chinese innovation that people sometimes deploy. That’s not the right mental model here.

The big difference with AI is that quantum is very much a hardware sport. This means iteration times are much longer. A lot of that diffusion is received wisdom and deep knowledge about how to fix optics to a breadboard and how these systems behave in different ways. It’s a very different science from that domain of zeros and ones.

Jordan Schneider: Which presumably would make the learning more frictionful across firms and involve a lot more people.

Chris Miller: Maybe the analogy is that in AI, the algorithms diffuse rapidly, but the know-how about producing the chips hasn’t. Perhaps the analogy is the same — that the algorithm layer, the software layer, and the research layer might diffuse rapidly, but the manufacturing know-how doesn’t.

Zachary Yerushalmi: Totally. And it’s at a much earlier stage. Each of these paradigms doesn’t have something like the transistor that you can base all your understanding on. Each way of building a quantum computer has deep expertise built around it.

But again, everything is double-edged. Because of the earlier-stage nature of the field, if there’s a real breakthrough in China around a particular domain, it’s going to be much harder to transmit that knowledge over to the US. That moat is just much stickier.

The Encryption Cliff

Chris Miller: One of the obvious uses of quantum computing that we’ve known about for a long time is breaking encryption. Now we’ve got post-quantum encryption standards that have been released by NIST, although it’s unclear how widely or rapidly they’re actually being deployed — probably not rapidly enough. Walk us through how you see us reaching a point in which all of our 2010-era encryption is easily broken by a quantum computer.

Zachary Yerushalmi: It’s pretty scary because NIST recommends all government systems be upgraded by 2028 or 2029, and consumer systems by 2035. That recommendation came out early last year. By the end of last year, folks were talking about having these systems online that can break these standards in 3 to 5 years from then — so by 2030.

That freaks me out because typically you want 10 to 15 years as an upgrade cycle for traditional security protocols, and we have 3 to 5.

Why these systems can do that is through this algorithm discovered by Peter Shor. It had nothing to do with the original idea behind a quantum computer. They are good at the hidden subgroup problem, and there are two prominent techniques of the hidden subgroup problem that classical computers struggle with, which is why they’re used as a basis for all these encryption standards.

One is factoring large primes. If you have 15 out there exposed as a public key, through a weird fact of math, it takes a normal computer a really long time to figure out that you could break that into 3 and 5. The bigger that number gets, the longer that computer takes.

The other one is elliptic curve cryptography, which is actually a similar problem, but with really cool math. It sounds like what it is — basically using elliptic curves as a way to find a hidden subgroup. That’s the basis for a lot of other types of cryptography, including helping secure the signature for Bitcoin.

A normal computer looks at these things and has a really hard time — age of the universe hard — to break them down and understand them. Whereas a quantum computer, because it has this exponential speedup, on a 3 to 5 year timeline, would be able to solve that hidden subgroup problem and break the cryptographic standards that we have.

What worries me isn’t just that we have a wildly short time to move to a new cryptographic standard. It’s that lattice-based encryption — which is the standard that NIST says we should move to. While the theory behind it is very good, and folks think a quantum computer would have a really hard time addressing those encryption standards, the implementation is really not mature.

We have to move faster than we ever have to a new encryption standard, but the one we’re moving to hasn’t been deployed at any real scale. You put those two things together, and it’s something I worry about. That’s something a lot of people worry about.

Chris Miller: On the encryption part, the thing I haven’t fully thought through is that for AI, there’s this AI race, but the fruits of it are kind of far out — it’s productivity enhancements. Whereas for decryption, the fruits are immediate if you get there.

How do governments in both the US and China think about this? If we’re six months away from breaking encryption — we’re never going to be six months away from the fruits of AI because it’ll always be constantly bearing fruit. But if you’re six months away from decryption, at what point do you just say all quantum computing resources must be devoted to this task, Defense Production Act style? It seems highly plausible China would do that. And it seems possible we’d do that too if those were the stakes.

There’s interesting game theory around that dynamic. “Bomb the data centers” was the not serious — or maybe some people thought it was serious — meme from 2022 or 2023 about what if AI gets out of control. But it starts to become a little bit more plausible in the quantum space if the stakes are that all cybersecurity disappears.

Jordan Schneider: Well, it seems harder to bomb a quantum computer though, right?

Chris Miller: Because it’s just one room you can put anywhere. And the know-how presumably continues even if you destroy the physical device.

Zachary Yerushalmi: I’m trying to figure out — there’s stuff around nuclear chemistry, which is really scary for quantum computers. It’s one of the many reasons that folks care about them. And again, none of this is on the high side. Do you think it’s that different from “we’re six months out from AGI”?

Chris Miller: Well, if you think that AGI is a threshold where before you have nothing and afterwards you have superintelligence, then the game theory is similar. But I don’t think we really believe that there’s an AGI threshold that has a dramatic before and after relative to an ongoing gradient where you get better and better capabilities with more and more productivity.

For most economic applications of quantum, it looks like — I don’t know if steady is the right word, but a trend over time. But decryption is this threshold dynamic where if you’re on the wrong side of it, the stakes are high.

Zachary Yerushalmi: The one thing I would say there — yes, absolutely. Stakes are super high. Global Western governments committed something like $23 billion to quantum in the last three years. I’m sure they’re really excited about molecular modeling, but they’re probably mostly really scared about the encryption side of it.

The one thing I would call out is that the first computer that gets there is not going to be very efficient at breaking those codes. It could literally, depending on the architecture, take a month or many months to break one code. Which means you have to choose your bullets very assiduously. That’s depending on the architecture. But I do think it is a little bit more akin to your AI model — just because you got there doesn’t mean that there’s more juice to squeeze.

Chris Miller: But isn’t the first code that you break the Chinese nuclear codes? That’s a pretty high-value code. There are some pretty high-value codes you could break right away and justify thinking about it as pretty important. I don’t know anything about the Chinese nuclear system, but that seems like exactly where one would go if one was going to think about the high-value code.

Zachary Yerushalmi: There are people with clearances far above my pay grade who probably know the concentration levels of Chinese high-value code. I don’t have a good sense of that.

What worries me most about quantum computers, aside from codebreaking, is their recursive nature and how they improve existing material science applications. Take high-temperature superconductors or nuclear chemistry, for example. If you had a system that could rationally design superconductors or chemical compounds, you would use that capability to lock down IP space and know-how in a way that blocks out adversaries and competitors.

From a moat perspective, it’s not just about building the system — it’s about securing the inventions that the system creates. AI is essentially sophisticated curve fitting, like stabbing in the dark. Quantum computers are fundamentally different. They’re not guessing — they solve problems from first principles and lock in on the correct solution. When I think about silent, mushroom cloud-level implications, that’s where my mind goes — it gets a bit scary.

The Bottlenecks: Talent and Time

Jordan Schneider: Let’s talk about that computer. It won’t be an engineering challenge like the Manhattan Project that costs 100 times more than any previous project, will it? Is it more likely to be an engineering breakthrough, or can you brute force your way to a useful quantum computer with enough money — one that could actually break nuclear codes?

Zachary Yerushalmi: Honestly, especially now that we’ve crossed the threshold where every additional qubit makes the system more powerful, you can just throw money at the problem. You’d build a wildly inefficient, expensive computer, but it could break RSA encryption a few times, which would be catastrophic.

How you spend the money is crucial. You could spend a trillion dollars on quantum computing, but the bottleneck is talent. You need humans to wire the refrigerators and set up the optics tables required to operate these systems. The prioritization of spending is everything. You could throw money at this problem all day, but without the right allocation, you’d overfeed the system and still fail to achieve your goals.

Chris Miller: It’s like AI — talent is the problem. Meta’s offering $100 million salaries for quantum researchers.

Zachary Yerushalmi: Talent is critical, but if I were to create a metric to track progress, I’d focus on iteration loops or cycle times. There’s the commercial cycle time — how long it takes to sell a company for significant returns. That’s important because it excites venture investors and attracts talented startup founders.

Then there’s the technology cycle time — how long it takes to go from idea to widget to product and test it in a relevant environment. Many bottlenecks exist here. While talent can be a constraint, access to technical services and capabilities often poses bigger challenges. These include superconducting fabrication facilities, specialized III-V semiconductor fabs, and scaled cryogenics systems. The bottleneck isn’t necessarily talent — it’s having the right policy framework to enable access to these resources.

Chris Miller: This gets back to the discussion of whether we’ve moved from a science phase to an engineering phase — not discounting the future science that has to happen — but do we have the right institutions for that scale-up? In the semiconductor space, there is agreement that it’s gotten way too hard and expensive to take an idea and translate it into a prototype. Prototyping is expensive, and you need exquisite equipment, materials, and so on. The same is basically true in the quantum space. Going from idea to prototype is hard because prototyping is expensive and needs this unique toolset. Talk to us about what has happened and what else needs to happen to facilitate that scale-up process.

Zachary Yerushalmi: It’s a good question — it was actually something I was chatting with Constanza about regarding her quantum supply chain paper. The short answer is we have the cards that we have in terms of the institutions in the US and the Western world. You have fundamental research, and we should be thoughtful about the things we incentivize with that. You have the free market and the private companies that are racing at this.

If I could create one institution, it would be an IMEC. If folks aren’t familiar with that, it’s canonical in the semiconductor industry — having institutions that are public, private, nonprofit, and they focus on this liminal intermediate phase after fundamental R&D but before it’s pretty competitive with the market. They just get good at that middle TRL phase. It turns out that you need to have institutions that all day long build that as a craft. That’s both an expertise, a capital structure, capital itself, and physical capabilities. You need specialized instrumentation that’s only good at that phase. From an institutional basis, that’s the one area I focus on as a missing potential piece.

Jordan Schneider: Given that Constanza is going to be next in our quantum series, why don’t you tease and pitch her work a little bit?

Zachary Yerushalmi: The teaser for this is — and Chris is going to be the emcee for the release of the report so there’s more than one person who can sing her praises — Constanza Bustamente is, on every dimension, the leading quantum policy researcher out there. She’s at CNAS. She did a definitive study on quantum sensing, which everybody on the planet should read. It went both breadth and depth — the best out there.

As a follow-up, while a lot of folks focus on quantum computing, which is great — right back to the drug discovery analogy where you need to focus on the individual drugs that cure cancer or whatever they do — she wanted to drill down and look at the quantum supply chain. What are the things that enable us to develop these quantum computers? She uses it as a framing how to stay competitive and how we lock down a capability in the supply chain. The report is coming out in March. Again, Chris, you’re actually closer to this than I am, but it is a must-read for anybody who cares about advanced technology policy and competitive advantages.

Reading Recommendations and Quantum Jobs

Jordan Schneider: Zach, I would love you to make a pitch for the syllabus that we’re going to put in the show notes. We already talked about the “But What Is Quantum Computing?” YouTube video, as well as Constanza’s recent report. What about the quantum-classical divide? Systems engineering bottlenecks?

Zachary Yerushalmi: The quantum-classical divide is just fun weekend reading on us being on the cusp — not just of these fault-tolerant systems, but really a better understanding of how the atomic world adds up and meets the classical world that we’re all used to. While probably not the most important from a policymaking decision standpoint, it’s pretty cool as a human being.

The system engineering bottlenecks actually provide one of the best breadth and depth, deeper views of quantum computers as a system and where they fall down and where we need to prioritize. I would say it’s more with a research academic lens to it. Costanza’s report is a really nice complement because it goes a little bit more into the policymaker dimension on what we have to prioritize.

There are a couple of others which are fun and very ChinaTalk-esque. When We Cease to Understand the World — quantum breaks your brain a bit. This book is probably the best that I’ve come across at capturing what it’s like to be in the mind of somebody whose brain has broken because of quantum.

Last, this one’s more weekend reading. But it felt like a very ChinaTalk recommendation because it’s a The Social History of the Machine Gun approach applied to the early history of nuclear fusion. Again, fun read. It turns out that through weird and wacky accidents, those can be the difference between life and death for some of the most important programs of our time, and I just love that little lens on the world.

Jordan Schneider: Are there good quantum podcasts?

Zachary Yerushalmi: Some can be more or less advertisements for companies, which are great. And we love these companies. But the one I like is New Quantum Era.

Jordan Schneider: Can you give us a little anthropology of quantum researchers? What brings you down this path? What kind of personalities do you get relative to other fields?

Zachary Yerushalmi: One of the biggest learnings in my career is that the people it takes to solve every particular chain of an innovation cycle — you need a different personality for every single type. There’s an infamous distinction between theoretical physicists and experimental physicists. Theoretical physicists are locked away in some room with a bunch of chalkboards, and their dopamine hit is them with chalk and paper or whatever it is.

Experimental physicists are different because they work in teams. What determines all of this is where you get your dopamine hits from. If you are a fundamental researcher, you don’t get your dopamine hit from reliably solving a problem, because the definition of fundamental research is that you don’t know when you’re going to solve that problem. You get your dopamine hit from asking an interesting question and finding something interesting about that.

Now that we’re transitioning into an engineering field, it’s a very different mindset because engineers often get their dopamine hit from solving a very specific problem that folks have solved before, and it’s working through it. How people find themselves — back to the top of the question — is starting with what motivates them and then mapping that to the right part of the technology innovation cycle.

What I would say for me, oddly enough, it’s different. I’m not coming from a science angle. A lot of it is trying to find the right mental model. It’s being curious and finding the thing that I can never fully scratch the itch of curiosity on. Then it’s trying to find the right mental model to meet that moment.

The last — this really transitions to a totally different phase — is folks with a sales discipline. There, it’s about winning deals. The fascinating thing about really anything that you’re trying to do that’s a team sport, but particularly with quantum, is you need to align folks with not just wildly different expertise. You need to align folks with wildly different passions and motivations and get them all to work together because you have to solve things all the way from the fundamental physics up through making a killer deal and a lot of money.

Jordan Schneider: That’s cool, but it’s also hard. Is this the path of least resistance if you’re a physicist and you want to do cool stuff nowadays? Do you see talent being drawn to the field thanks to the recent breakthroughs?

Zachary Yerushalmi: Yeah. Just not fast enough. There’s a famous stat that for every three quantum job postings, you only get one qualified candidate. There’s a lot of demand for this stuff. We need to address that badly.

The issue with addressing this, especially on these timelines, is that it takes five to seven years to get a PhD. If we’re going to surge resources to this, it gets back to the fact that you can spend infinite money, but you can’t compress the timeline for a PhD from seven years to two. We actually have to address this in a very different way from a talent perspective.

Chris Miller: What would you train more of today? You mentioned that you need salespeople who can sell quantum capabilities alongside the people who can actually do the fundamental engineering. If you could train X thousand people in discipline A or B, what would that look like?

Zachary Yerushalmi: Quantum system engineering would be wildly important. The other bottleneck would be technicians.

I found out something fascinating recently. If you are a technician or an undergrad, even a master’s level, and you’re trained in quantum, it’s all theory-based because the physical systems you need to do the training are so expensive and so exquisite. Nobody’s going to give a bunch of students access to something that costs a couple of million dollars if they break it. I get that.

But it also means that when you graduate and get hired by a company, you have to get trained from scratch because you haven’t had access to the physical system in the first place. I would be prioritizing those roles and things like physical access. It’s actually a good news story because we can do something about that. You can give access to these physical systems. You can spend the money and solve your problem. I love those sorts of problems if we can find them.

Costs and Cycle Times

Jordan Schneider: Claude tells me post-Willow, the realistic all-in cost for RSA-breaking quantum computers is $10 to $50 billion.

Zachary Yerushalmi: Cost per calculation is super important. It’s one of the factors that the government is trying to assess. Just on Willow architecture, absolutely. But some leapfrog capabilities would wildly bring that down.

Back to the biotechnology comparison — it costs $1 to $4 billion per approved drug. $5 to $10 billion for the first computer that can break RSA makes sense. The Human Genome Project took how many billions? It makes sense in my simple head.

Chris Miller: Your point about the cost of a computer is only relevant if you also know cost per calculation seems very important. How do we think about the spectrum of outcomes and time horizons on cost per calculation, and how that’s going to change over time?

Zachary Yerushalmi: It’s really early to say. One of the things the QBI is trying to examine is how much it costs to perform calculations across these modalities, and whether that cost makes sense for specific problems.

For example — and I’m making this up — if it costs $10 billion over 5 years to do a physics-based simulation of penicillin, that’s probably not worth it. But if we’re talking about an architecture that costs a week and a million dollars, then suddenly the economics look very different. Some architectures have hope of reaching that cross-profile, while others simply don’t. That’s one of the fascinating things to assess.

Here’s what I’m curious about, Chris. We alluded to this earlier, and I’d love your take on two things. First, is cycle time a useful metric as a North Star for industrial competitiveness? Cycle time as in how long it takes to make your product.

Chris Miller: I would consider that as one input into a broader rate of innovation and improvement — but only one input. Cycle time is important, but the differential between cycles also matters.

If you have a long cycle time but achieve huge improvements between cycles, that’s probably okay. For instance, if it takes TSMC a year to move to the next node, but that next node delivers significant improvements, maybe that’s fine. However, longer cycle times become problematic when your differential is smaller. Those are the two key factors I’d consider.

Zachary Yerushalmi: Let me give you a specific example. To build one of these systems for one of the main modalities, you need something called a photonic integrated circuit. Constanza discusses these extensively in her report. Think of it as the photonic equivalent to an integrated circuit — you really need this component to build scaled systems.

For some of the largest players, it’s not just about availability. They can actually get access to a PIC, but the cycle time can be 12 to 18 months. In contrast, if you’re located next to one of the fabs that manufacture these and have good availability, your cycle time drops to a matter of weeks.

This could be wrong, but when I look at China’s advantage in 5G and other photonic technologies, their defining advantage was the ability to build and test products an order of magnitude faster than American equivalents.

Chris, I appreciate your characterization that there are two factors — how fast you can make and test something, and how capable you are of learning from it. We Americans excel at the second one — learning — because we have talented people here and can leverage brains from around the world. But where we’re wildly far behind is cycle time. It’s not just an availability issue — it’s about how quickly we can actually get the thing.

For games like quantum computing, where we have a really small margin of error, my focus is on reducing that cycle time. If we think of the Euler diagram of what’s both important and actionable, wildly reducing cycle time would be the best engineering-style measure to cement a competitive advantage. It provides a non-market but very clear signal on where to prioritize — identifying what’s holding us back from building systems quickly and figuring out how to address those bottlenecks.

Chris Miller: What’s the limit to cycle time today? Is it that companies producing component X or Y don’t see quantum as core to their business? They have other customers buying at higher volumes, so they don’t want to prioritize quantum because it’s still science project-sized rather than commercial scale. Is that the main reason?

Zachary Yerushalmi: Yes, or there’s almost no market-clearing price that would make it valuable for them to do that. What we’re building with this federal award costs us $40 to $50 million just to stand up the fab. We hope to make about a million dollars a year from it.

If I were a commercial company or a VC investing in this, it would be a clear no-go. But because this is a nonprofit, the investment becomes really valuable.

Looking across the landscape, we need to ask — is there a short-term market for these components? If so, the private market can address it. If not, we need an institution that continues to focus on that intermediate TRL (Technology Readiness Level) step.

In semiconductors, basically every leading semiconductor ecosystem has institutions operating as public-private partnerships. On a VC basis, they don’t make sense, but for national competitiveness and the economic competitiveness of their partner companies, they make lots of sense.

IMEC serves this role, as does KAIST in Korea. Taiwan has their equivalent of IMEC. China has loads of these institutions.

That’s what worries me — on our current trajectory, this is the one area that could get overlooked if we don’t have cycle time as our defining metric.

Jordan Schneider: Thanks, Zach, for getting us started with our quantum journey.

Zach’s Quantum Technology Reading List

• Quantum Computing Fundamentals: But What Is Quantum Computing? by 3Blue1Brown — A visual, mathematically rigorous explanation of how quantum computers actually work, building up to a complete walkthrough of Grover’s search algorithm. The best starting point available for non-specialists.

• Quantum Computing Overview: The Map of Quantum Computing by Domain of Science — A comprehensive 33-minute tour of the entire field, covering algorithms, hardware approaches, applications, and the key obstacles to building useful quantum computers. The roadmaps are now dated but the modalities are still relevant (minus silicon spin, which has really taken off).

• Quantum Sensing: Atomic Advantage: Accelerating U.S. Quantum Sensing for Next-Generation PNT by CNAS — Dr. Constanza Vidal Bustamante’s 2025 report flagging quantum sensing as the most mature quantum technology today, with near-term national security and commercial applications in GPS-resilient navigation — an adjacent but urgent defense priority for governments globally. Constanza rocks.

• The Quantum-Classical Divide: Are the Mysteries of Quantum Mechanics Beginning to Dissolve? by Philip Ball, Quanta Magazine (February 2026) — A fun look at Wojciech Zurek’s decades-long program to explain how the quantum world becomes the classical one. Zurek argues that entanglement with the environment “selects” which quantum states survive into observable reality — a kind of Darwinian process that may finally explain quantum weirdness without invoking parallel universes or conscious observers collapsing wave functions.

• Systems Engineering Bottlenecks: Computer Science Challenges in Quantum Computing: Early Fault-Tolerance and Beyond by Jens Palsberg et al., IEEE Quantum Week (2025) — A 90-person community report arguing that the primary bottleneck in quantum computing is shifting from physics to computer science — compilers, architectures, and system integration. Notable for its candid assessment that industry roadmaps should be read as “aspirational, not predictive,” and its identification of the dequantization arms race, where classical algorithms repeatedly match claimed quantum speedups.

Further reading if curious:

• When We Cease to Understand the World by Benjamín Labatut (2021) — It’s been a while since I read this, but it’s a classic. Shortlisted for the International Booker Prize and a NYT Top 10 book, a blend of (mostly) fact and fiction to tell the stories of Heisenberg, Schrödinger, Haber, and Grothendieck . The closest thing to being in the mind of a physicist navigating the implications of quantum, genius, madness, and destruction their work has and will cause.

• Introduction to Special Issue on the Early History of Nuclear Fusion by M. B. Chadwick and B. Cameron Reed, Fusion Science and Technology (2024). Not really germane to new modern quantum tech but felt very ChinaTalk! Mark is a lovely human and the archives at LANL he has access to are fascinating.

Just as Pluto is technically not a planet, gallium is technically not a rare-earth element despite often being discussed in the same context. Like many rare earths, gallium is not directly mined from the Earth’s crust but rather a byproduct of aluminum extraction. Although not classified as a rare earth, the mineral plays a major role in compound semiconductors and has critical importance for the future of AI, defense, robotics, and more.

China has realized the element’s importance and has quietly shored up its supply chain while the U.S. has been asleep at the wheel. Now, the U.S. must secure this critical mineral and its downstream technologies before another lead slips from our hands.

The Problem

China’s recognition of gallium as a priority — both for domestic development and weaponization against adversaries — is unmistakable. As a result of their efforts, China is responsible for 99% of raw gallium production today.

Since the early 2000s, China has required domestic aluminum producers to also extract gallium, which has enabled the country to not just become self-sufficient but dominate the global market for gallium extraction. In the meantime, the U.S. has not shored up its supply chain insecurities, particularly in upstream extraction, leaving America vulnerable to weaponization of the mineral.

Such vulnerability is not just hypothetical. China noticed its leverage and imposed export restrictions on gallium (and the tools to extract it) since 2023. These export controls wreaked havoc on gallium prices in the global market, and firms have reported trouble in securing licenses for required gallium. As China builds up dominance over the products downstream from gallium, the United States should be worried about a future where industries are cut off from critical semiconductors and begin working now to ensure that such a threat is neutralized.

This is the current story for upstream gallium — the mineral itself. America’s dependence on China for upstream gallium has been covered excellently by other institutions like CSIS and the Atlantic Council. To address this dependence, the U.S. must actually follow up on its many ongoing projects to produce gallium domestically.

However, a less-discussed security issue is looming: the dangers facing downstream gallium — that is, the products made from gallium. China’s downstream gallium semiconductor industry has begun to encroach on the viability of American and allied companies. Instead of panicking when it’s too late, the U.S. must address its impending downstream gallium crisis in tandem with its already-existing upstream gallium problem.

The Downstream Competition

Gallium in Power Semiconductors

What is gallium used for, and why has China emphasized it so much? The mineral forms the backbone of semiconductors like gallium nitride (GaN) and gallium arsenide (GaAs) chips, which are irreplaceable for certain defense, power, and optoelectronics applications.

One of the most critical of these uses — and the one most under threat — is in power semiconductors, typically using gallium nitride (GaN). GaN chips used for power functions are often referred to as GaN high electron mobility transistors (HEMTs). GaN HEMTs, though currently a limited market, are increasing in popularity due to their use in EVs, motor control for robotics, and power solutions for data centers. Currently, their biggest market is the consumer end-market, focused on products like fast chargers for your laptop and phone. While consumer end-markets will likely remain GaN’s biggest cash cow, it punches above its weight in terms of irreplaceability for humanoid robotics, data centers, and EVs.

GaN, alongside silicon carbide (SiC), is considered a wide bandgap semiconductor, which endows it with properties better for power electronics compared to standard silicon. These properties include faster switching and better power efficiency. Although SiC chips are able to stand in for GaN in some contexts, GaN for power is largely irreplaceable due to its faster switching and better performance at lower voltages. Generally, SiC is used in heavy-duty applications like large industrial robotics, whereas GaN is used for lower-voltage applications like smaller humanoid robots.

Innoscience’s Rise

With respect to GaN power semiconductors, the U.S. has already lost its lead and is at risk of being pushed out altogether. Like the story with solar panels and electric vehicles, the U.S. (alongside Europe) built up a lead in the “higher-value” segment of products by being a first-mover, but the lead was promptly chipped away as sprouting Chinese companies buried American firms with unbeatable prices.

Here, the main competitor is Innoscience (英诺赛科), a Suzhou-based GaN integrated device manufacturer (IDM), whose prices are nearly 50% lower than competitors’. As a result, Innoscience now leads the global market for power GaN chips, beating out the American Navitas and EPC and German Infineon. Other players like STMicroelectronics and Onsemi have bent the knee to Innoscience by giving up packaging expertise, system integration, and their own manufacturing capacity in exchange for access to Innoscience’s production facilities in China.

As Innoscience continues to expand capacity, the situation risks shifting from one of market dominance to one of market monopolization. If trends continue, competition in the GaN power market will become a fiction, constituting a national security threat to the U.S.

So, how is Innoscience so much better than its competitors? The answer boils down to the synergy of in-house manufacturing, a stomach for unprofitability, government support, and genuine innovation.

In the GaN market, AMD co-founder Jerry Sanders’s adage holds true: real men have fabs. After Innoscience, the other two leading GaN makers include the American companies Navitas and EPC. Both are fabless. Both must rely on external foundries for their chips, which increases the cost of their final products.1

From the beginning, Innoscience decided to spend the money on R&D to make its own fabs, and its bet has paid off. Both Navitas and EPC have relied on TSMC for its fabrication, but TSMC is now exiting the GaN market entirely. Now, their business is getting punted off to Taiwan’s Powerchip (PSMC) and American GlobalFoundries because TSMC realized its capacity was better used for the more lucrative AI chip market.

Fab capacity for GaN is trending toward Innoscience holding all the keys. By being the first to mass-produce 200mm GaN wafers, the unit economics are in Innoscience’s favor. Compared to the previous standard of 150mm wafers, 200mm wafers allow for up to 80% more chip output at 60 to 70% of the cost. Further, by being first to the scene, Innoscience has had more time to perfect its process, achieving a yield of about 97% whereas others are stuck below 90%. Innoscience’s capacity also blows competitors out of the water, producing nearly four times as many wafers as second-place TSMC. With Innoscience having no intentions to slow down, the unit economics will just get better and better for the Chinese IDM and worse and worse for everyone else.

Companies like Onsemi and STMicroelectronics realize that the cheapest way to fabricate their designs is through Innoscience, creating a dynamic that essentially positions Innoscience as the TSMC of GaN. The question now is how much longer can Navitas and EPC find fabs that aren’t Innoscience to fabricate for them? And then in the long term, why would Innoscience ever want to fabricate for a direct competitor when it could instead monopolize the GaN power market? Even for Onsemi and STMicroelectronics, after market consolidation, Innoscience may devour its children.

Innoscience was able to become the greatest GaN company by being willing to stomach unprofitability. In 2021, the company was operating with a gross margin of over negative 266%. Unlike Western companies, Innoscience — and its funders — have been willing to eat bitterness while it figured out its manufacturing process, increasing yield and expanding capacity. American markets do not have the same willingness. Other GaN makers have been incentivized to maximize profit margins in the short run while Innoscience chased viability over the long run, leading to where we are now.

Now, Innoscience has been able to capitalize on its high-yield manufacturing process and exploding demand for GaN for high-tech applications to achieve positive margins for the first time in its history. Although the company likely won’t turn a profit until 2027, the upward revenue trend contrasts Innoscience with that of other GaN players. (Quarterly revenue from GaN sales alone is not available for some companies.) And if Innoscience was not deterred by negative margins in its early years, the company will most definitely not be deterred now.

Part of Innoscience’s perseverance in the face of negative margins is due to assistance from government subsidies. The combination of investments from national and provincial state-backed funds has totalled over 350 million dollars of financial support at minimum for the then-burgeoning Innoscience. That is more than double the company’s gross losses since 2021. By the time of its IPO in 2024, the company had established enough capacity and was already poised as the best option for large-scale GaN manufacturing. Other companies like STMicroelectronics realized this, and they decided to become a cornerstone investor in Innoscience with a $50 million investment and further fund the GaN giant.

But before we lazily blame the evaporation of Western market share on government subsidies, we must reckon with the reality that Innoscience has also simply played better than the U.S. Competition in the GaN power market is more intense for individual voltage ranges. Some companies, like EPC, focus only on the sub-350V range. (Products in the sub-100V range are used for motors in humanoid robots, sensors and ADAS for electric vehicles, and motherboard power conversions in data centers.) Most companies expand that focus up to 650V or 700V. However, Innoscience is the only company that both designs and manufactures GaN power chips across the whole spectrum, from 15V up to 1200V.

And they are not low-quality chips, either. For example, Innoscience designs and produces 650V and 100V GaN products for rack-level power conversion in AI data centers. Innovation in this increasingly critical segment enabled Innoscience to become Nvidia’s sole Chinese partner for this power architecture. The 800 VDC power architecture is touted as the best option for the “next generation of AI factories” because it allows better power efficiency and less reliance on copper cables. Although large companies like Nvidia will always qualify more than one supplier for diversification, Innoscience will likely emerge as a primary supplier if its prices and quality remain preeminent.

Lest I risk fearmongering, it is important to note that none of these 800 VDC GaN designs by any company have been qualified as of this piece’s publication. They are all simply reference designs that Nvidia has requested from these companies. A rudimentary analysis also suggests that Innoscience’s competitors have created better products for this application; for example, Navitas’s product supports an output of down to 6 V, suggesting better capabilities for handling high current. It is unclear how important this functionality is and what the cost differential is for these products. If any reader with a background in GaN would like to provide answers, please comment or reach out to aqib@chinatalk.media.

Regardless, such innovation cannot be swept aside and blamed on government subsidies; the U.S. must contend with Innoscience as a company with the ability to both produce at scale and innovate. These characteristics enabled Innoscience to establish its partnership with Nvidia (and now Google) for the future of AI data centers.

And regardless of the extent of government subsidies enabling Innoscience’s rise, the U.S. cannot just call foul play and say it isn’t fair. There is no referee. We must take fate into our own hands and fix the problem ourselves. The U.S. has prided itself on government programs such as DARPA shepherding critical technologies like GPS and the Internet before they were profitable. Can we not do the same for manufacturing critical technologies like GaN?

We now find ourselves in a position where the snowball is forming. If we do not prevent it from getting bigger, makers of robots, EVs, and data centers may reasonably be dependent on a single Chinese company for its power chips. Do we seriously believe these technologies will become less important in the future? In the next trade war or diplomatic spat, this is worrying leverage that China could use to bottleneck critical industries. Does this not mean we should be trying to stimulate GaN production, not throw its carcass to the vultures?

The Solution

Fortunately, it is easier to fix the problem now, when we still have some GaN players, compared to later, when the outcome is set in stone. To ensure the U.S. is not overreliant on China for critical GaN products, we must support allied industry to make producing GaN a profitable venture. We should perhaps limit competition in the short term to create healthy competition and stable supply chains in the long term. This does not mean the extermination of Innoscience, but rather the protection of market competition.

Policy should also recognize its limitations, however. The U.S. cannot and should not spend obscene amounts of money to compete with China on capacity. Instead, we must focus on winning on efficiency, innovation, and other methods that give us the edge besides raw buildouts.

Patent Infringement Cases

The quickest relief is through the judiciary. Both EPC and Infineon have filed patent infringement cases against Innoscience, and the results of those cases could limit Innoscience’s ability to compete in the American market. Although EPC’s claims were invalidated by the USPTO, import restrictions imposed by the ITC continue to be enforced. The Infineon case will be finally decided on May 7 by the ITC as well.

The ITC’s determinations, however, will not be a panacea. The patent infringement punishments only apply to certain products, and Innoscience would be able to design around them to continue to sell in the U.S. Further, the determinations would not be able to restrict finished products containing Innoscience chips. Especially when the current money makers — consumer end-products — are largely produced in China, the case determinations may not produce a serious impact. This route is also not a policy position, as the judiciary should not bend the rule of law for policy goals.

The Race to 300mm

Outside of the judiciary, the U.S. can support innovation and the commercialization of the next generation of GaN power semiconductors. Here, the best options for champions are Texas Instruments and Infineon. Both companies have dedicated foundry space for GaN power semiconductors, and both have piloted the production of 300mm GaN wafers. Where Innoscience was able to achieve superiority in unit economics from the shift from 150mm to 200mm wafers, TI and Infineon can perhaps achieve it in the shift from 200mm to 300mm.

However, the gains from 200mm to 300mm may not be as large as the gains from 150mm to 200mm. Although 300mm wafers produce about 2.25 times as many chips per wafer compared to 200mm, the throughput for processing may not be as high. For epitaxy, 300mm wafers currently require single-batch processing due to strict requirements for wafer uniformity and robustness, whereas 200mm wafers allow for multi-batch processing. Development of multi-batch 300mm wafer tools is almost certainly ongoing, but no progress is yet visible. The overall cost savings and throughput advantages of the 300mm transition are still unknown, but they may not be as impressive as the previous 200mm transition. The step to 300mm is a step toward the ultimate objective for GaN manufacturing — cost-parity with silicon — and it is an important step toward reducing dependence on Innoscience. However, it is not a panacea.

America’s export controls on the metal-organic chemical vapor deposition (MOCVD) tools required for GaN epitaxy (ECCN 3B001 a.2.) may enable the 300mm wafer lead to be enduring. Infineon and TI have been able to achieve pilot production because they have been able to purchase the relevant MOCVD equipment from the German AIXTRON and American Veeco, whereas Innoscience must wait for domestic suppliers like AMEC to develop a solution. AMEC has no visible progress toward 300mm GaN, so export controls will perhaps give TI and Infineon more time to develop and mature process flows for 300mm GaN.

To goad TI and Infineon on, the U.S. may fund projects through the CHIPS Act to support the quicker construction (or conversion) and operation of 300mm GaN fabs. By accelerating the timeline to mass production, homegrown companies will more quickly improve yields and unit economics so Innoscience’s explosive capacity expansion would not be so oppressive. We cannot build as much as Innoscience, but perhaps we can build better.

Ecosystem Stickiness

The most enduring solution would be to create ecosystem “stickiness” for end-customers so that they are more locked into purchasing from allied companies. The West again has an inherent advantage here, with allied GaN makers (mainly U.S.-based Texas Instruments and Germany’s Infineon) being IDMs across the semiconductor stack; unlike Innoscience, they do not solely focus on GaN.

For end uses more complicated than fast chargers (e.g., data centers and robotics), GaN becomes less of a commodity and more a question of integrated solutions and technical capabilities. End customers would be more willing to work with GaN suppliers that could tailor their manufacturing solutions to the customers’ power architecture, which presents an opportunity to reduce the importance of Innoscience’s price lead.

For example, when a company wants to purchase a GaN power HEMT for their humanoid robot, they should be incentivized to purchase a system, not just the product. If they are already using a TI MCU, it should pair best with TI’s gate driver ICs, TI’s sensor chips, and TI’s GaN HEMTs. By contrast, there is no such thing as an Innoscience MCU. When the full-stack comes with so many advantages, customers are incentivized and better served by sticking with TI, rather than considering redesigns to drop in a cheaper Innoscience product.

Innoscience simply does not have this ecosystem capital outside the GaN stack, and unless they quickly partner with Chinese companies across the stack, they will not accumulate such capital soon. Currently, they must rely on products from companies like TI and Taiwan’s YAGEO for reference designs of motor drives.

To capitalize on this ecosystem advantage, the U.S. could consider providing modest funding for better open reference designs for applications like robotic motors, EV onboard chargers, and data center power topologies. Companies are already incentivized to pursue this, and TI already does this well, but coordinated government funding could reduce barriers and promote better designs. If the U.S. produces powerful reference designs that perform well with potential robotics MCUs, data center power topologies, POL parameters, and vehicle architectures, then end-customers may not care about the marginal savings of Innoscience’s GaN HEMTs.

The primary source of pessimism with this strategy, however, is that American reference designs may not matter if the end-customers are Chinese. If Unitree and BYD are the main end-customers, they will likely work with Chinese MCUs (like ARTERY) and be incentivized to work within the Chinese ecosystem. The American GaN market will miss out. This is not a fait accompli, however. Chinese carmakers like Changan Automobile have opted for American Navitas GaN chips for their onboard chargers, meaning Chinese OEMs can be incentivized to pick American products over Chinese ones.

Further, larger companies like hyperscalers tend to have their own engineers who do not need to rely on the easy reference designs given to them; they make bespoke designs in house and take the best products for each segment, prioritizing cost savings and performance over ease of use.

Still, funding design is significantly cheaper than funding factories, and better reference designs may trickle down to benefits for start-ups in the robotics industry where the major players have yet to calcify.

Flexible Fabs

Lastly, though most vaguely, the U.S. should incentivize companies to make it as easy as possible to convert legacy fabs into GaN fabs if the need arises, just as we did with factories during World War 2. Although this would mostly be easy, as GaN wafers can be processed by the same equipment used in depreciated legacy fabs, the biggest obstacle would be ramping up the epitaxy for GaN wafers. In this case, possible options include encouraging a GaN wafer stockpile or promoting expedited production of MOCVD equipment for GaN epitaxy.

Conclusion

The U.S. is largely aware of its upstream gallium dependency, and 99% dependence is a difficult ditch to climb out from. But let’s ensure that we do not fall into the same ditch when it comes to GaN.

The U.S. can accomplish long-term viability in the GaN market now before Innoscience makes it too difficult. We can accomplish this through innovation and flexibility, not expensive buildouts, via the pursuit of 300mm wafer adoption, ecosystem stickiness, and flexible fabs. These are not the only tools in the toolbox, but they are feasible options that the U.S. government could readily pursue.

We also do not need — and probably should not want — to banish Innoscience. American and allied companies like Onsemi and STMicroelectronics work with Innoscience, and punishing one would be punishing the whole lot. Instead, we should focus on preventing Innoscience from becoming a monopoly and encourage companies to work within the American ecosystem instead of compelling them to settle for a Chinese one. A world with Innoscience and at least one allied viable alternative is a win.

Instead of sleeping at the wheel (again), the U.S. can prevent GaN from going the way of solar panels and EVs. If we want to secure our supply chains, we can start with GaN.

The author would like to thank several GaN industry executives for their contributions to this piece.

Marrying her was the best decision I ever made. Starting a company with her may be the best one you make!

Once upon a time, I saw the Komodo dragons in their natural habitat. They were majestic, imposing and the chill they sent down my spine was unforgettable. I am no dragonologist, but a friend and I decided to meet up midway, and that landed us on the Komodo island.

For me, it’s the adventures that give life its magic and the most epic ones start with companions who expand our worlds and draw us to unexpected wonders.

Today, I am seeking a partner to embark on the ultimate commercial adventure: entrepreneurship.

Perhaps you are about to team up with another developer to build another dev tool, but let me open the door to a whole new world for you. If this goes well, we’ll spend the next decade together with an exceptional team of fellow adventurers to push the frontier of a real industry.

You might be a good fit if you are:

1. Based in NYC

2. Already committed to starting a company

3. As good at your craft (ML+data) as I am at mine (see below)

4. Open-minded but decisive

5. Ambitious; our approach will be frontier and radical, not midwit or marginal

6. Down for a 10-year adventure, which may or may not involve riches or glory, but will definitely involve great people and honorable work

About me:

1. I am 10 years out of college, had some big responsibilities professionally, and am now also raising a toddler with Jordan!

2. As a private equity investor for 7 years, I evaluated hundreds of companies (and founders). My portfolio companies exceed $5bn in total valuation, ranging widely from $3mm to $2bn. On this adventure, I developed clarity of thought, a sense for order of magnitude and taste in people

3. As an operator in recent years, I defined strategies, closed deals, opened new markets, ran payroll, delivered hiring and firing decisions, etc. On this adventure, I recognized how much I have to give. One startup I coached received a ~100% valuation increase after I rebuilt the narrative and told them the numbers to ask

4. Over the years, I earned the trust of some people who think the world of me

5. I go very deep into very niche topics, in order to do that I have a high bar for what’s worthwhile; I will work 24/7 when called for, but not 996 every week

6. I consider the ultimate job of a CEO to defy gravity, as most companies see their upside shrink overtime. I have watched this happen up close twice, painfully. If I get this right, it will not be me alone - although I do have great instincts - but because our trust in each other keeps us honest, in perspective and relentless

The idea space:

You may have guessed, dear smart reader, that if the idea space was “sexy” I would have put it in the first paragraph. But remember, I am looking to expand worlds (yours and mine); you don’t need to be an insider because you have me.

This is a multi-trillion dollar industry with highly specialized talent. It was Franz Kafka’s day job and the core engine of Warren Buffett’s empire. High walls surround it but the people inside are generous, smart and fun!

Welcome to the world of insurance.

I have invested in insurance companies and operated within insurance companies as General Manager and VP of Finance. We won’t be selling insurance policies. We’ll solve structural, heavy-lifting, legacy tech debt that can build us a data moat to move the entire industry to the agentic era. Few trillion-dollar industries have this opportunity still on the table.

I have a hypothesis and people to call; we will work together on the strategy, product and plan. And if you have a better idea outside insurance, let’s pressure-test both.

Interested? Use this form to take care of logistics.

Oh, and I have a substack too if you want to follow along.

I work with Sid Sijbrandij, a technology entrepreneur who has taken a radically personalized and high-agency approach to fighting his osteosarcoma (bone cancer). Before meeting Sid, I was a product lead at 10x Genomics, a sequencing technology company developing novel tools for understanding biology. Sid was the first person I met who had used 10x tools to inform their care. I now run the enterprise of Sid’s care, pursuing a strategy of maximal diagnostics, making personalized therapeutics, and doing treatments in parallel rather than one at a time. Against the odds, Sid has had no evidence of disease for almost a year now. We are scaling this approach for others, both by starting companies and through philanthropic efforts.

Elliot Herschberg wrote an excellent and approachable post on Sid going “Founder Mode” on his cancer on his blog, Century of Biology. We recently gave a talk at the OpenAI forum on Sid’s journey and our approach. More details can be found at sytse.com/cancer, and 25TB of data and Sid’s treatment timeline are available open source at osteosarc.com.

Last August, Sid Sijbrandij and I traveled to Beijing for an experimental scan to look at a biomarker that’s specifically upregulated in his cancer.1 At that time, the only place we could do this was in China, using a molecule developed by Yang Zhi (杨志)’s group at Beijing Cancer Hospital. So that’s where we went.

We were stunned. The whole experience — from international patient check-in, to preparation of the radiotracer, to injection, to imaging, to discussing the result with the physician, to leaving with a glossy printout of the whole-body scan — took two hours. Even in Germany, where clinics are experienced in using developmental tracers, this process would take most of a day. Beijing broadly and the hospital specifically were surprisingly straightforward to navigate for foreigners such as us who speak no Chinese.

This experience inspired me to return to China in search of a deeper understanding of what is happening at the forefront of biotech and medicine. I often read and hear that it is becoming more difficult for American biotech to compete with what’s happening in China. I wanted to understand specifically what was going on, and what the implications were for a patient seeking the world’s most innovative care.

I spent a week in China at the end of March, visiting 5 cities in 6 days.2 I had over 25 meetings with biotech companies, investigators, contract research organizations (CROs), and contract development and manufacturing organizations (CDMOs). I came away impressed. Medical tourism is likely to invert, with patients flying to China to seek cutting-edge care. And I hope that we in America can learn from the sensible steps the Chinese ecosystem has taken and speed up our own innovation cycle. Patients deserve it.

A Marketplace of Reputation

The “investigator-initiated trial” (IIT) is an important fundamental concept to understand. Through IITs, individual physicians at major hospitals in China can propose and run studies for cell and gene therapies under the oversight of local scientific and ethics committees. There’s no need to clear a single, centralized national gate before enrolling patients. Compare that to the United States, where early trials are usually company-driven and require formal approval from a national regulatory body (like an IND filing with the FDA) before anything can begin. The tradeoff is pretty straightforward: the US system emphasizes uniform standards and upfront rigor, while China’s IIT model pushes decision-making closer to the doctor and the patient, making it easier to start trials quickly and iterate as data comes in. Carvykti (ciltacabtagene autoleucel) is perhaps the most striking example of what this model can produce. The BCMA-targeting CAR-T therapy first entered human clinical trials through an IIT in China, and has since gone on to reshape the standard of care for relapsed or refractory multiple myeloma — a disease where treatment options had long been limited for patients who had already cycled through multiple prior lines of therapy.

For more on Carvykti and China’s biotech coming of age see this feature we ran last year:

Based on what I heard on the ground, it takes about 6 months to go from a first conversation between a doctor and a patient to that patient getting dosed, and people noted that exciting programs with support from senior investigators can go even faster. This means new therapies can get to patients faster, and companies and physicians start learning from and improving the underlying therapies earlier in their development.

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.

The emergent system is a more pragmatic and cost-efficient one than what we see in the US for equivalent trials. In China, institutional ethics committees set their own manufacturing and pre-clinical data standards for project initiation. Because reputation is on the line, the standards are strict but sensible. They demand robust manufacturing controls and toxicity studies, but not at the level typically required in the US for trials of this stage. While the details can differ, the standards are similar enough — and the volumes at each institution are high enough — that CROs and CDMOs have set up IIT platform processes sufficiently mature that they could quote me approximate prices.

China’s State Council has recently adopted Decree 818 (国务院令第818号) to streamline IITs for cell and gene therapies. Prior to this regulation, IITs were popping up everywhere (particularly around regenerative cell therapies), leading to uneven data quality. With the goal of making data quality more systemically robust, 818 restricts the authority to run IITs to a pre-selected set of Tier 3 hospitals and requires Good Clinical Practice (GCP) certification for investigators. Interestingly, 818 opens the door to bring therapies to market very very quickly. Once ~10-15 patients have been treated with a therapy at a given hospital, that hospital can apply for the right to charge patients for access to that therapy. Essentially, the combination of therapy and institution is being approved. Data across institutions can also be leveraged for national approval down the line.

All of this makes sense! The system leans on the reputational sensitivity and naturally risk-averse incentive structure of academic medicine to regulate which medicines move forward to human trials. By putting trust in clinicians’ and hospitals’ judgment, the system is able to bring therapies to patients quickly.

Momentum is All That Matters

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.

In Shenzhen, I made a curiosity-driven comment about the instrumentation being used in an automated data foundry we visited. Two hours later, the head of the instrumentation company was waiting for us at the coconut chicken place we visited for dinner. If I wanted to put together a lab, he could do it.

A friend of a friend joined us in the afternoon in Shanghai. At that point, we’d only exchanged about two sentences in direct conversation, but she was there as I discussed supply chain considerations for personalized medicine projects with my main host. As I headed to the train station that evening, I got a WeChat message from an Executive Director at ATLATL, a well-connected incubator in Shanghai that serves as somewhat of a “scientific embassy” for international biotechs looking to explore doing business in China. The friend of a friend had suggested we meet. The next day, I was in the ATLATL offices for lunch with their founder, Dr. PC Zhu 朱鹏程.

These experiences are reflective of the connectedness and sense of urgency I saw at every step during my week in China. Whether with CROs and CDMOs in Suzhou or deep elements of the reagent and instrumentation supply chain in Shenzhen, local ecosystems were dense and highly connected. Competition is fierce at every level — destructively so, according to many people I talked to. Quality is very high for those in the know. People are responsive and flexible.

Labor costs, medical costs, and infrastructure costs are lower compared to the US and Europe. There’s apparently a local discount, too — I heard from one company with operations in both the US and China that the Chinese operation’s quotes from local providers are half what gets quoted to American companies (this gives them an advantage in capital efficiency). But the most striking dynamic I observed was the speed. For companies that know how to navigate (read: have relationships, know whom to trust, and possess pre-built trust with those people), there’s a vibrant, redundant, end-to-end supply chain that can be tapped on demand with a high degree of responsiveness.

The ability to go from zero to patient data in 18 months is an advantage that will compound, as companies and the ecosystem writ large will be able to get to the real learning (testing drugs in patients) faster and iterate. Many of the companies I talked to were primarily (if not solely) funded by local capital markets and domestic government support. But the local Chinese pharmaceutical market is not enormous, with prices substantially lower than in Western markets and many patients paying for drugs out of pocket. I got the sense that the ecosystem sees preclinical development and clinical data generation as an important export market, with China serving as the innovation and proof of concept generator for medicines that will help patients around the world. Capital is already starting to flow to support this vision. At multiple stops, my visit was preceded or followed by VC firms and multinationals looking to feed and tap this innovation engine. This is good! Humanity needs more medicines — and right now, the path of least resistance to generate more medicines seems to lead through China.

A Note on Good Intentions

In Shanghai, I met the founders of a company working on pH-sensitive modulators of oxidative phosphorylation as therapies for chemo- and radiation-resistant cancers, based on work the Chief Science Officer (who is also a professor at a local university) had done during his postdoc in the US. They’ve focused on high-grade gliomas, and we met because they think their approach may be a good fit for osteosarcoma as well. They plan to open their first clinical trial in early summer in Shanghai and Beijing.

I mentioned that I am in touch with a glioblastoma patient who’s going ‘founder mode’ on his own cancer, and asked if it was possible for an international patient to access their medicine. The two founders of the company looked at each other a bit sheepishly and then shared their story. A friend of one of the founders, also a biochemistry professor, reached out to them on behalf of his mother. She had high-grade glioma, was in a coma, and was running out of options. The founders were hesitant, but the professor insisted they try. So they did, outside the auspices of any official trial. The protocol was 5 days of temozolomide, followed by 10 days of the experimental therapy. The mother is on her third dose of the drug. She emerged from her coma after the first dose. The company doesn’t have trial-grade evidence of the efficacy of their drug, but they do have more confidence. The professor-friend is grateful.

I share this story because I found in China what I find everywhere — people do biomedical research and drug development because they want to help patients. Scientists were friendly, helpful, and precise. People readily made introductions. They offered to help navigate further. They want to help patients.

I would personally feel comfortable flying to China for care. If a loved one or I got cancer, I would look at the IITs happening in China (with local help to try to get the full menu of what’s available). There’s a reasonable chance the most innovative option is being developed there.

Closing Thoughts for America

One of the smartest, most innovative scientists I spoke with on my trip noted that the process China currently follows is inspired by the one America once used for cell and gene therapy. As Dr. Ruxandra Teslo has very cogently laid out in her work on Clinical Trial Abundance, the early work on CAR-T’s in labs like Carl June’s at the University of Pennsylvania followed a playbook that seems similar to what’s happening in China now. Meanwhile, efforts to extend the efforts that led to the Baby KJ gene editing triumph seem to be running into stringent standards that are unrealistic for academic groups or narrowly scoped efforts to overcome. Dr. Teslo has put forward a number of specific policy proposals that would help America return to agility in early-stage clinical trials. I suspect we could lean more on a marketplace of reputation to keep clinical research in check in the US. The Clinical Trial Notification Pathway recently proposed by the FDA would be a good step in this direction.3 Some states, including New Hampshire, Montana, and elsewhere, are also moving in this direction; if not precluded federally, state-level innovation could serve as a laboratory of governance to test different versions of reform prior to wider-scale implementation.

It was interesting to observe that the expanded access/single-patient IND pathway is more suited to flexibly get an individual access to a potentially important treatment than anything I heard about in China. It was encouraging to see Dr. Marty Makary say recently on X that he’s signed every compassionate use request that’s crossed his desk. This is great! We should continue leaning in on single-patient INDs, with situation-appropriate standards that reflect the risk of inaction for a patient in dire straits.

How can America go faster? The default is to leverage the CRO/CDMO infrastructure that’s available in China to develop Western IP, which is clearly happening. What about parallel infrastructure? My mind goes to companies like Plasmidsaurus, Adaptyv, and Aequita, which are building highly automated, fast, pay-by-credit-card offerings for specific high-volume assays, and earlier-stage analogs in manufacturing like Nature’s Toolbox, Harton, and Exthymic. I’d love to know what else is out there.

It is ironic to me that the ‘marketplace of reputation’ that seems to govern China’s IIT ecosystem is more market-oriented than the regulatory apparatus we use to govern early-stage trials in the US. Every system has its strengths and drawbacks, China’s included. The parts I saw up close show how the Chinese ecosystem is leaning into its strengths — velocity of science and engineering, urgency, close-knit relationships within the ecosystem, compassion for patients. I’m hopeful that, as a country, we can reflect on and actively lean into our strengths as an ecosystem too.

For more, please visit sytse.com/cancer. Please write to us with thoughts or questions at cancer@sytse.com.

The Big Three memory makers — SK Hynix, Samsung, and Micron — have dedicated increasing capacity to memory for AI, or HBM. High-bandwidth memory (HBM) is a product that stacks multiple DRAM dies for AI memory. The increased allocation toward high-margin HBM means that not enough capacity is reserved for memory chips for consumer products. Thus, products like phones, laptops, gaming consoles, routers, tractors, and hospital equipment may experience price increases and shortages, perhaps as late as 2028. Adding memory capacity is a years-long operation, and in the meantime, the people will suffer.

As a result, there is murmuring amongst everyone, from the Pentagon to Apple and to individual gamers: perhaps the U.S. ought to turn to Chinese memory for consumer products. China’s leading DRAM company, CXMT, offers a compelling additional supply source. But the thought may scare conventional wisdom in D.C. Haven’t we been trying to decrease reliance on China? Why would we now open the floodgates on Chinese memory? In that case, perhaps the U.S. should instead ban or limit Chinese memory before the market creates unwanted dependencies.

Which is the right answer? Should Chinese memory be welcomed or restricted? This piece tries to answer the question by presenting both the case for and against Chinese memory. Ultimately, after balancing the impacts on the economy and national security, this piece believes that the U.S. should welcome Chinese memory — for products destined to the Chinese market. If customers can qualify CXMT for DRAM, then this would also lead to lower prices for American companies and consumers. The second-order benefits would be myriad, while the potential risks for market dependence and national security would be mitigated. Some risks, including assisting CXMT’s technological advances, are real but not sufficiently compelling.

The Case for Chinese Memory

Market Function

The most straightforward argument for allowing Chinese memory is to let the markets do what they will. Allowing Chinese DRAM from CXMT to compete with the Big Three will drive down prices for all. A naive calculation suggests that allowing CXMT unfettered access to American markets could increase global commodity DRAM supply by over 25%.1

However, the American markets will not be flooded with Chinese DRAM. First, CXMT’s capacity is already fully utilized by orders from Chinese customers like Xiaomi, Lenovo, and Alibaba Cloud. Although U.S. customers may be able to outbid other customers for limited capacity, this would likely be constrained in effect. Some Chinese customers have ongoing long-term contracts, and others would likely retain a preference for customer relations and governmental reasons. Thus, American customers would likely only be able to secure capacity for products destined for the Chinese market; for example, Apple is considering qualifying CXMT for iPhones only for Chinese consumers.

The real purpose of permitting CXMT is to offer bargaining power to customers in the immediate term. The advantage is not in securing orders, but in possessing the ability to secure orders. By qualifying CXMT DRAM, customers present a viable alternative and threat to the Big Three. The credibility of that threat is again uncertain, but it is likely credible enough for the Big Three to partially trim margins on commodity DRAM for customers.

The Big Three have moved away from fixed-price long-term agreements (LTAs) for DRAM and instead use post-settlement deals where suppliers can adjust the price after the orders have been delivered; this pricing structure benefits memory suppliers, but the inclusion of CXMT as a possible supplier could potentially promote a reversion to fixed-price LTAs or at least lessen the costs of post-settlement prices. This already seems to be the philosophy of leading PC makers and Apple. In this event, we would still be living through a shortage, but one that does not harm retail consumers as much.

The exact extent of price moderation in a world with CXMT memory is impossible to pin down — rough estimates must do. The extent would depend entirely on negotiated prices between customers and their memory suppliers, which would vary depending on the customer. The LTA Apple would get would be very different from the spot-price deal a small-time OEM would. Savings could also decrease if CXMT skyrockets DRAM price to align strategy with its market competitors, furthering the memory oligopoly. However, by adding more usable bits to the market, the price increases of memory in the coming months could decrease from anywhere from 5% to 15%.

Regardless of the exact number, these are real savings that pass on to the rest of the consumer economy. The RAM shortage is making the bill of materials for common products like smartphones and routers balloon, and allowing CXMT as a competitor will depressurize the market. Families needing laptops for school, offices needing PCs for workers, businesses needing cloud computing for operations — they all benefit in this world.

The persistence of the memory shortage also supports the need for alternatives outside of the Big Three (at least until H2 2027). Although everyone is currently spending heavily to expand capacity, fabs take years to come online. Further, as demonstrated below, the demand exceeds supply for HBM too. The capacity that the Big Three is building is for HBM, not for commodity DRAM. So while we wait for the Big Three to have the capacity and incentives to supply both HBM and commodity DRAM, CXMT can fill in the gap.

This situation is not hypothetical. Samsung’s planned memory expansions in its P4 fab and greenfield P5 fab are destined for HBM, not commodity DRAM, so such expansions will likely not alleviate the memory crunch. The story is similar for SK Hynix and Micron. Further, much of what “commodity” DRAM is manufactured by the Big Three may actually go toward AI applications, given server DDR5’s usage in the prefill phase for AI inference. By contrast, CXMT’s ramping production in its Shanghai megafab will be predominantly focused on commodity DRAM, not HBM or products for AI applications.

It is worth noting that the Big Three’s capacity allocation and expansion can play out in one of two ways: either they largely stick to their planned HBM roadmap, or they pivot to shift more allocation toward commodity DRAM. In the former situation, CXMT plays a helpful role moderating the market, pacifying the consumer economy until the Big Three have enough capacity in 2028. This scenario opens up greater risks of market dependency, which are explored in the case against Chinese memory below.

The latter scenario, while possible, is unlikely. Shifting allocation from HBM to commodity DRAM is not at all difficult; one just needs to swap out the masks in front-end fabrication, but all the equipment is the same. Shifting from normal DRAM to HBM is the more difficult transition, though, given HBM’s unique back-end processes. In this light, it makes sense that the Big Three’s expansions are all nominally targeted for HBM, as doing so gives them flexibility. However, shifting from HBM to commodity DRAM carries its own risks. By switching to commodity, fabs would effectively be losing money by underutilizing the tools that should have been used for HBM’s back-end processes. For semiconductor fabrication, where unit economics is king, downtime on tools is a cardinal sin.

However, some commodity DRAM products’ profit margins are superior to HBM3E, so perhaps more companies will dedicate more capacity as their HBM contracts are fulfilled. But anyone who says they know how the allocation will shake out is lying. Companies have some incentive to persist in HBM production even in the face of better commodity margins, as AI demand is more stable than the cyclical commodity market. Over the long-run, HBM yields better profit margins, even if DRAM booms cause the balance to shift temporarily. But perhaps a company will miss out on some HBM contracts or have process issues, making the allocation to commodity DRAM a better route. This is a dynamic process, involving variables ranging from the fate of the global economy, AI progress and developments, contract agreements, and individual business decisions. But if companies allocate more toward commodity DRAM than originally perceived, then the need for CXMT declines, circumventing potential concerns of market dependency.

After 2028, the crisis will likely have passed, and we can return to normal. At this point, the other fabs will have introduced enough capacity to render CXMT obsolete. Projects like SK Hynix’s Yongin megafab and Micron’s Boise and Tongluo fabs will be able to alleviate more of the demand. Further, commodity memory is notorious for being a glut-to-drought cyclical industry. By 2028, no one should be surprised if demand for commodity DRAM or even HBM dries up, causing a crash in prices instead of a continued surge. This is partially why memory makers are so reluctant to invest in commodity DRAM. (The emergence of LTAs and now so-called strategic customer agreements with five-year contracts is intended to lessen this risk, but we will see how much of an impact they have.)

Such cycles are dependent on consumer demand — a fickle variable tied to the global economy — and how important and memory-hungry AI will continue to be. The answer to the latter has been debated ad nauseam, and this piece largely follows Derek Thompson’s assessment of AI: nobody knows anything. Regardless, the odds are that by 2028, customers will not want to turn to CXMT for memory anymore.

Geopolitical Advantages

Another line of reasoning suggests that allowing Chinese memory into the American market may actually further our national security interests. By giving CXMT access to a lucrative market for their DRAM, they may be less incentivized to invest in HBM. After all, HBM would be a high-risk venture with certainly low yields (and thus, lower margins) in its early days.

CXMT is expected to dedicate 20% of its increasing capacity to producing HBM3 this year, but perhaps it can be incentivized to move away from the AI market. Already, commodity DDR5 margins are exceeding profits from HBM3E among the Big Three. Considering that HBM3E is already achieving mature yields, imagine the incredible profit comparison for CXMT’s DDR5 versus a pilot HBM3 it has yet to start. Rough estimates indicate that a TSV yield of near 60% is the inflection point where HBM becomes more profitable than commodity DRAM, and a yield of upwards of 70% is required for the margin percentage to be better. However, given some estimates that CXMT won’t break 40% until the end of 2026, CXMT seems to be a far cry from reaching that inflection point.2

Given the importance of AI to Chinese customers and governmental actors, it is ludicrous to think that CXMT will give up HBM altogether; after all, the company may be able to realize a better profit on HBM over the long run if it increases yield and finds a way to keep progressing (an uncertain prospect). This is the same logic the Big Three are currently following. However, in the short run, bales of cash may induce CXMT to temporarily prefer commodity DRAM over its HBM ambitions. CXMT is not exactly like other Chinese chip companies (like Innoscience) that can run large deficits without care for revenue. Building DRAM and HBM is expensive on orders of magnitude greater than that of compound semiconductors or mature-node chips. Capital matters, and CXMT knows it. Thus, instead of a 20% allocation to HBM, CXMT could be tempted to lower that number to something like 15%. That would be a win.

It would be truly a difficult decision for CXMT to make. CXMT DRAM would be a competitive product internationally, allowing the company to grow more rapidly and have greater market penetration. HBM, while deemed a critical product, would have no global market; the rest of the world is already on HBM4E, whereas CXMT is stuck two generations behind. CXMT’s HBM would be for domestic markets only, and CXMT would have to perform a balancing act between domestic mandates and international growth.

This argument is not certain, however, and prompts objections that CXMT earning more in commodity DRAM can actually support their HBM ambitions; more cash translates to more resources for HBM development and process perfection. This argument is explored in the succeeding sections.

Splitting capacity between American and Chinese customers also causes negative externalities for other parts of China’s AI industry, such as SMIC and consumer-facing companies. Every chip going to an American customer is one not going to a Chinese customer. China’s leading chipmaker SMIC has already announced that its own orders are lagging; because customers don’t think they can secure enough memory chips for a finished product, they don’t bother ordering with SMIC for the logic chip. Further capacity allocated to American industry exacerbates this trend for the Chinese industry. If one believes that the U.S. buying CXMT DRAM supports their HBM ambitions, then by this logic, it would also hurt SMIC’s advanced node ambitions. With fewer orders, they would have fewer resources to develop past 5 nm.

Although a slowdown in the Chinese economy is not inherently an advantage for the U.S., the fewer dollars dedicated to SMIC’s advanced-node developments and Huawei’s AI processors are in America’s interest. Of course, it is again unclear how much capacity American customers would receive, and the Chinese government would certainly clamp down on attempts to leave Chinese companies empty-handed while American companies receive whatever they want. However, the allocation would likely be greater than zero, and the increased tension between the company and the government can only serve American interests.

The Case Against Chinese Memory

Market Dependency

The leading argument against allowing unfettered Chinese memory is predicated on real concerns of market dependency. The U.S. has taken pains to reduce its economic dependence on China for critical industries like rare earths, semiconductors, telecommunication infrastructure, etc. Why would we now allow that dependence to again fester in the form of memory chips?

Even if CXMT does not dominate the American market in the beginning, the company’s foothold in the American market has the potential to skyrocket. No one can serve demand right now, and everyone is attempting to expand capacity. As demonstrated in the previous sections, CXMT will be the first to significantly expand capacity for commodity DRAM. Could this not lead to a long-term, increasing dependence on CXMT? Although the immediate term may lead to helpful bargaining power without real allocation, the future may lead to real allocation that causes genuine entanglement.

It is possible that the Big Three become increasingly “HBM-first” companies, allowing CXMT (and later, YMTC) to take up a bigger share of the commodity market. The trends in wafer allocation could support this claim. The revenue that CXMT generates from this increased market share could be reinvested into R&D, capacity expansion, and even advancement in HBM.

However, it is highly unlikely that Chinese memory companies will play a role larger than end uses constrained to low-performance applications and/or in foreign markets. First, the Big Three will always make commodity DRAM. Large-scale production of DRAM dies helps the companies improve their yield for newer nodes, which are to be used later for HBM. The commodity DRAM is always the first step of the HBM process; they cannot be separated.

Second, customers will always want commodity DRAM from the Big Three, such that the economics will always tilt toward the Big Three maintaining some amount of commodity DRAM production. The Big Three’s DRAM nodes and performance are leagues ahead of CXMT’s. The Big Three are perfecting their 1c/1γ nodes while CXMT is still on 1y/1z, at least three generations behind. Even significantly cheaper CXMT DRAM is not so attractive given the use cases for memory in consumer products. Apple does not release a new generation iPhone with worse memory, even if it is much cheaper. The same goes for XBOXs and PCs; while some focus on the lower-cost market, the bifurcation of markets for low-cost and high-cost products can only serve consumer interest.

Betting on CXMT not to catch up or plateau is not a bet against Chinese innovation (a poor bet indeed), but rather a bet that export controls on EUV lithography and equipment required for DRAM advancements are effective. China’s domestic EUV capabilities will likely not be realized until 2030 at the earliest, and restrictions on EUV lithography have been an enduring American policy throughout both Republican and Democratic administrations. The $400 million machines are colossal and monopolized by ASML, meaning smuggling is not as serious an issue as it may be for individual chips.

Export controls condemn CXMT to only applications that do not require cutting-edge memory. That is an important market segment, but nothing near an impending monopoly or concerning supply chain risk. And even in these segments, other companies like Taiwan’s Winbond and Nanya will have room to compete and prevent a Chinese monopoly.

Lastly, this world would cause CXMT to constantly be tugged away from allocating more capacity toward HBM. Although revenue generated from the commodity segment may help CXMT build more and research better, they will be faced with tough decisions in wafer allocation. The market disincentivizes companies from building enough capacity to perfectly satisfy both commodity and HBM demand, as no one wants to be left holding the bag on a $20 billion fab once the cycle declines or if the AI bubble bursts.

Cautious expansion is the philosophy for everyone in the chipmaking space, for reasons well explained by Asianometry with respect to TSMC, the bullwhip effect, and the beer game. In brief, small variations in demand from retail consumers or AI players cause the greatest volatility for the suppliers at the end of the chain. If everyone in the world starts buying one more candy bar from the gas station, the gas stations feel it slightly, but the candy bar factory gets slammed the most with orders. If everyone starts buying one fewer candy bar, then the gas station barely feels it, but the candy bar factory can go broke. When the memory industry inevitably experiences a demand downturn — no matter how small — the memory makers will suffer the brunt of the fallout.3

Geopolitical Disadvantages

The stronger argument against Chinese memory is a geopolitical one. Every dollar going to CXMT and YMTC, regardless of how it benefits the American economy, would also be benefiting companies widely considered national security risks.

Although American policymakers have a tendency to think that Chinese companies have open checkbooks from the Chinese government, they need a great deal of supplemental funding to support their ambitions. CXMT’s recent IPO (and YMTC’s impending one) demonstrates the need for billions of dollars more capital to fund capacity expansions and R&D. American companies giving CXMT money for DRAM, thus, is America funding China’s HBM ambitions.

One of the Big Three’s biggest advantages currently is their ability to spend on capex in a way that CXMT cannot. These dollar amounts go toward node migration and capacity expansion — the reasons we’re ahead right now. Perhaps allowing CXMT to proliferate in the market will reverse this advantage.

Another concern is the reality that American customers would be helping to perfect CXMT’s processes by qualifying them as a supplier. For example, if Apple desires to qualify CXMT for its LPDDR5X in iPhones, then Apple will work with CXMT to make its processes more reliable and better performing. Apple engineers would literally assist CXMT’s products to outperform the JEDEC standard and meet rigorous requirements for metrics like thermal performance and consistency. Do we want American engineers helping Chinese companies in this way? It’s a hard pill to swallow. These technological advancements directly translate into CXMT building better HBM for AI demand.

And once qualified, ecosystem stickiness poses a problem. Even if the Big Three have capacity available once again, companies will have already gone through the trouble of qualifying CXMT as a supplier. Why not stick with them as a significant supplier, specifically for low-cost applications or in markets where price matters more than performance? How this plays out is again impossible to predict, but believing that CXMT will remain a major player beyond this memory crisis is a real possibility.

Aqib’s Verdict

Ultimately, the risks associated with permitting CXMT market access are grounded in more exaggerated doomsday scenarios rather than rigorous analysis. Giving CXMT money and qualifying it sounds scary, but the downsides seem to pale in comparison to the benefits. We shouldn’t care more about making sure China stays inside a box than the welfare of American citizens.

The fear of CXMT represents a prevailing American paranoia of anything associated with the five-star red flag. China is an adversary, but each decision should be predicated on rigorous cost-benefit analysis, not blanket anathema. Buying from CXMT will certainly help them in some way, with increased funding and a level of technological progress, but this is a far cry from China catching up or posing a threat.

First, CXMT’s progress will be more stymied by American export controls than benefited by customer revenue. Cracking future nodes of DRAM is more a technical problem than a financial one. Second, the technological benefits from being qualified by American customers should not be overstated. CXMT is already qualified by major players like Chinese smartphone, PC, and cloud computing companies. These companies already push CXMT to progress beyond industry-required minimums. An Apple partnership will perhaps move the needle a bit, but it is not like the U.S. would be helping them discover fire.

The ecosystem stickiness argument is the most defensible, but this piece does not weigh it as heavily compared to the myriad benefits. The Big Three produce better memory compared to CXMT on a performance basis, mainly due to the yield superiority gained from technological advancement and export controls. By 2028, just like 2026, CXMT will not be in the same league as the Big Three in terms of memory quality. Without advanced tooling, they cannot reach yields or performance specifications like the Big Three can.

The semiconductor industry is unlike electric vehicles or solar panels. Although China may offer cheaper products, the risk of market dependence is not so serious. Export controls and the difficult science of semiconductor manufacturing indicate that CXMT will be behind for years. The current market crunch is not a permanent state of affairs by any evaluation, but rather a temporary pain that requires a temporary solution. The players in the industry are also not engaged in a race to the bottom that China will win, but rather a race to the next node that China will lose.

The options are also not binary. The U.S. can permit CXMT now, when the benefits are attributed more toward bargaining power than to actual capacity allocation, and slam the door later. Other policy options to reap the benefits while managing the downsides exist, including greater tariff impositions, requirements on customer allocation ratios, etc.

Permitting now but slamming the door on Chinese memory later can also have added benefits. If CXMT expands capacity to satisfy American demand now, being shut out later would leave the company holding the debt of underutilized fabs and equipment. Of course, this is a severe oversimplification, and CXMT is not this dumb, but the regulatory uncertainty adds a layer of benefit.

Perhaps the best means of managing the benefits and downsides of permitting Chinese memory is allowing it in limited contexts. If the U.S. permits American companies to qualify CXMT solely for products destined for the Chinese market, then the scope of the “exposed” market is narrowed. For example, only iPhones for the Chinese market could contain Chinese memory, and the overall savings may be distributed throughout the market worldwide. However, this policy option, in my view, is worse than the aforementioned ones.

Banning CXMT is the least defensible policy position right now. The memory crunch is here, but it is not here to stay. Let’s allow the market to do what it will in the interests of our own people. Seeing ghosts in national security threats here discredits real national security threats elsewhere, so let informed policy reign, and let the DRAM flow.

To discuss, we’re joined by Cat Buchatskiy, Director of Analytics at Snake Island, a military analytical group, along with Chris Miller

Our conversation covers:

• How battlefield pressure forced Ukraine to build a drone war machine from scratch — from a handful of soldiers flying off-the-shelf drones to domestic assembly at a massive scale.

• Ukraine’s industrial legacy and how whole-of-society mobilization repurposed its civilian tech sector into a wartime industrial base.

• Why modular design, frontline reassembly, and tight feedback loops allow Ukraine to iterate faster than traditional defense systems.

• The constraints of global supply chains, the impact of export controls, and how China is playing both sides of the war.

Listen now on your favorite podcast app.

Building a Drone Industry From Scratch

Jordan Schneider: Let’s start off with a very brief overview, Cat, of the accomplishments of the Ukrainian drone industrial base. What is getting pumped out on a monthly basis at the start of the war?

Cat Buchatskiy: We’re right around the time where it’s almost exactly two years of full-scale war, and the drone industrial base has been completely transformed at a pace that we really haven’t seen in basically any other country. Necessity is the mother of invention.

In February of 2022, we had about 3,000 drones total being produced in Ukraine — FPV (First-Person View), UGV (Unmanned Ground Vehicles), sea drone, anything of the sort. Ninety-nine percent of them were imported as entire systems from China.

In 2026, we basically have 99% being assembled in Ukraine. Now, just the FPV industry alone is cited to be able to produce up to 5 million FPV drones per year. That doesn’t include our massive industry of heavy bomber drones, ISR (Intelligence, Surveillance, and Reconnaissance), loitering munitions, or UGVs, which is now a booming industry in Ukraine as well.

But the most impressive thing isn’t necessarily just those numbers — we went from about 3,000 systems being made in February ’22 to 4 million FPVs alone. It’s the actual localization of that final assembly and the way that Ukraine has been able to completely transform its drone manufacturing industry. Now we’re at a point where 99% of the systems are final assembly in Ukraine with a lot of components being imported, but basically no final systems being imported from China anymore, which is a massive accomplishment.

Jordan Schneider: Let’s turn the clock back to understand the arc of industry, military, and government that made that possible. It’s February 2022 and war has broken out. If you wanted to fly a drone for Ukraine, how would you procure one?

Cat Buchatskiy: February of 2022, we actually hadn’t really seen the introduction of drones on the battlefield quite yet. If you wanted to fly a drone for Ukraine, first of all, you’d have to convince the Ukrainian military that it was even something worth exploring because it took us a little bit. It was not really until summer of 2023 that we saw the massive boom in FPV use and started seeing the pioneering of that industry.

From 2014 to 2022, there were sporadic incidents of Ukrainians toying around with Mavics. But in February 2022, if you wanted to fly a drone in Ukraine, you would probably have to bring in a DJI Mavic and get it to the front line yourself. We had essentially no homegrown drone industry whatsoever, and it was used in very small batches, basically all imported by volunteers that were buying them up in bulk from DJI.

Jordan Schneider: That really is a fascinating moment in retrospect. It wasn’t necessarily that the technology didn’t exist, but that folks on both sides of the war hadn’t quite clocked into the potential of drones. Cat, can you take us back to all the light bulbs going on that this stuff actually had an enormous amount of utility?

Cat Buchatskiy: By 2023, when we were thinking about potentially another wave of counteroffensive, Ukrainian soldiers and volunteer networks started buying up more and more drones initially to perform ISR functions. Then pretty quickly — if you’re a soldier fighting in an existential battle, you’re going to do absolutely everything you can with the tools at your disposal. People realized that they could strap explosives onto these things and just fly them directly into the enemy, which was huge.

What was huge here for Ukraine was the asymmetry of using drones. First, because we were strapped for cash. The cost asymmetry of being able to put a payload and ammunition onto a drone that would only cost you a few thousand dollars was huge for a country that’s at economic disadvantage and fighting against Russia, which has one of the largest military industrial complexes and military budgets in the world.

Second, the asymmetry of being able to protect our soldiers and pilot these drones remotely was also huge because you’re never really going to be able to go person for person with the Russian army. They’re always going to have more people. In a war of attrition, which we pretty quickly realized it was going to be, we weren’t going to be able to hold the line with as many infantry and as many soldiers as Russia would. Being able to send remotely controlled tools to perform certain functions instead of putting human life at risk was another huge benefit for the Ukrainian side.

Pretty quickly we went into overdrive to produce these drones just by the sheer necessity of not having as much money to buy different systems and not wanting to put our people’s lives at risk. From 2023 to now, it was just a huge industry boom. We got to where we are today because we realized that was one of the major things keeping us in the fight — our ability to leverage unmanned systems as opposed to putting our capital and our people’s lives at risk.

Industrial Mobilization

Chris Miller: Could we dig into the industrial dynamics of that? Ukraine is fascinating in that it started the war with this pretty sizable defense industrial base — Motor Sich and others, legacy firms with really substantial capabilities. Then we’ve seen, as I understand, the last couple of years have brought this extraordinary boom in new firms emerging, first sourcing then producing the drones that Ukraine requires. Tell us about how the defense industry in Ukraine has shifted and transformed over the past couple of years.

Cat Buchatskiy: Ukraine had always had a rich industrial background, particularly during the times of the Soviet Union. It was a massive producer of jet engines. Dnipro was a huge rocket factory in Ukraine, and we had that background in place.

The issue was that during the time of the Soviet Union, one of our major contractors was the Russian Federation. After the collapse, without that centralized funding from the USSR, the Ukrainian government was responsible for upkeep of those manufacturing facilities and for continuing those production lines. It struggled tremendously with that because our state budget was strapped.

We quickly realized that we were going to be in a tough spot if we continued to invest in our defense industry, because Russia was not going to be happy about that. It was clear in the 2000s, after Ukraine had gained its independence, that Russia was going to put pressure on Ukraine to do everything it could to cripple its defense industrial capacity. After 2014, it became a major point of contention.

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.

It was similar to what happened in the US during World War II, where you tapped into this massive civilian talent and massive civilian production lines and directed them to contribute to the war effort. The tapping in of the civilian industries, which was supported in large part by our government and its state policies to encourage more companies to direct their efforts into defense, was what kept us afloat.

Now you have millions and millions of dollars going into this. You have dozens and dozens of different companies. I saw a stat recently — we’re releasing a report about it this week, actually — that there are over 40 component manufacturers in Ukraine, and that’s just components alone. I can’t speak to their effectiveness or the scale they’re producing at, but that’s a massive number for a country of about 30 million people in active war. The fact that there are dozens and dozens of these UAV companies for every single category speaks volumes to the amount we’ve been able to mobilize all of society and have this defense tech renaissance.

Chris Miller: This is really fascinating. In the last ChinaTalk episode I was able to join, we were discussing — in a worst-case scenario, would you rather have a Xiaomi or an Apple? In other words, a hardware company or essentially a software/supply chain company?

What you’re describing with Ukraine is not a labor force that was skilled in drone manufacturing, but a labor force that had a lot of software know-how and was able to quickly repurpose itself to build a bunch of essentially brand new drone companies in a way that I probably wouldn’t have predicted would have been nearly as successful as it actually has been.

Can you tell us about the typical entrepreneur who started up a new drone company or runs a production line? Who are these people and what makes them good at their jobs?

Cat Buchatskiy: They come from completely different backgrounds, which is super interesting. You have some people who, in their past lives, used to be top software engineers at B2B SaaS companies. You have one of the biggest defense tech VCs right now supporting the entire industry, who used to be the chief marketing officer at a workflow automation company. Some people weren’t in tech at all and became CEOs, stepping into it from working at video game companies.

The video game overlap is actually quite real — that pipeline exists. I actually used to play a lot of video games and learned drone operating from that. Many of them were working across the industry at places like Uber Ukraine or other rideshare companies. There are a few examples of that.

It became unimaginable for most people in Ukraine after February 24th to work on anything except this. It’s something that you really can’t replicate unless your country is at war — and not only at war, but in an existential one. It’s extremely difficult for any other country to imagine.

When I talk to my friends abroad about the fact that most people I know now in Ukraine follow this pipeline — you’re in high school, you’re in college, you want to do something for the army. You either want to join the army, you’re going to work on loitering munitions, or you’re going to do something else. It’s a type of society-wide mobilization that’s very hard to imagine and is only comparable to maybe Israel.

Chris Miller: I love the dual-use chief marketing officer. I wouldn’t have guessed that, but that’s an extraordinary anecdote.

The other interesting dynamic is that you had this assembly spring up very rapidly by comparison to almost anything else I could imagine. Can you walk us through what a typical drone assembly factory looks like circa 2026? We’ll probably spend some time later digging into the components, but when you get the components together, what does one of these factories look like? What is their scale? What’s the time to assemble a specific drone? What do the economics look like?

Cat Buchatskiy: I want to debunk a myth here that I see a lot in online articles and conversations about Ukraine’s supposedly “garage shop” drone industry. It’s not that. While it might have been in 2022 or early 2023, and while we still have many small startups operating that way, for the most part, these are massive manufacturing facilities.

We’re talking about warehouses spanning hundreds of thousands of square feet, multiple stories high, with many completely underground to protect against Russian strikes. You have these massive underground bunkers with production lines and hundreds of employees assembling drones by the thousands per month. One particular drone factory produces 10,000 units monthly.

While it’s not quite at DJI’s level yet — the drones don’t fly themselves to the next assembly station like in those videos — we’re getting there. It’s comparable to what you’d see at SpaceX’s Starlink production line. This is not a garage shop industry. Our prime contractors (though we don’t have traditional primes) are producing tens of thousands of units per month and generating tens of millions of dollars in revenue.

The production looks like any other assembly line — carbon fiber frames come off the line, then workers attach the motors, props, and other components before shipping. What’s interesting is our reverse cycle approach. Many manufacturers send completed systems to the front line, where units have their own assembly facilities. They actually disassemble the drones and reassemble them according to their specific battlefield needs.

The process from parts to combat-ready drone involves assembling components in the shop, shipping to the unit, where they’re disassembled and reassembled in their own production lines before deployment. This is something many Western countries don’t comprehend — it’s almost impossible to ship a finished system that flies straight out of the box. R&D shops and assembly lines operate across locations closer to the front lines, run by the military doing their own assembly work.

Jordan Schneider: This reminds me of a past ChinaTalk episode we had with Christian Brose from Anduril a few months ago. I opened by asking him why he was building his new factory in Ohio — a state without mountains to hide factories in. It was somewhat facetious, but hearing that all Ukrainian drone manufacturing happens underground makes me worry that American drone manufacturing is just happening in open fields somewhere.

Chris Miller: If I wanted to start up a factory producing 10,000 FPVs a month, how long would that take to get up and running? What does it cost to build a factory like that? Are there specific types of machines or tools that are hard to access, or is that not really a challenge?

Cat Buchatskiy: I can’t speak to the exact cost, but in terms of estimated time to set it up, our biggest FPV primes took 2 to 3 years to get things up and running. By 2025, they were already hitting those numbers, which means it was about 3 years from the beginning of the full-scale invasion to scale production.

They achieved this without venture capital funding, which is interesting because we didn’t really have VC in Ukraine at all. It was purely state contracts and bootstrapped companies. One of the big companies that recently announced a huge joint venture with a German firm was completely bootstrapped. In 2022 and 2023, they were basically producing no systems out of their garage and almost went broke. Then they managed to pull it together and are now one of the biggest Ukrainian drone primes that just opened a factory in Germany. Their turnaround took about 2 to 3 years.

Regarding machines that are difficult to access, CNC machines are a challenge. But honestly, it’s less about the machines themselves and more about the fact that logistics are super challenging for Ukrainian companies. We’re dealing with export restrictions from China, so the actual components we’re trying to access are severely bottlenecked.

Additionally, getting logistics to a wartime country is difficult. We don’t accept direct deliveries and you can’t fly in. You have to load trucks in Poland and take them across the border, or from any other country, dealing with customs and security clearances. It’s a significant lift and heavy operation to establish. That’s why most of our companies are in Western Ukraine — it shortens the logistics cycles considerably.

These are things people don’t really think about when they consider Ukraine scaling its defense industrial base. The fact that we are in wartime, isolated from traditional shipping routes, makes our achievements all the more impressive.

Jordan Schneider: You have this line in one of your reports from late last year that “Ukrainian startups can assemble and ruggedize, but they cannot easily reproduce decades of specialized chemical material or electronic expertise.” Before we get to the second half of that statement, let’s explore the assembly and ruggedization aspect. What has that unlocked for Ukraine? Why was it important to have that domestic capacity developed in the first place?

Cat Buchatskiy: The ruggedization was crucial because Ukraine is fighting a war of attrition. Modularity is incredibly important, which is why in-house assembly matters so much.

As I mentioned earlier, when systems get built in the factory and sent to military R&D labs, they essentially disassemble and reassemble them. The reason is that manufacturers can’t predict which features the frontline will need by the time products ship out.

VTX systems are almost always switched out — it’s completely unpredictable what frequencies you’ll need to fly on. That’s never flown straight off the factory line. Soldiers always switch it out themselves. There’s also modularity in terms of payload attachments, munition requirements for specific missions, and flight conditions.

Having the actual assembly capability allows you to disassemble and reassemble according to specific mission requirements. This is huge because one downside of the massive Chinese drone imports into Ukraine is that you don’t have the same relationship with manufacturers who can react and respond to real user needs.

This is also why most Western companies fail in Ukraine — the headlines you see reflect this. They don’t have an engineering presence or assembly presence to react fast enough to soldiers’ needs. When soldiers report getting jammed and needing to swap out VTX systems, or when they need specific payloads for particular missions, these companies can’t do anything about it. They can’t phone California and ask them to fix something if they don’t have an assembly line in Ukraine to react.

This gives us tremendous mobility and leverage because we can adapt the tools to whatever mission we’re conducting. When people talk about short innovation cycles in Ukraine, this is mostly what they mean — the ability to have continuous R&D and for soldiers to get hands-on and adapt modular systems to their needs. We’ve nailed this down, and it’s been incredibly important.

China’s Calculated Neutrality

Jordan Schneider: Nice drone industry you have there — shame if some export controls were to happen to it. I want to read in full this opener from a Financial Times piece from about a month ago, which featured one of your interviewees —

“On his numerous visits to the factories of southern China, Oleksandr Yakovenko finds that his hosts increasingly plan his arrivals and departures down to minutes and seconds. They sometimes ask him to wait nearby for a while or usher him through side doors, down service corridors, or into empty conference rooms. It took the founder of TAF Industries, now one of Ukraine’s biggest drone producers, a while to realize why his arrival at the head office of a camera developer or battery maker required such opaque rituals of schedule juggling and extreme punctuality. It was because the Russians had just been there, or they were on their way, or both.

‘Our suppliers make an effort to manage the Ukrainian and Russian customers. They try to make it so we don’t have to be in the factory at the same time,’ he told the Financial Times. ‘They invite us for one time slot, but they invite the Russians for a different time. As soon as the car with the Russians drives away, the car with the Ukrainians goes in,’ he adds.”

What an unbelievable situation we’re in. It’s truly surreal. There have been other times in history where arms manufacturers sold to both sides — actually, the more I think about it, it’s not that uncommon. But the fact that we’re having this iterative technological race, as opposed to just selling some AKs to this side and some AKs to that side with a shrug, is really weird.

Cat, can you start by telling the story from the Russian side as well? How do both sides of this war have significant drone dependencies on what comes out of factories in China?

Cat Buchatskiy: It’s definitely a very bizarre scenario, especially for our manufacturers dealing with this. Both sides have a dependency because most of the critical components for the drone industry are based in China. The whole world really has this dependency. For both sides to produce the unmanned systems we need at the scale we’re going through them, it’s impossible to do without China.

Jordan Schneider: What are some historical examples? There’s this example of the British selling dreadnoughts to the Argentines, Brazilians, and Chileans all at the same time.

Chris Miller: The dynamic here isn’t just selling finished weapons — it’s this supply chain iteration. That does seem pretty unique.

The defense supply chains today look very different from those of 50 years ago. Rather than all production occurring within one country, we now have actual multinational supply chains.

There’s a book called Producing Security that discusses how defense industrial supply chains have become multinational. It’s no longer a simple matter of one country producing battleships while another produces theirs — you can’t manufacture your battleships without components from their country. We have Chinese rare earths in our fighter jets and Chinese components in our drones, while they have US AI chips in their supercomputers simulating weapons explosions. This represents a much deeper integration in defense industries than we’ve ever seen before.

Cat Buchatskiy: It’s interesting because Russia and Ukraine have two very different relationships with China when it comes to manufacturing dependency. China will absolutely sell to both sides, which is actually one of the reasons why several of our manufacturers aren’t particularly concerned about China and their supply lines.

When you ask them whether they’re worried about China, quite a few will tell you no, because it’s big business for them. The Chinese are primarily businessmen. Although China is definitely trying to tip the scales in various ways, there’s a belief among some manufacturers that China will never completely cut off Ukraine. They don’t see it as a real threat because China can’t afford to lose that business — no one else other than Russia is buying at that scale. About 30 to 40% of Ukrainian manufacturers really don’t see China as a threat, which is interesting.

One of the key differences between Russian and Ukrainian manufacturers’ relationships with China is that the Russian manufacturers are integrated at a level where they’re localizing those supply chains within Russia and moving them into special economic zones. China is allowing them to buy entire assembly lines and relocate them, which is scary because it means Russia is slowly developing a domestic industry.

They’ve already done this with the Iranian Shaheds, localizing production in the Alabuga Special Economic Zone. Sometimes, there will be assembly lines in China working on a particular component that Ukrainian manufacturers don’t have access to for a year at a time. Then they’ll come back saying, “No worries, we just finished — the Russians partially relocated it to Russia, and now we can accept orders again for what we have left.” This quote appears either in the Financial Times article or in our report.

Ukraine doesn’t have access to replicate that type of relationship. We can’t bring those assembly lines over, partially because we don’t share a border with China, and secondly, because we don’t have that special economic relationship.

This situation is going to compound over time. I’m very worried that in a few years, the Russian industrial base might be significantly less dependent on China for certain manufacturing know-how and capabilities they’re able to localize. They’re going to be able to outpace us — they already mostly outpace us in most things. It’s genuinely concerning that this is spreading beyond China and that these industrial capacities are localizing into other parts of the axis of evil. That special relationship stands out as very different from what Ukrainian manufacturers have with China.

Jordan Schneider: I want to stay on this very interesting series of decisions that Chinese central policymakers have made over the course of this war. Famously, Putin postponed the invasion of Ukraine until after the Beijing Olympics, and right before then, they declared they’re best buds with the best strategic partnership ever.

As Cat walked us through, for the first year or so of the war, drones were not the central dynamic. Everyone was talking artillery, artillery, artillery. There was still focus on infantry manpower, tanks, and fighter jets. During that time period, it would make sense that this wasn’t necessarily something front and center for Chinese policymakers. If a few drone manufacturers could make an extra 8 figures here or there, great, why not?

But as drones increasingly become critical to Ukraine’s ability to defend itself and Russia’s ability to inflict pain both on the front line and deeper into Ukraine, you have a slow creep of export controls. Clearly, this effort has not been pushed to the extreme that it would be if China were directly in conflict.

We famously have some quotes from Wang Yi from this past summer saying that if China wanted to end this war the next day by fully supporting Russia, it would. But he also said that Beijing didn’t want Russia to lose the war in Ukraine because it thought the US might then shift its focus to China and Asia.

My conspiracy question here for you — do they also not want the Ukrainians to lose? Selling drone parts to Ukraine is not a central pillar of the Chinese economy. There has to be some larger strategic calculus going on to allow this number of parts to continue to flow to the Ukrainian drone base.

Cat Buchatskiy: I don’t think that China wants Ukraine to lose. The reason being that I don’t think China and Russia are real friends, and I don’t think China minds depleting Russia’s arsenal. China doesn’t really need Russia for the most part. While there’s a lot more economic interaction with Russia now, in terms of defense and the role that Russia plays in the world, China sees it as a defeated, has-been power.

China understands that, frankly, the US also doesn’t see Russia as its greatest threat. Read the recent national security strategy — it’s barely in there. The US is all focused on China.

I don’t think Xi really cares if Ukraine is able to continue to attrit the Russian defense industry. For them, playing both sides is a win-win scenario because they keep their biggest ally dependent on them. The Chinese defense industry is going to be stronger than the Russian defense industry, and Russia is going to continue to need to buy parts from China.

Frankly, they’re exacerbating the divide. If you’re thinking about great power politics, China’s only getting stronger, Russia’s only getting weaker, and it’s not going to be a tripolar world between Russia, the US, and China. They’re going to want to make it a bipolar world, and Russia is going to be dragged into that orbit as long as Ukraine continues to weaken its global position and sanctions continue to be held.

This might be a bit of a hot take in Ukraine as well, but the fact that China did put in place export controls to both sides — technically export controls to Russia as well — and the fact that China was not extremely explicit in its support for Russia in the war and has mostly maintained an outward political neutrality means that China doesn’t really care about what happens to Russia that much. They particularly care about their standing with the United States and how they compete there.

I don’t think that China wants Ukraine to lose. It’s in their interest to keep it going.

Export Controls

Chris Miller: Could you talk about the export controls in detail? On one hand, you’ve had these controls put in place over the last couple of years, yet Ukraine’s drone industry has grown by every measure. However, your research shows that doesn’t mean controls haven’t mattered. In fact, they’ve been disruptive in several ways, and there’s been a rerouting of components. Walk us through what has actually been the impact of those restrictions.

Cat Buchatskiy: The biggest impact of the restrictions is the logistics challenge. We have difficulty with a lot of suppliers buying directly and shipping directly from China, especially when buying at scale. We’ve been forced to reroute — we can’t have mass DJI Mavic procurement from the government. We have to rely on volunteer networks.

We have a Mavic shortage on the front line right now, and we’re taking a big hit because of it. It’s widely talked about in Ukraine that we don’t have enough Mavics, and our entire ISR ecosystem is basically dependent on having a continuous, steady flow of Mavics flying on the front line. The fact that we cannot procure them at the government level at scale is quite a challenge, forcing us to rely on these spread-out volunteer networks.

Second, it can sometimes be like playing a game of whack-a-mole. We have shell companies selling to Ukraine, but once someone finds out on the other side that they’re selling to Ukraine, they have to shut down and reopen under a different name somewhere else. Russia can play this game as well, putting in reports claiming that a company is selling to Ukraine through a long chain of command. In certain high-tension political situations, this has resulted in companies no longer being able to sell to Ukrainian manufacturers.

I want to be wary of overstating this — Russia doesn’t pull that many strings. China is very much in control of this. Russia doesn’t have as much leverage as it would like to have. But our inability to work directly, particularly on the systems front, makes DJI Mavic procurement very difficult for us. We feel a significant impact when we can’t get certain components fast enough to overcome bottlenecks. This leads to less coverage and less visibility on the front line, which results in less successful strikes. It also means more deaths on our side. The impact is very tangible when we have these problems.

Chris Miller: It’s interesting to think through this situation. The US and Europe have export controls on the transfer of technology to the Russian military, and China’s got them on Ukraine’s drone industry. I wonder whose controls are more impactful. There’s plenty of leakage in both.

Subscribe now

Cat Buchatskiy: The US and Europe have export controls to Ukraine as well. That’s something that isn’t really as well spoken about. Our manufacturers have had trouble procuring from European component suppliers because they don’t let the systems go to Ukraine.

This is another aspect of the Russia-China partnership that’s interesting. There’s a tendency in the West to worry about how much China can control Ukraine’s drone industry and how much leverage China has, without realizing that Europe could probably help Ukraine a lot more if it didn’t have certain export controls. We could get European components and decouple from China by procuring from the West, but we can’t because of ITAR and European export controls.

I don’t know if this anecdote is in the report, but we had a situation where one of our drone companies wanted to procure a certain component from the French. The French refused to let them have it because it had some sort of super-classified super glue on the component. The knife cuts both ways.

Chris Miller: Cat, you’ve helped us understand how Ukraine now assembles millions of drones per year. But as you’ve noted, while some of those components are sourced locally, many are not. Walk us through how the component supply chain has evolved over the last few years.

Cat Buchatskiy: The component supply chain has grown significantly. If in 2022, we were importing 99% of finished systems, this trend has completely reversed. As of 2025 — the last data we received — we’re now importing 99% of components instead. We’re talking about tens of millions of dollars in imports from China per year for these components alone.

The supply chain has also evolved significantly in another way. China has recognized that this is becoming a massive industry in Ukraine and that domestic production and vertical integration of unmanned systems manufacturing is now a massive priority for us. In response, China has begun manipulating policymaking to affect our supply, not only through export controls.

One of their tactics is to affect pricing, making it significantly more attractive for us to buy final components rather than the individual pieces needed to build components from scratch. They want to prevent us from learning how to source specific chemicals and parts needed for ground-up manufacturing.

Currently, the Ukrainian components industry has become quite proficient at final assembly of motors, battery packs, and similar items. We’re now focusing on mining rare earth minerals and achieving even greater vertical integration.

At the end of 2025, the Ukrainian government launched a major initiative. BRAVE-1, a mechanism by the Ukrainian Ministry of Digital Transformation, has been pivotal in stimulating the Ukrainian drone industry. They finally turned their attention to components, providing various grants and funding incentives for manufacturers to start producing components in-house.

When we pursued this direction, we quickly discovered that the price difference between buying a finished motor from China versus buying just the magnets needed to make the motor is significant. This makes establishing that supply line even more challenging. There are also many more export controls on raw materials.

The trend for the Ukrainian industry this year will be a continuation of this pattern. The Ukrainian government is pouring significant money into securing raw material supplies. I’m hoping that when we chat again in six months, I’ll be able to report that we’ve invested so much money and implemented policies that we now have actual rare earth mining and raw materials production, with help from the United States. However, I’m not yet sure what that will look like.

The main changes in the supply line are the massive intake of individual components that we simply weren’t seeing in 2022 or 2023. This has become a critical backbone of the industry, especially as it’s become more decentralized. Ukrainian R&D shops now need to source their own components, do their own assembly, and work on drone modularity.

It’s a hugely decentralized industry, which is actually part of why we’re able to navigate around export controls. With so many different buyers rather than one centralized node, people find ways to work around restrictions.

Chris Miller: We could start with the motors and dig into a couple of key components. You mentioned it’s possible to import a finished motor, and it’s also possible to try to import the parts to the motor. Pretty interesting that the parts cost more than the motor itself, if I understood you correctly, which is the opposite of what you’d expect. What’s the hardest thing to make inside of a motor, and walk us through the pathway to indigenizing that capability?

Cat Buchatskiy: The hardest thing to make inside a motor is the magnets. It’s not really something you can make — it’s a matter of whether you have that natural resource in your country. Having neodymium reserves and being able to mine them to produce magnets is a bottleneck. It’s essentially a choke point.

Ukraine actually does have the natural resources to do a lot of this. We have large reserves of lithium as well, because lithium for batteries is another bottleneck. We have many of the natural reserves, but we don’t have the know-how or the infrastructure to mine them. This is why the rare earths deal with the United States was such a huge deal for us — we’re hoping to get support from US industry and the US government to help us establish that capability.

We basically don’t have an established structure or the IP to access those raw materials. It’s quite difficult for Ukraine to do this at scale during wartime, because many of our critical minerals are in the occupied territories in eastern Ukraine. Even if we wanted to invest significant infrastructure into setting up raw materials production and processing plants, much of it resides under occupied territory. A lot of it would require pretty heavy lifting on our side that would certainly become targets for Russian attacks.

We already have a huge problem with our defense industry being crippled by Russian attacks, with production capacity getting completely wiped out. It’s a significant challenge for Ukraine to do the work needed to access neodymium for making motors ourselves or lithium for making batteries ourselves because of these myriad factors.

Chris Miller: If you dig into the sensors that are on drones too, you’ve got optical sensors, infrared sensors, and all sorts of chips managing communications. Those supply chains are heavily centered in East Asia. Is there an effort to domesticate some of that, or can you get the resilience you need by sourcing from multiple suppliers, not only Chinese ones?

Cat Buchatskiy: We’re sourcing right now. We’ve seen an uptake in sourcing from South Korea, Japan, and Taiwan. We’re obviously always looking to diversify. There’s an effort to manufacture PCBs in Ukraine.

One of the most interesting trends is that we haven’t necessarily started setting up those factories yet, because establishing a PCB plant in Ukraine that can produce at the scale we need is going to be an extreme challenge. What I have seen is that we’re starting a few steps back from that with educational programs.

Some of our major universities now have master’s degree programs in PCB manufacturing, where you can actually get a degree and learn how to do it. We’re investing in the long game. Another interesting master’s program I saw recently focuses on UAV manufacturing, design, and component design.

We’re hoping to raise this next generation of engineers who will not only know how systems work and how to design products, but will really get down to the granular level. We want to raise the next wave of engineers who can make PCBs, chips, motors, magnets — everything in between.

Get back to us in a few years and see where our industry is at. I’m hopeful that we’ll be able to build that know-how, even if we don’t have it right now. Those programs are made in collaboration with engineers from around the world. The lead engineer on one of those master’s programs is from Sweden.

We’re making a long-term investment in Ukraine’s defense industry, from the microcomponent level all the way to final systems. This will be Ukraine’s biggest leverage, our biggest brand, and our biggest export in a few years. We’re investing in making this long-term infrastructure within Ukraine.

Chris Miller: How do you guys think about cost in this context? Obviously, if China is going to sell you a finished motor for cheaper than you can buy the components separately, you’re spending more money. Over a million drones, I’m sure that starts to add up. Walk us through the trade-offs you face around cost versus indigenization and how different companies think about that.

Cat Buchatskiy: The biggest barrier to Ukraine’s component industry is the trifecta of speed, scale, and price. When you talk to manufacturers and ask why they haven’t switched from Chinese components, even if Ukrainians can nail down the speed and scale, the price remains a significant barrier.

However, this will likely change in peacetime. If we reach a ceasefire agreement and establish a recovery rebuilding mechanism where Ukraine’s budget receives more support, we might be able to focus more on investing in this area — something we aren’t considering at all right now.

During wartime, it’s basically impossible to think about making those kinds of investments. Building long-term infrastructure with those price trade-offs is really challenging because the biggest priority on the ground is immediate relief and support to the armed forces in a way that won’t wreck our budget while keeping the economy afloat.

When you’re a policymaker thinking about this, you know it’s going to pay off in the long term for a myriad of reasons. But investing several more million dollars into buying raw materials right now won’t have an immediate payoff because we don’t actually have the processing power to intake those raw materials — we haven’t done that before. We need to understand those manufacturing processes and what to do with this raw material.

Unfortunately, war is a game of immediate payoffs. You need results now, otherwise you’re putting people and lives at risk. I don’t see this being a major investment while Ukraine is still in active conflict, but I do see it as a major investment during a ceasefire scenario. That’s also why we’re investing more heavily in education — we understand that in a few years, we’ll have an educated workforce able to step into that role and focus more on the vertical integration aspect of manufacturing that we simply cannot afford to look at right now.

Jordan Schneider: Cat, what’s up with the controllers?

Cat Buchatskiy: Flight controllers are an interesting one. We’ve been slightly more successful at diversifying from Chinese sources. We’re getting quite a lot of flight controllers from Taiwan, although I’m not sure if that’s less of a supply chain risk than China, depending on what happens there.

We’ve also been able to do some local production of flight controllers, luckily. I don’t know what it is exactly about our workforce, but Ukraine’s software talent and engineering talent has seen more success in the flight controller world. Again, scale is important, which we haven’t quite nailed down yet, especially considering the millions of systems we go through. That’s going to be the next challenge. But in terms of know-how and ability to execute, it’s less of a challenge than the know-how required for motors and batteries.

Jordan Schneider: And Chris, flight controllers are legacy chips.

Chris Miller: Legacy chips and displays, printed circuit boards, and many other components are needed to put these systems together.

Jordan Schneider: Let’s zoom out and talk about broader lessons. What are your big takeaways, Cat, after watching this industry develop over the past four years?

Cat Buchatskiy: My biggest takeaway is that investing in R&D is super important because your drone or final system is never going to be stagnant in an actual wartime scenario. You have to have flexibility baked into your actual assembly process. It’s not enough to have a separate branch that works on R&D and develops separate product lines — you really have to have the flexibility and modularity to adapt to whatever the end user’s needs are at the time. This is really hard for a country that’s not at war.

The US gets a lot of flak for not having the same innovation cycles as Ukraine and not moving as fast. I empathize because it’s genuinely difficult. You’re not going to be able to have that R&D if you don’t have someone using your product every single day in real-world conditions saying, “Hey, this is the next big thing. This is how we need to change it. This is what we’re going to need next.” The US just isn’t in a conflict scenario to have to take in all of those lessons, so it’s hard to replicate.

For US defense companies looking to scale and provide real solutions to the Department of Defense, it should be a priority to form early R&D relationships with the end-user units that will be using their products. Without consistent communication and the ability to adapt — without being locked into any one particular final product — you need to ensure your products are easily assembled and disassembled. This allows you to plug and play with whatever payload, VTX, or camera a particular mission might need.

It’s hard to say one particular lesson like “you should build this thing” because in six months it could look very different. The lesson is that you should be able to build a lot of different things in a flexible, timely manner at scale. That’s probably the biggest takeaway.

Chris Miller: It’s crazy — we get back to what might seem like a facile analogy, but being customer-focused and understanding what the end user actually wants is really important. I came away from this conversation remembering that Jordan was on the side of Xiaomi, not Apple, if I’m remembering our last conversation correctly. These supply chain management and customer understanding skills end up seeming shockingly important.

Jordan Schneider: Even as long as you can get access to the parts, right? That’s the unique piece of this case — the supplier is willing to maybe not give you 100%, but give you 70 or 80% of what you need, and you can make the most of that.

There are conflicts where China wouldn’t really care if we needed a lot of drones to invade Mexico or something horrible like that. But for an Asia contingency, this dynamic isn’t necessarily the case. While the US would have a better chance of spinning up that chemical material and electronic expertise that has atrophied relative to, say, Ukraine, it’s still not a straightforward thing. There’s also a time lag that would be associated with that, which could prove decisive.

I found your research on this broader arc fascinating — building an entire core pillar of your national defense in an existential war on imported parts. I’m not entirely sure... Clearly, there are some lessons, but that fundamental dependency, which Ukraine has been able to live with over the past few years — not really just on China for drone parts, but from the rest of the world for a lot of other inputs to its defense posture — it’s different when we’re talking about World War III and Taiwan.

What’s next in the pipeline for you, Cat? What other research are you guys working on for this year?

Cat Buchatskiy: We’re about to release — very timely — in about a week, we’re about to release our big overview of Ukraine’s defense tech industry as of February 2026. We’ll have updates on what components are gaining traction, what industries are popping up.

Then we’re mainly going to be focusing on doing a little bit of work with Russia’s supply chains and Russia’s defense tech industry. As much as we talk about Ukraine and other countries being dependent on Chinese components, it’s really worth looking into and mentioning the fact that Russia uses a lot of Western components in its drones as well, which is important. They haven’t been fully export-controlled and fully sanctioned enough to cripple those supply lines for them. We might be seeing a big report soon on how Russia still sources Western components and what we can do about it.

Mythos independently identified and patched a 16-year-old vulnerability in the online media library FFmpeg. It also escaped a restricted sandbox and leaked information to the open internet. Anthropic says “it’s about to become very difficult for the security community” and is not releasing Mythos Preview for general-public users. Instead, it is setting up Project Glasswing to share a limited version of the model with AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, for “defensive security work”.

Jordan covered the national security implications earlier this week on the podcast with two former US officials, but China is also an essential part of this conversation. While Claude models remain officially unavailable in China, Chinese researchers and the wider AI community there have followed Anthropic’s work closely. Below are some reactions to the Mythos news from Chinese analysts and technologies, featuring:

• Wrestling with Anthropic’s theory of ethics;

• The Mythos case against Chinese labs’ business models;

• Why you don’t need to worry about your WeChat wallet being stolen — yet;

• And how Project Glasswing puts China on the backfoot.

Translations were drafted with the assistance of Claude Opus 4.6, and then edited for accuracy and fluency. Bold markings were added by the editor.

Does Mythos Incentivize Safety?

Dario Amodei is likely the American AI CEO with the sharpest publicly-expressed attitude towards China. From supporting export controls to repeatedly ringing alarm bells around AI-enabled dictatorships in his essays, it’s no wonder that in China, Claude is sometimes known as “anti-China AI” 反华AI.

But curiously, this hasn’t exactly stigmatized Claude among actual AI users in China. Programmers there are fans of Claude Code, and while no official data exists for the size of Anthropic’s general Chinese customer base, social media content about Claude is everywhere. The US-China rivalry is taken for granted as general context, but individual users don’t feel pressured to switch away from “anti-China AI” yet.

This context is important for understanding why Chinese tech media’s coverage of the Mythos release is not particularly cynical. The mood is closer to slightly-grudging esteem, with few observers loudly doubting what the company has claimed about Mythos’ capabilities.

In fact, some Chinese outlets are quite sympathetic to Anthropic, especially in the aftermath of their face-off against the Department of War. GeekPark 极客公园, an entrepreneurship-focused tech outlet, published an op-ed by a pseudonymous author who defended Anthropic’s decision not to publicly release Mythos. Beyond already-well-articulated safety concerns, the piece situates Mythos in the context of other recent corporate strategy adjustments from Anthropic and analyzes how the lab might be balancing multiple priorities.

On the very same day Mythos was released, Claude’s service experienced a large-scale outage. Today, on April 8, connection issues still haven’t fully recovered, with hundreds of users reporting login failures and chat errors. … A bit earlier, in late March, Anthropic accidentally leaked nearly 2,000 source code files and over 500,000 lines of code during the release of Claude Code version 2.1.88. Security researcher Aaron Turner’s assessment was rather chilling: the leak compressed the timeline for adversaries to replicate America’s strategic advantage, making it a geopolitical accelerant in the agentic AI arms race.

…

Putting these events side by side, and Anthropic is fighting on three fronts at once: infrastructure stability, the boundaries of its business model, and — the hottest issue of all right now — just how dangerous the thing it built actually is.

The way Mythos was released is, in a sense, a high-stakes bet on Anthropic’s “responsible AI” doctrine. They chose the most conservative possible method to unveil the most dangerous possible model — telling the world “here’s what it can do,” while refusing to “let it do it.” The logic behind this move: only by publicly disclosing the threat can you drive defensive action; but opening up the capability itself could trigger a chain-reaction catastrophe.

Whether that judgment is correct, nobody knows yet.

Chinese Model Doomerism

Founder Park, a subsidiary of GeekPark aimed at a founder audience, wrote about the implications of Mythos for AI as a global business. The piece doesn’t mention China outright, but is clearly pessimistic about the prospects of open labs with less-capable models in a post-Mythos world. It lays out an interesting case against the possibility of a democratized future for AI.

[ChatGPT] has locked us into the assumption that flagship models will be supplied and sold abundantly, at a price that tens of millions of people can afford. Building on top of this assumption, we imagined MaaS [Model-as-a-Service],1 the token economy, and how agentic coding would help or replace programmers — but once the spiral kicks in, this assumption no longer holds.

Anthropic’s current annualized revenue is $30 billion. Suppose Mythos really does have the ability to sweep through and uncover system vulnerabilities at scale; why would Amodei make it public? Selling MaaS makes money, and charging membership fees makes money; however, collecting protection money makes money too. Think about how Amodei could easily unveil Mythos with a five-point statement:

1. AI now has the capability to discover and exploit system vulnerabilities at scale;

2. Evil nations and organizations are about to acquire this capability, and they’re only six months to a year behind;

3. But our Mythos is ready;

4. As long as you’re an upright company that cares about human civilization and shares Anthropic’s values, Mythos will come protect you;

5. Next, please wire payment to Anthropic. After reviewing your values, we’ll decide the priority order in which you receive Mythos protection based on your payment amount and our internal values-alignment score.

…

This is the first model that wasn’t immediately made available via API, and it therefore represents an entirely new commercial reality.

…

To elaborate a bit: some might say that AI is currently flourishing, and other companies (especially Chinese ones) will soon catch up.

This, too, is an illusory assumption born of the past three years — really, the past one year. Once flagship AI stops being offered publicly, [labs that trail in capabilities] won’t just be unable to distill flagship AI; even finding out how flagship AI works or how it solves problems will become increasingly difficult. Internal opacity at AI companies will also inevitably keep rising in order to prevent leaks.

Will this day come? If so, we’d better pray that current AI technology isn’t yet enough for the spiral to hold, that technological progress isn’t fast enough, and that AI companies still have to publicly offer flagship AI services to build momentum and capture more profit.

Mythos is Anthropic’s forceful attempt to break into “the next act of the LLM.”

Anthropic is committing $100M in model usage credits to Project Glasswing. After these run out, it plans to charge $25/$125 per million input/output tokens for Mythos access by approved participants. Cyber Zen Heart 赛博禅心, a well-known tech influencer account (previously featured in our WeChat AI Field Guide), put together a summary of the various signals Anthropic might be sending with this approach. It’s a more moderate interpretation of Anthropic’s thinking, with close analysis of pricing and revenue strategy in anticipation of its potential IPO this year.

Product line expansion

Claude’s product line has gone from three tiers to four. Above Haiku, Sonnet, and Opus, a new Mythos/Capybara tier has been added. This change itself matters more than any single benchmark result. It means Anthropic’s model capabilities have already opened up a gap wide enough to require a new price band to absorb it. Based on documents leaked via Fortune, Capybara is internally defined explicitly as a new tier “larger than Opus”, representing a structural expansion of the product line.

Leading with the safety narrative

Mythos is a general-purpose model that’s strong in coding, reasoning, and search, and could easily have taken the standard benchmark-release route. But Anthropic chose the “too strong to make public” narrative, giving access to only 12 major enterprises. This reflects both genuine consideration of safety risks and a statement about pricing power and ecosystem control. Want to use the strongest model? Join Glasswing and buy tokens at $25/$125.

Anthropic is choosing not to let you use its strongest model, but it’s telling you exactly how strong that model is.

Pricing signal

The $25/$125 pricing is about 67% more expensive than Opus 4.6’s $15/$75. If Mythos-tier models do eventually go public, this price band will become the new anchor. For anyone who believes token prices will only keep getting cheaper, this pricing is a counterexample: when capabilities are strong enough, prices can move upward.

Timeline

On April 4, the subscription channel for OpenClaw was shut down. On April 7, Mythos was released. On one hand, Anthropic is tightening control over the open ecosystem (you can no longer use a monthly subscription to run third-party agent frameworks without limit); on the other, it is releasing the strongest model to enterprise partners. Just three days apart—the cadence is tightly choreographed.

Perspectives from Technologists

Robin Li 李彦宏, founder and CEO of Baidu, noted infamously in a 2018 speech that “Chinese people are … less sensitive about the privacy issue.” The comment prompted major backlash, but struck at a certain level of truth. Forcibly thrusted between corporate and governmental surveillance — neither of which is easy to opt out of — many Chinese internet users have resigned themselves to surfing the web with little privacy protection.

Of course, Chinese people are far from alone in suffering from the consequences of corporate data leaks and feeling powerless in the face of pervasive data collection. But some of the Chinese internet’s regulatory and commercial attributes arguably make it fertile ground for vast cyber threats. To sign up for accounts with many Chinese apps — including popular AI chatbots and tools — you need phone numbers tied to Chinese national IDs. Social media platforms have been required to institute real-name authentication. Super-apps integrate users’ financial, governmental, and interpersonal lives. Sensitive data about millions of ordinary people is increasingly concentrated in the hands of a few companies, and regulations still struggle to catch up.

Cybersecurity experts were already unnerved by what havoc AI agents might wreak for the security landscape — and then came Mythos. A consumer-focused tech media outlet, 差评X.PIN (chàpíng, literally “negative review”), interviewed an anonymous cybersecurity researcher about Mythos’ implications for regular people’s online safety. Their response:

If everything in [Anthropic’s red-teaming technical blog] is true, I feel like half of the people working in internet security right now should just jump into a river

“Wen’an” (the researcher’s pseudonym) clarified to X.PIN that they were being hyperbolic, but also offered some sobering analyses of how models at the Mythos level will reshape cybersecurity as we know it.

… these vulnerabilities haven’t yet reached the level of clearing out your Alipay balance or splashing your WeChat chat logs all over the internet.”

But the crux of the issue is this: the reason Anthropic released these cases publicly wasn’t to show off “how nasty the exploits are,” but to demonstrate that AI — without any plug-in tools, relying purely on its own knowledge base and cross-domain reasoning — can dig up brand-new vulnerabilities on its own.

So in Wen’an’s view, Mythos at this stage isn’t “a stronger hacker tool,” but rather a lowering of the barriers to entry for cyberattacks.

In the past, whether you were a legitimate security professional or someone working in the gray/black market, you at least needed someone who knew what they were doing to run the show. Pulling off a real, serious cyberattack meant holing up in a dark room for months on end.

But going forward, it might be enough for that pudgy village loiterer to shout a couple of voice messages at an AI while picking at his feet.2

This kind of “if you’ve got hands, you can do it” low-entry-barrier setup is inevitably going to attract hordes of thrill-seekers and outlaws looking to have a go.

That’s why Wen’an thinks it actually makes sense for Anthropic to roll out the Glasswing program first.

After all, traditional security tools are like rigid gatekeepers: they only check whether you’re carrying contraband, and they’re useless against insider jobs. AI, on the other hand, can trace the threads, understand business logic, and spot the kind of move where John Doe uses his own key to open Dave’s door.

Letting the big enterprises self-audit and trial the tech in advance lets them get a head start on building network defenses, running vulnerability sweeps, and preventing problems before they happen.

Not everyone believes that Anthropic is being entirely altruistic, of course. In their commentary about how the cybersecurity challenges introduced by Mythos relate to US-China competition, China-based analysts for the consulting firm IDC did not mince words. They see Glasswing widening the capabilities gap between America and China’s AI industries, as well as threatening the entire technical foundation of the digital economy in China:

One core challenge is asymmetry in technology access, which creates a clear technological gap with overseas peers. Participants in Project Glasswing can prioritize leveraging Mythos’s powerful capabilities to conduct vulnerability discovery, threat detection, and defensive system optimization, while simultaneously sharing related security research findings and open-source resources — enabling rapid iteration of defensive capabilities. Chinese vendors, by contrast, are completely excluded from this collaborative framework. Unable to directly access Mythos’s model capabilities or related security resources, they can only rely on their own efforts to develop the relevant technology. This creates an inherent gap in the iteration speed of AI security technology between China and its overseas peers. Especially in high-end domains such as zero-day vulnerability discovery and AI adversarial techniques, this generational gap may widen further, and closing it in the short term will be difficult.

The second core challenge is a dramatic escalation of cybersecurity threats and a sharp increase in pressure on the defense of critical infrastructure. In China, critical infrastructure sectors such as finance, energy, government services, and healthcare make extensive use of various open-source software and general-purpose operating systems, and Mythos has already uncovered a large number of high-severity vulnerabilities in these systems. Overseas vendors participating in Project Glasswing can use the model to quickly obtain vulnerability information, generate fixes, and complete system hardening in a timely manner. Chinese vendors, however, cannot access the corresponding vulnerability information or remediation guidance, and must rely on self-directed investigation and self-directed patching. This not only sharply raises defensive costs but also prolongs vulnerability exposure windows, significantly increasing the risk that critical infrastructure will be attacked. At the same time, as Mythos’s capabilities proliferate, the state-level APT attacks and black-market attacks China faces will become more covert and more efficient, with a wider variety of attack methods—further compounding the difficulty of cybersecurity defense and posing a serious challenge to the nation’s cybersecurity baseline.

In addition, China’s cybersecurity industry faces derivative problems such as a shortage of AI security talent and significant pressure on investment in indigenous R&D. While Project Glasswing fosters a healthy ecosystem of “technology sharing + talent collaboration,” China, by contrast, suffers from an insufficient supply of high-end talent in AI security. Its indigenous R&D lacks mature technical reference points and ecosystem support, which further constrains the improvement of defensive capabilities, leaving China in a more passive position when responding to the AI-driven attacks that Mythos makes possible.

Mercor isn’t top of mind in DC today, but the expert-data ecosystem it represents is a core determinant of AI capabilities growth and the U.S. frontier AI advantage. Last week’s hack and data breach at Mercor, and the broader security challenge it represents, may shape the future of tech competition. It also could hold the keys to understanding two rapidly vanishing U.S. assets in the field: the specialized-data moat and a differentiable frontier itself.

What is Mercor?

Mercor is a $10Bn startup that builds specialized human-expert datasets. Firms like Mercor, Surge AI, and Turing1 are key to AI capabilities growth, helping labs unlock high performance in new domains. These datasets are one large reason why model performance continues to improve even after foundation-model labs scraped most of the web many training runs ago.

Models are only as good as their data. But one of the greatest challenges in AI development today is finding a way to make new, less-documented fields legible to models. While today’s frontier models have achieved incredible capabilities, they still struggle with “generalized” reasoning when asked to function in domains where they lack robust training data.2 Thus, one of the key drivers of capabilities growth, especially for the type of white-collar workflow automation critical to helping models complete high-value tasks, is the collection and curation of specialized datasets in expert domains. For example, radiology models train on large corpora of X-rays.3

Mercor specifically conducts these operations through a hiring platform.4 As Mercor built this platform’s talent base, it accumulated an impressive set of specialties from biotech research and interventional radiology to corporate law and international business development. This data is incredibly valuable for the labs; when it helps models generate new insights, it unlocks whole new production domains.5

Demand for Mercor’s products has grown alongside the AI fundraising flywheel. As foundation labs raise more money to complete ever-larger training runs and reach ever-more users and ever-more-demanding benchmarks, demand and available capital for specialized data rise, too.6 Labs spend tremendous amounts on specialized data. An exclusive with The Information suggested that Mercor’s annualized revenue had recently reached $1 billion.

The Most Important Hack DC’s Never Heard Of

But recently, Mercor was hacked. On Monday, March 30, a group known as Lapsus$ claimed to have stolen 4TB of Mercor’s data. The hack allegedly included everything from candidate profiles and personally identifiable information (PII) to video interviews with experts, source code, and other proprietary information and secrets.7

The Mercor hacks suggest that expert data companies could be a weak link to copy or steal a tremendous lab investment in data. Suddenly, critical data from a company built on proprietary data was available for purchase, allegedly at a price of $1 million for nonexclusive use. For scale on how good a deal this represents on such assets, Mercor pays its contractors over $1.5 million every day to build them.8 Mercor project lengths vary widely, but if the hack contains data from a substantial share of projects, the contractor costs alone may exceed the price of the leaked dataset many times over.

Mercor’s close partnership with the labs also raises the concern that the hack might have exposed secrets about how foundation-model labs manage their product development. Within days of the hack, Meta had paused its contracts with Mercor, reportedly due to such a concern.

We don’t know specifically what data was leaked in this hack. (For example, it is unclear how much annotated data or process secrets were exposed, as opposed to data about the experts or less important procedural concerns.) But the specific fallout from this hack might prove less significant than demonstrating that these kinds of hacks are achievable. The most consequential concerns might be future hacks of expert-data startups and the threat of them.

Such concerns are especially acute in the wake of Anthropic’s announcement this week that they would withhold release of their new Claude Mythos Preview model due to its immense potential to conduct cyberattacks. Even if Anthropic refrains from releasing its model today, other close followers might not be so judicious. This development raises concerns that advanced cyberattack capabilities are coming, fast.

Expect Fast-Followers to Pounce

Fast-follower foundation-model builders — especially in China — will surely try to access this incredibly valuable data. (From both this hack and any future attacks.) Why? Beyond these firms’ sophistication, they’ve also conducted far more controversial operations recently. Take, for example, Anthropic’s recently publicized allegations of mass distillation of Claude models by Moonshot AI, Minimax, and DeepSeek. OpenAI raised similar concerns, including in a letter to the US House Select Committee on the CCP. Obviously, data curation strategies extend far beyond distillation attacks.9 But these incidents suggest the inventive methods Chinese firms are willing to employ to close the gap with leading U.S. labs. They’ve shown a willingness to use even nominally closed-source models to develop their own, often open-source alternatives. Alongside fighting Anthropic and OpenAI’s distillation defenses to build a synthetic dataset, it is also surely worth playing around with the datasets.

Two Key Takeaways

1. Increased Urgency of Strong Export Controls.

Threats to one moat increase the importance of protecting another. If one can distill models and steal critical expert data, then access to compute becomes an even more important long-term differentiator.

2. A Higher Premium on Security for National-Interest Expert Human Data Startups.

This incident strengthens the case for more active state involvement in bolstering cybersecurity for strategically significant AI companies. Cyberattacks like this could disincentivize innovation: Why invest in high-quality, bespoke datasets if hacks will let non-paying firms free ride? The need to protect a critical asset like national-interest AI startups that help power our labs justifies federal security assistance. (Put another way, perhaps such calls to secure the foundation-model labs need to be expanded to key partners like expert human data firms.)

Such an arrangement would allow these national-interest entities to leverage existing state infrastructure to provide, among other things, unique threat intelligence; testing; and incident response and support with the state’s unique authorities, scale, and visibility. Such services already have precedent. For example, the NSA offers cybersecurity services to private contractors working with the Department of Defense War. Similarly, the FBI’s Business Alliance Initiative offers support to private firms in the form of counterintelligence vulnerability assessments, information on specific threats, and advice on a variety of infiltration scenarios.10

If export controls and enforcement11 are key to maintaining U.S. AI advantage, then leveraging state security strengths that can incentivize firms like Mercor to keep developing datasets that advance the frontier in critical professional work is key to continuing U.S. AI innovation and capabilities growth.

Ben Buchanan, former senior advisor for AI at the White House now at SAIS, and Michael Sulmeyer, former Assistant Secretary of Defense for Cyber Policy now at Georgetown, join the show to break it all down.

Full disclosure: Ben advises Anthropic.

We discuss…

• How Mythos found 27-year-old bugs in code everyone thought was secure

• The offense-defense balance: whether a Ukraine with Mythos and a Russia without it changes the war

• Project Glasswing and Anthropic’s attempt to build a private-sector vulnerabilities equities process

• Why critical infrastructure patching is about to become a nightmare

• What happens when ransomware gets vibe-coded

• Why bio won’t be far behind

Listen now on your favorite podcast app.

Thanks to the Hudson Institute for supporting this work around AI and the future of war.

Has the Atomic Bomb of Cybersecurity Just Been Discovered?

Jordan Schneider: So how big a deal is Claude Mythos?

Ben Buchanan: This is a big one. I’ve been thinking about cybersecurity and AI for more than a decade. I think a lot of us who were thinking about AI and cyber back then imagined that a day like this might come where you could see automated vulnerability discovery. It does feel like something that had long been imagined is actually now finally here, and it’s up to all of us to figure out what that means.

Jordan Schneider: So what can the model do?

Ben Buchanan: What this system does at its core is it takes a general-purpose capability — it is not a cyber-specific model — and applies it to the business of vulnerability discovery and exploit development. As Michael can attest very well, these are fundamental tasks in cybersecurity: finding a weakness in a piece of computer code and then figuring out how to exploit that weakness to do something as an attacker that you’re not allowed to do.

The evidence is very clear that Claude Mythos is by far the best automated system in the world ever to do this, and is better than even some of the best expert humans — or close to some of the absolute top-tier expert humans — at this task of vulnerability discovery and exploit development. The proof is in the pudding. It found vulnerabilities in code that all of our operating systems and all of our browsers are running. Those vulnerabilities in some cases had lurked there for multiple decades. In some instances, we thought that code was secure. Millions of automated tests had been run on it, and yet Mythos found ways to exploit it. There is a real raw capability here that is vital.

The question is, what’s the analogy for that? That is really an important question.

Jordan Schneider: Let’s stay on the finding of a 27-year-old bug in a piece of open-source software that the entire world uses. Michael, how wild is that?

Michael Sulmeyer: That’s pretty wild. I ended up talking to one of the original developers of some of that software. And it was just silence on the other end. Because everyone thought this was almost axiomatic in computer software development and in cybersecurity — that this piece of code was secure. Knowing that at some point this day would probably come where they’d find problems in it, but that today was going to be the day, and it would be a machine that did it.

Jordan Schneider: The point being that this type of thing — the entire world has been looking for it in this library for decades. You would think that someone would have been able to find or exploit or patch it, given this level of proliferation. This is not the sort of thing where Apple pushed a new update three months ago and we’ve got to work the kinks out.

Ben Buchanan: No, this is long-standing code, there for decades. The core credo of the open-source software movement, which I should be clear I totally support, is: with enough eyeballs, all bugs are shallow. Basically, if enough smart people are looking, they will find everything that is to be found. I think the answer for this moment is we need to have machines look too — or at least, a machine of this capability level can find things that a lot of good humans looking for a long time didn’t find.

The Nuclear Analogy

Jordan Schneider: So let’s come back to the nuclear analogy I started with. The US, of course, invented the atomic bomb and then had a good four-year run of exclusive access to this power. No one else has this model. Just for the record, where are we? Is the US government not allowed to play with this at all?

Ben Buchanan: I would ask the US government. I don’t know what the particulars are between Anthropic and the US government.

Jordan Schneider: Okay, but there’s like a six-month whatever. So TBD on that. Because my first thought was, this is almost like U-boats’ 1942-style happy times. If you’re the one person in the world who can use the offensive version of this, where on the other side you now have Project Glasswing and the whole world trying to harden their systems. Michael, how much fun would a nation-state doing offensive stuff potentially be able to have with this power and no one else having it?

Michael Sulmeyer: “Fun” is probably not the word I would use in the official setting…

Ben Buchanan: You’re out of government now, you can say what you want.

Michael Sulmeyer: That’s true. I think when you think about what is the fundamental responsibility for the kind of role you have in government, it’s to bring options to the senior-most decision makers. What something like this allows for is a new set of options — if used for offense and exploitation purposes — a new way to really scale those options for decision-makers. Whatever the expected outcome is, for better intelligence collection or other types of purposes, it really opens up the opportunity space.

What I think remains the same is that success in cyberspace generally has come down to a race — a race from when the offense or the exploiters know about a problem and how fast they can get at it, compared to how fast the defenders can identify, fix, and then disseminate the fix as broadly as possible. So part of the answer is: if you’ve got the offense, you’re the only one, and defense doesn’t know, it’s pretty open season.

Ben Buchanan: I know Michael agrees with this, but Rob Joyce, who was the head of what was then called Tailored Access Operations — the pointier part of NSA — gave this talk at USENIX in 2016, which I actually used as the basis for a paper in 2019 or 2020 when I was starting the cyber-AI project. The basis of the talk is walking through the steps of offensive cyber operations. This is the first time someone from NSA is out there saying, essentially, how NSA at a conceptual level goes about its business.

The conclusion we came to in 2019 and 2020 was that at least theoretically, at each step of that offensive operation process, AI could help. Now I think with something like Mythos, that conclusion is just far more robust. We saw the glimmers of it in 2019 and 2020, but Mythos is really doing it — not just in vulnerability discovery, though that’s a key part of it, but throughout the process. There’s something in the system card for Mythos where it carried out a simulated network exploitation that would have taken a human 10 hours. So there really is evidence now that what cyber operators call the kill chain can be transformed by AI capabilities.

Now, a separate question for Michael is, of course — will the governments be able to adapt? That’s a whole other thing. But as a technical matter, it seems to me we’ve crossed that Rubicon.

Russia, Ukraine, and the Offense-Defense Balance

Jordan Schneider: Let’s talk about the status quo ante. Russia-Ukraine is maybe the best analogy, because that’s the conflict where we’ve presumably had the most no-holds-barred cyber going on between two countries in a hot conflict. When you’re ranking the things determining battlefield progress or morale, cyber is pretty low on the list. Michael, is your sense that these two countries are equally sophisticated and the technology leads you to fight yourself to a standstill? Is a Ukraine with Mythos today performing radically differently if Russia doesn’t have it?

Michael Sulmeyer: It’s a really good hypothetical. I think it could give Ukraine the advantage. I also really like how you distinguished cyber operations from electronic warfare, which is a common conflation. You see a lot of battlefield use of EW, which has had important battlefield effects. You’ve seen much less of the kind of cyber attacks-type work. It was the opening shot in some sense of the conflict with the Viasat compromise, but it wasn’t really exploited and leveraged.

I’d say you’re probably still seeing a lot of cyber intrusions — I’ve been out of the business for a while — but that’s different from creating destructive or disruptive effects that would degrade and disrupt morale. We shouldn’t think that there isn’t a lot of aggressive, malicious cyber activity going on between Russia and Ukraine right now.

Ben Buchanan: It seems to me that if you had something like Claude Mythos as a state, you would probably want to use it for your intelligence operations or your pre-positioning, because you are almost by definition going to find vulnerabilities no one else knows of with this system. And you don’t want to make a lot of noise about that. You want to go in, set up a persistent, quiet presence. My view for decades has been that the advantage of cyber is not the whiz-bang sky-is-falling blackout — though you can do that sometimes — it is the slow, insidious shaping of the environment and collection of information. A capability to find vulnerabilities and exploit them autonomously would really help on that side of the ledger.

Michael Sulmeyer: Can I ask Ben a question on something he just said? You mentioned shaping, and you have a great piece from many years ago where you questioned in a lot of ways the difference between signaling and shaping. I think the answer is pretty clear on how well Mythos would help with shaping. Does it help the ability to signal through cyberspace at all, for the crowd that’s obsessed with deterrence?

Ben Buchanan: The backstory here, as Michael alluded to, is my pitch back when I was a cyber academic — even before the White House — was that cyber operations were suited to shaping: stealing a card, stacking the deck, rather than changing how the other side plays its hand. I don’t think Mythos changes that.

The broadest thing you could say about a capability like this is, in the abstract, it has some brandishing value or maybe even deterrent value because it bolsters the status of the nation that has it. But I imagine a government who truly wanted to play offense would want this kept quiet so that people don’t go looking for it. Anthropic has very clearly come out and said their goal for this technology is not to play offense — their goal is to tilt the balance of power in cyber operations to the defender. Michael mentioned Project Glasswing, where Anthropic is trying to give access to some critical software developers — Apple, Google, and the like — to make sure systems are secure before a Mythos-like capability proliferates.

The bottom line for me is this is incredibly important for understanding the landscape of modern cyber operations, but it does not fundamentally change their character, which I think is still one of shaping rather than signaling.

Project Glasswing and the Proliferation Clock

Jordan Schneider: How messy is this going to get? You brought up Project Glasswing — basically, the idea is let’s give this to the adults first and let them play with it for an undetermined amount of time. There were about 40 companies, a pretty awkward line in the press release saying they’re open to partnering with federal, local, and state governments. TBD on that one.

At some point, the past five years of model development have shown that this is not something that only one company with a certain view about how this capability should be rolled out is going to keep under wraps. What happens when someone else who’s not only giving this to folks who want to patch up holes gets access to this technology? That could be someone training a new model, someone releasing this in the US, someone releasing this in China, or someone hacking Anthropic’s servers.

Ben Buchanan: I buy the premise here. Anthropic says it in their press release — this is going to get out there at some point because folks will catch up. Let’s not overstate that though, which is to say it appears by all accounts that Mythos had a huge compute requirement. I’m sure we’ll get to export controls at some point, Jordan, because you and I always do. This is not the kind of thing that you could just train out of the box without a lot of computing power. Things like export controls will constrain who has access to this level of capability for a while. How long? I’m not sure, honestly.

Michael Sulmeyer: I do think there’ll be a transition period. And I do think Anthropic’s model of the situation is right — that we want to, during that transition period, patch as much stuff as we can, find the vulnerabilities and patch as much as we can. The consortium, the Glasswing thing, I think is pretty interesting because you have like 12 named members and then a broader group of companies, many of which are fierce competitors, all coming together and saying this is a systemic threat and we have to get ahead of it. That’s exactly the kind of response we need in a transition period. Because to your question, Jordan, we don’t know how much time we have. It’s probably not a ton, even though I think it’s more than some people expect.

Ben Buchanan: What I also take away from it — there is a bureaucratic process which I’m sure every single listener is well aware of, Jordan. After the Snowden disclosures, there was a study and report done, and they recommended the White House create what was called a vulnerabilities equities process: what should the government do when it discovers a vulnerability in software, and how does it make the cybersecurity trade-off versus the exploitation or offense trade-off?

What you have through Glasswing is, I think, one of the first efforts by a private-sector company that has developed a capability that finds these vulnerabilities to figure out its own almost multinational vulnerability equities process. It’s a remarkable effort to manage those equities and do it in a responsible way. It’s tough in government, and I think it’s going to be real tough outside in the private sector to do it too.

Jordan Schneider: Let’s talk about that. That whole government process has had some ups and downs. And now we have the private-sector version of it, because this White House is fighting with Anthropic and maybe just didn’t quite believe that this transformative capability was right around the corner. Michael, let’s talk about the pluses and downsides of the federal government seemingly taking a backseat for now.

Michael Sulmeyer: A vulnerabilities equities process in government really works well, and I’m a big supporter of it, when there’s a commitment to action and actually fixing with urgency the problem that’s pointed out. Where it doesn’t work and where it feels disappointing is when you go through the whole effort to do the right thing, make sure you warn the company that’s got a problem, and it doesn’t feel like the urgency is there to fix it.

Jordan Schneider: Urgency on the part of the vendor?

Michael Sulmeyer: Yeah, the vendor. You’ve handed them the understanding of what they’ve got to do to fix it, but now they’ve got to go fix it, and they’ve got to do it quickly. They can’t just sit on it.

A great thing you’ve seen in Glasswing is Anthropic saying they’re going to come back in 90 days showing which vulnerabilities have been fixed. That’ll be a good test of the urgency on pickup by all the partners, and also a way to improve the process going forward.

The Patching Nightmare

Michael Sulmeyer: I do think this is an important point — historically, sometimes even once the patch was issued, people wouldn’t apply it for a long time, or the company would know the vulnerability and take its time issuing the patch. The whole process from discovery of the bug to development of the patch to deployment of the patch — that’s going to have to go so much faster in a post-Mythos era, because stuff like this will proliferate and folks will be looking for these things and maybe they can reverse-engineer patches. The IT industry and the backbone of critical infrastructure is going to have to level up in speed because of Mythos.

That probably is a harbinger of what’s going to come in AI — that where we have the things for societal resilience, we’re going to have to get more resilient faster for individual cycles because AI is going to accelerate the offensive side of the ball.

I would also note — on a good day at a software company where you’re talking about a vulnerability found in actively supported software where the developers are still employed, it’s still difficult. Now you also have to factor in the situation where you find a vulnerability in software where all the people who wrote it are gone, and the company said, “We stopped supporting this thing years ago.” Just thinking about how you’re going to manage the scale of vulnerabilities that’s going to come through here in the near term across software — whoa.

Ben Buchanan: There’s a flavor of that in critical infrastructure as well. Even if the critical infrastructure companies are still in business or the software providers are still in business — you tell me, it’s closer to your world than mine — but my understanding is that is a messy set of systems to patch. It is not meant to take critical infrastructure down every week or every two weeks to apply a software update. Sometimes the uptime is measured in months or years.

So if one of the effects of this new world is that AI systems find vulnerabilities in critical infrastructure software with a much higher cadence, that’s going to be its own complexity. And of course, the consequences of failure are pretty high there.

Michael Sulmeyer: Right now, you use Chrome, and the smart people at Chrome force Chrome to reboot to apply the patch after a certain number of days. Apple figured out that if you get people new emojis, they’ll update their iOS and get some good security vegetables with it.

Ben Buchanan: There’s a whole swath of software and associated hardware that is not subject to that kind of patch cycle. If a vulnerability is found there, it’s going to be a real problem. We thought a lot of software was secure, but then again, we thought some of the software that Mythos found vulnerabilities in was secure as well. And clearly it wasn’t.

Offense vs. Defense in the Mythos Era

Jordan Schneider: So what does Mythos tell us about the offense-versus-defense dynamic with accelerating cyber capabilities? The hope is, all right, well, maybe I don’t have to hire these software engineers back. I can just press a button.

Ben Buchanan: I think in the near term, putting Anthropic’s efforts to benefit the defensive side — if Mythos were just dropped in the world for anyone to use, a capability like this would clearly benefit the offense. And I think Michael should talk about some of the ways in which it can benefit the offense at each step of the kill chain.

Michael Sulmeyer: My hope is that we can get through some kind of transition period in which it benefits the offense, mitigate that as much as possible by differentially privileging defenders, and then we end up in a world — I don’t know how long this is going to take — where new code is secure, Mythos has found most of the vulnerabilities that are out there, and we have patched those.

The counter-argument, which is a pretty compelling one, and which is why this is a hope and not necessarily a prediction, is as Michael said, some companies don’t exist and their software is not going to get patched even if a patch was developed, just because there won’t be anyone to push it out. Critical infrastructure is harder. In the long run, it’s going to be messier, but you can tell yourself a good-news story here if society can use this technology to its fullest extent.

Michael Sulmeyer: I’m skeptical we’re going to be able to manage it, but that’s the good-news story. You can’t give up. Glasswing is the best way at scale to give defenders a fighting chance. I cannot think of a different or better way to deal with it, but that doesn’t mean there aren’t still structural issues that are going to make uptake more challenging, especially with critical infrastructure.

Ben mentioned some parts on the offense. If you think about that old Rob Joyce / Ben Buchanan framework of what the offense kill chain is — from reconnaissance to gaining initial access into a system, to persisting in that system, lateral movement to get where you want to go, and then finally generating an effect — those are the five parts of what Rob Joyce talked about and what Ben wrote about as well.

AI, even without Mythos, probably helps you along each one of those. If you wanted to say which one it helps the most and which one the least — there’s an argument that AI would help you with persisting really quite a bit in novel ways, because once you’ve broken in, you have to make sure you don’t get caught. Finding ways to blend in with what normal looks like within a system and to adapt on the fly — that’s pretty cool if you can do that, and it’s a hard thing to do remotely for humans. Probably the one that may not benefit the most — still benefit, but maybe not like persistence — might be effect. There are still ways to improve how you extract data, but you’re still extracting information. You might do it in more creative ways or extract more in a shorter amount of time, but I’m not sure that’s where you’ll see the step change on offense from AI.

Ben Buchanan: There is an interesting notion here. If you look at the most powerful cyber attacks historically — and here I’m saying attacks very deliberately as opposed to espionage — there is already, before we even get to the machine-learning era, an automated component. And it’s that automation that gives it scale.

Whoever attacked the Iranian centrifuges in Stuxnet — that clearly has an automated component that lets it spread from system to system. If you flash forward to 2017, the WannaCry attack from North Korea — not clear it was meant to be an attack, but was an attack — had automated propagation. The Russian attack NotPetya in 2017, probably still on a dollar-value basis the most damaging destructive cyber attack in history, probably more than $10 billion worth of damage, clearly very automated.

So there is an intuition we can develop in which automation in cyber operations, even before the machine-learning era, can yield the power that manual operations can’t. And there have been some near misses. One of the most overlooked cyber attacks in history was the Russian blackout in Ukraine in 2016, a case called Crash Override. The Russians, for context, had been in Ukraine in 2015 — December 2015, they caused a blackout, very manual, this beautiful symphony of all these different pieces of the operation coming in at once, but totally manual. A year later, the Russians come back, they try it again with Crash Override, but in a totally automated way. Maybe the Russians being the Russians, they screw it up — the effects were not world-changing, powers out for an hour or something.

But one of the questions that came to my mind when I started seeing the cyber capabilities of Mythos is, how would this work in targeting critical infrastructure? Could this actually manipulate a programmable logic controller or industrial control system in a way that has a kinetic effect, in the way that the Russians tried and failed to do? The answer might be yes. Or if this one can’t, the next one can.

I do think there’s an element in which the automation of the kill chain yields more overall power for the attacker. On the flip side, though, for defense — put aside AI. There are things we’ve known for years that you could do to frustrate efforts like that. Air-gapping an OT network from an IT network, for example. That doesn’t take AI to do on defense.

Michael Sulmeyer: It doesn’t necessarily take AI to exploit it. But the fact is that kind of foundational step isn’t done nearly enough. It isn’t implemented nearly enough. I wouldn’t want to lose the point that foundational cybersecurity measures — just doing what we know works — doesn’t go away just because of Mythos. You should still, maybe all the more reason, urgently do what you know you probably should have done a while ago, because that will help you. It may not totally protect you — it was never going to totally protect you — but it will make a model’s life harder to jump an air gap at a critical infrastructure system.

I totally agree with that. Basically where we’re both landing is Mythos is a game-changer for cyber operations in that it’s going to change how sides play the game, but it’s still the same game, and the core tenets of what works in cybersecurity — I think those are going to hold for a while. One of the chief ones is the defender has this huge advantage that so few defenders actually realize and claim for themselves: they set the terrain. They get to decide where the operations are going to take place within their network in terms of protecting the boundaries and air-gapping.

I’m optimistic that Mythos can, in the right hands, aid defenders by making those networks more secure, by finding the configuration errors before the attackers do and remediating those.

I’ve seen it work really well, unfortunately, only after the organization has been had. Once you get nailed, then you find that it turns out you could just air-gap these two networks, or you could implement these kinds of changes. You could have done it before, but now you’ve got the urgency, now you’ve got the resources. Even though all the evidence in the world said that was the thing to do, you needed to be attacked first to be convinced. And I hope we can get over that hump.

A bunch of American banks — JP Morgan in 2012 got a lot of Iranian incoming, DDoS and intrusion stuff. That was one of the seminal moments for the financial sector. You talk to them five years later — how did you guys get serious about this? They recognized a very clear business case with credibility, and they felt the incoming in 2012. They were lucky that those were not extremely destructive attacks, but it was a galvanizing moment for the industry.

Michael Sulmeyer: What Anthropic would say, and certainly what I would say, is Mythos should be a galvanizing moment for every industry. A reminder that now the clock is really ticking before the offensive side of the ball levels up in cyber, and defense has to get there first.

On those Iranian DDoS attacks, what I think is also really important to note is that those institutions — whether they knew it or not at the C-suite level — at the CIO level, I think knew they could have worked with and been under Akamai’s protection to have resisted that kind of activity anyway. They just hadn’t wanted to do that. It was all totally preventable. They just, again, needed to unfortunately go through it. Not to pick on them, because it’s hard to adjust — I get it — but that just furthers the point. They know what to do. We can give Jordan his show back here.

Jordan Schneider: No, this is amazing. Thank you, guys.

Ransomware, Voidlink, and the Non-State Threat

Jordan Schneider: There were some really interesting lines in the red-team report by outside cybersecurity experts that were almost like an emotional plea, saying: we’re in for it, strap in, this is going to be really messy and really painful. Ben, you mentioned the $10 billion of damage from a Russian cyber attack that went awry. I’ve been asking ChatGPT for ransomware and cyber-extortion numbers, which seem really low — in the tens of millions or like $120 million in 2025.

Ben Buchanan: I’d have to imagine it’s more than that globally, in the aggregate.

Jordan Schneider: That seems like it’s going to change. How much more mischief can you get up to if your goal isn’t just finding schematics for fighter jets or poking around systems like China has been reported as doing over the past few years, but really just messing things up — in a “don’t care what happens” way, or in an “I want to extract enormous amounts of money from desperate organizations” way?

Ben Buchanan: There’s no doubt that a capability like this in the wrong hands would allow a lot of that. It’s hard to put a dollar figure on it, but you could probably do billions of dollars of damage if you were going no-holds-barred or if enough groups had access to it.

An interesting case — I think from maybe January or February of ‘26 — is a case called Voidlink, where it was a ransomware group and the defenders teased apart the code and realized the code itself had been all written by, I think, Claude — one of the AI systems. Rather than a bigger ransomware operation, it was just a small number of people, maybe even one person, that had essentially vibe-coded a ransomware operation and was carrying it out. Even before Mythos, this was a capability that was coming online.

In some sense, I think we, society, crossed the Rubicon with Opus 4.6 in January or February, when Anthropic found 500 high-severity vulnerabilities. They weren’t as big a deal as what Mythos found, but it wouldn’t surprise me if we look back with the benefit of hindsight and say that’s when the exponential really started to take off, and Voidlink was in the mix for that, showing how non-state actors could pick that up.

We’ve become, rightly so for a lot of reasons, very focused on China as the peer competitor. But what happens when Russia or China gets their hands on something like this? For Ukraine, sure, the Russians are going to use it to bully and harass and attack them. But more broadly, it’s a really compelling espionage tool, which means you don’t use it to screw things up and make yourself known.

However, in focusing on great-power competition, we’ve as a result put counterterrorism on a deprioritized basis. In some sense, it was the fifth priority: China, Russia, Iran, North Korea, terrorists. The goal was keep number five as number five. The challenge is that the terrorists have all the incentive in the world to screw things up and make things be very disruptive.

Jordan Schneider: And you’re rounding down — or cartels or the Houthis. It’s a pretty broad group of non-state actors. And even North Korea. This is their game, right? Just trying to make money off cyber hacks.

Ben Buchanan: Scamming, yes, for sure. Beating the sanctions. I would distinguish the scamming and financial exploitation from more destructive and disruptive things that I think terrorists and probably cartels would have more reason to do than North Korea would. They’d want to use it for their own purposes, which is more about scamming the system.

The Next Turn of the Crank

Jordan Schneider: What does the next step of the exponential look like in six or nine months?

Ben Buchanan: I have to assume — it’s almost an article of faith at this point — there are vulnerabilities out there that Mythos does not find, cannot find, and that a better system would find. I doubt we’re at the ceiling of AI cyber capabilities. We’ve saturated every benchmark. Anthropic has reported it’s very hard to measure Mythos’s capabilities because it aces every single test or close to it. But I have to assume there’s a little more headroom, maybe a lot more.

One way to think about it — this is not original to me — is that all of the vulnerabilities Mythos finds are vulnerabilities that are immediately legible to a human once explained. Some are very clever, to be sure, but there’s not a lot of doubt about them. It is sometimes the case that a human historically could find a whole new class of vulnerabilities, where it’s a weakness that shows up again and again because we didn’t recognize it was a weakness. At first it looks alien, then once it explains it, it makes a lot of sense. Maybe the next generation starts finding more of these vulnerabilities that are less intuitive to us.

Michael Sulmeyer: From a cyber and AI development standpoint, that’s absolutely right. When you step back — how do most citizens look at or relate to this technology? I think most of the citizenry looks at it and says, “Why do I keep getting these preposterous phone calls asking me for my credit card number, or some prince in Krablokistan is offering me $5,000?”

It’s very possible that at that level, they don’t perceive much of a change. Those who get their hands on Mythos and want to cause abuse may have bigger fish to fry, but the ransomware gangs and the scammers keep at it. And on defense, we don’t look at stopping the scammers that hit vulnerable populations like senior citizens. So I think there’s a big segment of the population whose life is still being severely irritated by cyber scammers one way or the other.

That’s where we get to the defensive side of the ball — where can we use Mythos to raise the bar for what it takes for an attacker or scammer or spy to achieve their objective? Frankly, I think it’s an open question. The Glasswing thing is a really noble undertaking, and I commend everyone involved, but the press release is not the point. The point is, if we sit here in six months, have they patched 10,000 or however many high-severity vulnerabilities in the collective ecosystem, and has that actually had the effect of meaningfully raising the bar for intruders? That’s a very open question. That would stave off a crisis. But may not have any real impact on senior citizens.

Ben Buchanan: If the phones aren’t secure, then it doesn’t matter.

Jordan Schneider: Speaking of senior citizens, we’ve had all these hospital ransomware things over the past few years. If NYU Langone every two weeks is getting another “pay us $500 million or we’re going to delete your entire system,” then this is a very real thing. Cyber extortion was a sexy news story and a few school systems and hospitals got screwed up, but this was not a society-shattering trend over the past 10 or 15 years — in the way that once you have this proliferated, it may end up being.

Ben, you mentioned this idea of new classes of exploits, things that aren’t legible to human beings. Is there a theoretical limit? Is it possible to be sure that code is secure? I guess the answer is no, right?

Ben Buchanan: No, the answer is actually yes. There’s a branch of computer science called formal methods, which essentially gives a mathematical guarantee, a provable guarantee that a particular piece of code is secure. Now, right now we cannot do very much with formal methods — they’re fairly limited. But I can imagine we are sitting here in five years, 10 years, who knows, maybe given AI acceleration even less time, and we’re shipping secure code because we have used formal methods and AI can help with that. That is a possible end state. That’s a very desirable end state that I would love to get to. We are nowhere near that right now, but at the mathematical limit, it leads us to formal methods.

Norms, Bio, and What Comes Next

Jordan Schneider: It’s interesting that the most dramatic AI national-security application — well, I guess targeting is kind of TBD — is cyber, which is a space where it’s kind of no-holds-barred sitting here in 2026. Whereas if it was something around bioweapons or chemical weapons, maybe because there was already a global norm that this stuff isn’t cool, there might have been a richer path to have an international dialogue on potentially not exploring this tech tree. Given that the big one is coming in cyber first with AI, what other thoughts or implications do you have?

Ben Buchanan: I think we are very fortunate that cyber is coming first. I think we should use cyber as a lesson for what is coming next at the intersection of AI and other fields. Bio will not be far behind. At some point we will have a Mythos moment for bio. I’m not smart enough as a biologist to know what that looks like, but I’m confident that is the direction of travel. Maybe the norms save us — I kind of doubt it, especially when it comes to non-state groups, but who knows.

One lesson we should take away from Mythos is not “wow, this means AI is really good at cyber” — it’s that AI is really good. This is a general-purpose system that happens to be good at cyber. If you read the Anthropic system card for Mythos, it’s also really good at bio. I imagine the next version is going to be even better. There’s been a lot of debate for the last five years about how good AI systems are going to be. Obviously folks like me have argued for a very long time that they’re going to be very good, faster than people think. I’m biased here, but this feels like a pretty big piece of evidence that should update us towards taking AI risks seriously — in cyber, yes, but also in things like bio, because those are not going to be far behind.

Michael Sulmeyer: You mentioned norms, and the cyber community had a multi-decade effort to try to figure out what kind of international normative commitments could be made among countries about peacetime behavior. That was a noble effort. But I remember some Israeli colleagues telling me at one point, 10 years ago, “You missed the boat on starting a normative effort. You want to start the effort when you have enough of an advantage that the other side doesn’t quite see it yet, so you can get everybody to commit to maybe tying half a hand behind their back because people don’t quite see it as so detrimental to their own self-interest. You start too late, everybody’s so invested in trying to use the technology to pursue their objectives — it’s hard to get those kinds of commitments.”

A question I have — and I wish Joe Nye was still here to talk to about it — is: have we already missed that moment in AI? Not saying we should or shouldn’t be spending a ton of effort on a normative regime, just that if you’re asking about norms, that would be my question. Is it already a little too late?

Jordan Schneider: The thing with chemical weapons is they didn’t win you World War I. When it’s still kind of an open question, there’s a lot more excitement and incentive to explore the possibility space.

Ben Buchanan: It’s pretty clear to me the next wave to crash in terms of big societal national-security things is going to be bio. I hope people who are skeptics of AI look at Mythos and what it does for cyber and say, this should cause me to rethink my prior views when it comes to AI and bio.

Deepfakes, Persuasion, and the Information Ops Frontier

Jordan Schneider: Michael, at the very beginning of the show you alluded to the personal phishing-type cyber stuff. The last show I did with Ben, I tried to sell him on the US-China AI companion race and the potential implications — AI-powered case officers, recruiting spies, getting people to do things. AI being able to radically improve if you have a capability and people aren’t ready for it. Mythos doesn’t necessarily give you the video call with your mother — that’s probably the true frontier. But I’m curious for your general thoughts on how AI is going to ramp up that human-relationship-establishing side of convincing a soldier not to fight in a war, or someone to give you their secrets, or a country to revolt against their leadership.

Michael Sulmeyer: It’s a great way to talk about that nexus between cyber and information operations. If cyberspace is the delivery system — the way of getting to the information operations — then what does the message say? What does the content look like for the purpose of trying to convince you to not do something, or to do something?

You didn’t need Mythos to see how much more convincing deepfakes were becoming. And there’s an international-security dimension, but there’s also a very at-home dimension. Post-government, I’ve been helping K-through-12 school districts look at the kinds of new security threats and challenges they’re facing. Deepfakes of students against other students, by students against teachers — it’s scary how this is playing out because it’s so convincing. It’s very difficult to have a technological solution, to have AI figure out if that’s an AI-generated message. And so the opportunity for known human validation and follow-up requires a level of discipline and process that I’m not sure our institutions have really developed.

Ben Buchanan: It’s deepfakes for sure in the images and audio and video, but one of the really surprising things to me about AI — and I’m sure Mythos is good at this too, and frankly the whole class of systems including Google and OpenAI — is just how convincing they can be with text alone.

This was the last academic project I did before I went to the White House. At Georgetown, we used GPT-3 — before GPT-3 was released to the public, an early version — to see if it could persuade people on two political issues. One was, should the US be more aggressive towards China? The other was, should the US withdraw from Afghanistan? This was the summer of 2020. It could measurably, in a statistically significant way, write single-shot text messages that would change people’s minds. That was 2020. If you look at what’s happened since — a Nature study, a Stanford study — it’s pretty clear AI systems have only gotten better.

The one that’s so striking to me — I’m going to butcher some of the stats, but this is pretty close — is there’s a subreddit called Change My View, where people post an opinion and it awards points (I think they call them deltas) to folks who give compelling counter-arguments. Some researchers used an AI system in 2024–2025 to post on Change My View. I think it scored in the top 1% of earning these points and changing people’s minds.

We’ve strayed a little from cyber operations as narrowly defined, but it gets to the broader point: in a renewed competitive information environment, an AI system can be useful for a wide range of aspects of national competition — cyber operations on offense, cyber operations on defense, but also the adjacent category of information operations.

Closing: Bureaucratic Uptake and the Race Ahead

Jordan Schneider: Michael, you brought up information operations, which folks will debate, but I think have been somewhere between ineffective and national embarrassment over the past few decades. Maybe we can close on bureaucratic uptake for these tools. It may be hard to hire someone who speaks Tagalog well enough to push narratives into the Philippine political system. But if all you have to do is press a button, it starts to be a lot easier to do some of this. Aside from hardening systems, is there anything different from the slate of recommendations you were pushing during the Biden administration versus what you’d want to give the US government in a post-Mythos era?

Ben Buchanan: Jordan, you know this, but near and dear to my heart is building an American lead and a democratic lead in AI. This whole conversation we’ve had — the cyber dimensions and everything else — reaffirms the importance of doing that. Obviously, that’s an export-control conversation, a domestic-investment conversation, an infrastructure-buildout conversation. One of the things we tried very hard to do in the Biden administration was to ensure democratic preeminence in AI, in part because we had high conviction that things like Mythos were coming, that this technology would be useful for national security and geopolitical competition.

Frankly, I think a lot of folks — don’t take my word for it — if you look at what Dean Ball wrote, a former Trump administration advisor, when he left the government, he commented on basically how non-AGI-pilled his colleagues were, how they don’t believe in a world of very powerful AI systems. I’m hopeful that for them and for other people, things like Mythos and the broader development of AI capabilities can be a reminder that we really want America to lead in this technology. And it would be so much worse for the world if China had this. I doubt if China had this, they’d be giving it to defenders first and making a lot of noise about the need to patch systems. They’d be doing exactly the kind of espionage we’ve been talking about.

Jordan Schneider: It’s an interesting Jensen proof point, right? He’s been talking for years about how this stuff is entirely revolutionary but has been pretty quiet about it from a national-security perspective, saying they’ll be able to get all the chips they need, it’s not going to change the world when it comes to the sharp end of what governments want to do. And here’s another proof point against that. Michael, take us out. Any closing thoughts?

Michael Sulmeyer: You have to hope that the leaders of America’s war fighters know that this is a technology they have to adopt to really make sure that our offensive cyber operators maintain and extend a competitive advantage. As Ben said, this is important to win, and this is an important tool to extend that advantage for the nation in cyberspace.

A large debate that probably has not played out enough on defense is what kind of autonomy our leaders are comfortable with for a model to run for cyber defense on sensitive military networks. It doesn’t really work to have the model merely alert a human that there might be a problem. And yet you’d be right to have some caution about just turning the keys over to the model to say, “Hey, keep us safe.”

I worry that between the CIOs and these kinds of bureaucracies, the instinct is to maintain human accountability and not disrupt the business process. But increasingly viewing it as an operational matter, as a contested domain, where you have to put more weight on autonomy to defend faster — that’s where I worry there’s a really tough conversation coming, and there’s going to be risk that has to be taken to lean on AI to keep us safe.

Jordan Schneider: That was excellent. Thank you two so much for being a part of ChinaTalk. Is this WarTalk? It’s ChinaTalk. Is this ModelTalk? I’ve got so many verticals now, it’s horrible.

Ben Buchanan: Jordan, I told Michael that your show used to be called ChinaEconTalk. And then you dropped the econ and it became ChinaTalk. And now you do so much stuff on AI and other things — you’re just going to be Talk.

Jordan Schneider: We’re just Talk Talk. We’re Talk Talk.

Toll is the only historian I’ve read who is so good with his material that you’re not compelled to read more about the subjects he raises because you’re that confident that’s he done his homework and surfaced the most interesting angles.

Iran War parallels abound!

Once the Algerians decided they weren’t getting paid enough tribute and hijacked an American boat, “Vessels bound for the Straits of Gibraltar could not be manned at any rate of pay. Maritime insurance premiums doubled and then tripled, U.S. government bond prices collapsed, and merchant houses were bankrupted.”

Adams and Jefferson discussing whether or not to fight the Barbary states or give tribute:

Adams did not believe the American people or their leaders were ready either to rebuild the navy or to fight a war in the Mediterranean. “We ought not to fight them at all,” he wrote, “unless we determine to fight them forever. This thought is, I fear, too rugged for our People to bear.” The more likely outcome, Adams predicted, would be that the United States would fight for years at great expense, only to pay for peace in the end.

Hegseth wishes that this worked for the Iran War:

America used to give other countries shiny things to bribe them to leave us alone. How the tables have turned!

Oh and we even got some torpedoes and undersea mines that the Brits during the War of 1812. The established powers though thought they were dishonorable (because anything that challenges what works well for you always is).

Paul Revere was in the strategic industry of copper sheet-rolling. Also, two other guys did the ride with him. “History would forget two other men who rode alongside Revere: William Dawes and Samuel Prescott. Their names did not rhyme with ‘hear.’”

Jefferson tried to DOGE too:

Dueling between bored adolescent Navy officers was a big deal. “The most trivial disagreement was liable to trigger a challenge. One midshipman was offended when another entered the wardroom wearing a hat. Another challenged a messmate because the offender had spilled some water on a letter he was writing. A pair of midshipmen nearly dueled after arguing whether a bottle was green or black.” But here is perhaps the dumbest one in history, involving legendary Stephen Decatur.

Boat captains challenged each other to boat duels.

Did we have a better regime strategy for Tripoli in 1804 than for this Iran war?

Lots of child soldiers everywhere! Up to a quarter of the crews on these frigates were boys.

At night when you pull up to a ship and you don’t know what country it was from apparently you just ask, and if you don’t like the answer you start to shoot:

These battles were horrific:

And then you boarded the other ship and started cutting people to pieces with cutlasses.

The UK had export controls to colonies on arms manufacturing: “Britain had forbidden the manufacture of heavy cannon in the colonies, and there were no domestic foundries capable of smelting, refining, and casting big naval guns.”

America’s big innovation was to build longer, meatier frigates that could defeat British and French frigates and run away from their heavier battleships. Joshua Humphreys alone came up with this new in-between class of boat, and had to convince people with drawings and vibes rather than any fancy simulation software that this was the way.

Instead of centralizing production, the US farmed out their manufacture to six shipyards to gain more support for the Navy. Some states went overboard. Everyone wants giant masts and giant guns until you actually have to sail with them.

To incentivize the captain and sailors to live dangerously and not become pirates, they were paid a percentage of the ships they defeated and were able to salvage.

Fighting was very much a skill issue. Boats with fewer guns could, if trained well, fire twice or three times as fast as adversaries. Human capital came into play—empressed Americans weren’t particularly eager to get shots off from the British boats they were forced to sail on.

Dueling

Abigail Adams, when stationed with her husband in London, was not a fan of the child street fights:

he was appalled by the boxing matches she witnessed in the streets of her neighborhood, where she had “been repeatedly shocked to see Lads not more than ten years old striped and fighting until the Blood flowed from every part, enclosed by a circle who were clapping and applauding the conqueror, stimulating them to continue the fight, and forcing every person from the circle who attempted to prevent it.” She associated the brutality of the street hooligans with the invective of the English newspapers. “Bred up with such tempers and principles, who can wonder at the licentiousness of their Manners, and the abuse of their pens?”

Thanks to ChatGPT for surfacing a whole thesis on the ‘plebian honor fight’ that played out in contrast to the gentlemen’s duels and prize fights.

Dismissing developments that don’t align with your worldview has a storied history! Didn’t need a deepfake to convince Jefferson that everything was going great with the French Revolution even after the mob really got going.

Eric Robinson, , , and join me to score the Iran conflict.

We discuss…

• Whether Iran’s Strait of Hormuz toll booth is a Trump card or a wasting asset

• How the administration fumbled the messaging on the war’s most heroic moment — the JSOC pilot rescue deep inside Iran

• America as Prussia 1806: the great military machine that can’t learn new tricks?

• Colby’s bizarre knife fight with Pope Leo

• McMasterism, dereliction of duty, and why no one is pushing back

Listen now on your favorite podcast app.

Who Won the War?

Jordan Schneider: Who won the war? Here we are. WarTalk, rebranded from Second Breakfast. We got a special guest, Secretary of Defense Rock — hereby dubbed Mr. Secretary — along with regulars Eric, Tony, and Justin. Should we take a vote? All in favor of Iran winning? What’s the judge’s scorecard here?

Eric Robinson: We got to set out a rule. What determines victory? Did you realize your ambitions that you set out at the start of the conflict? Do you come out of the conflict stronger or in a better negotiating position? Do you retain your combat power?

Justin: The only people who won out of this were those who held long-term commodities futures. On the Iran front, yes, it’s clear we didn’t win. It is very clear the United States didn’t win. Best reporting from the New York Times and elsewhere says Iran’s economy is only a few weeks away from probably fully collapsing without some sort of aid and income. Fiery loss? I don’t know.

If they really were a few weeks away — let’s say they were a few months away from total collapse of their economy prior to this conflict. Obviously there was a lot of damage to the economic infrastructure that occurred during the conflict. All that did was speed up the economic collapse, but now you’ve potentially given them this ability to tax. I saw the report today in Forbes — something like if they did the toll on the toll booth at $2 million a ship, that’s like $9 billion a year at normal transit numbers, which is not chump change. So it would be a victory that potentially hands them an economic win unless we’re going to go back later and actually force the strait and not enforce the tolls. That’s one of the reasons I lean towards Iran is in a materially better place as a regime at the end of this.

Tony: Which regime though, right? Is the regime that now exists still the regime that exists in a couple of weeks? Because if the IRGC is particularly upset with this deal — their idea of victory is very different from the regime’s idea of victory — then it’s in an even worse position. I’ll also say, talking to some investors, the VLCCs, the very large crude carriers, those that go through the strait will do fine because it’s probably only a $1 increase per barrel. Anything short of that, and you are looking at substantial impacts to profit and revenue. And when you’re talking about a reduced flow anyway — we’ve seen it with the Red Sea, there are fewer ships going through — this has longstanding effects if they allow that toll to go through.

Secretary of Defense Rock: The big winner is Robert Pape and probably book sales for Bombing to Win if I had to pick one. The way he’s kind of broken out on Substack — it is pretty remarkable.

As far as the amount of munitions that were expended, the amount of assets we had to bring in that were used — again, it’s sort of just what was actually accomplished. When military officials are talking about how many meals and energy drinks are consumed and how many targets we hit, it’s like, what are we doing? It feels like this throwback to Vietnam — we hit this many targets and...

Tony: We joined the war on substance addiction on the side of substance addiction.

The 2018 NDS Scorecard

Eric: I want to establish a benchmark because we are speaking through the last six weeks of arguably high-intensity combat in the Persian Gulf. If we go back to 2017 with the Trump administration’s first National Security Strategy, followed by the 2018 National Defense Strategy — this is a principal document that reset American defense away from the excesses and the meandering of the global war on terror and reoriented it towards four large nation states: the People’s Republic of China, the Russian Federation, the North Koreans, and Iran. And in the intervening eight years, two of those four have suffered catastrophic battlefield reverses — the Russians in Ukraine and now the Iranians in their own skies and on the seas.

Did we inadvertently accomplish goals of the first Trump administration’s National Defense Strategy? I’m obviously hedging because the strategic circumstances have shifted — we’re post-COVID, we’re in a different world. But among those big four, are those potential disruptors to global order in a better situation now? Or are they in a fundamentally weaker position than they were at the outset of this period? I don’t know that the 2018 NDS is the right measuring stick, but if we could get Bridge Colby on here, I would certainly like to ask him. Just to give us an eight-year period — when the United States came out of this time of trying to patch together Iraq and Afghanistan, reoriented towards nation-state competition, now we sort of know what it looks like. And for better or worse, we have a couple of case studies.

Justin: They chose to abandon that with the most recent NDS. So if that becomes a fulcrum by which they’ll say, look, this is victory because we said this back in 2018 — they had given that up. The other thing I would say is when we look at the messaging, that’s one of the really important things this administration didn’t do a good job on. The Department of Defense, the whole U.S. apparatus kind of failed pretty hard on the messaging part, in part because we didn’t have really clear goals and expectations at the beginning of the conflict. But even as we continued, we came to a point where people were openly questioning within the United States if what the president said was right or if what Iran said was right — which, given everything that’s ever happened with this administration, that’s a new thing even for the Trump administration.

They lost the ball on the messaging in a very real way. And then we have them openly talking about potential Title 50 action to help protesters who were trying to overthrow the regime in Iran during the initial uprising — potentially sending weapons and support to these protesters. That fundamentally validates everything the regime has said since 1979. Whether it did happen or not, whether it was just something he said in a one-off, they validated the legacy of the narrative of CIA intervention in Iran since 1979.

Eric: And talking constantly about oil reserves didn’t help either. It’s perfect advertising for bin Ladenism.

The Moral Low Ground

Jordan: Tuesday we had Art of the Deal, Genocide Edition, right? It’s sort of unimaginable that we could end this on the moral low ground against the regime that just shot 30,000 people two months ago. But he did it. Hats off to him. That’s a high degree of difficulty move.

Eric: It’s depravity. Justin, I think it’s an important framework to think through this about messaging — how do you communicate in times of war? In previous episodes, we’ve talked about how warfare is a form of bargaining and diplomacy is a component of this, next to armed force. When the president communicates through social media, he’s communicating not to one audience but to a multitude — Iranians who are pro-regime, Iranians who are soft, Iranians who are anti-regime, American forces in the field, countries in the crossfire, and the American public. Calibrating your messages so that it impacts those different audiences the way you want is extraordinarily difficult and it’s not always effective.

There’s a degree of recklessness about the way this administration spoke about their most fundamental responsibility, which is warfare — with the president dangling his normal sensationalism on Truth Social, the Secretary of Defense talking about this being a firm religious obligation, the pilot who was recovered being emblematic of Jesus returning to life on the third day, and then Raisin Cain talking about soldiers crushing MREs. It is difficult to determine what is the actual signal. And to Jordan’s point, the president did break through a substantial amount of noise when he said, I’m going to wipe out a civilization — language you would leave for Nikita Khrushchev in the heights of the Cold War. To hear it from an American president in the 21st century is jarring, and it discredits every one of us.

The Rescue That Got Lost

Tony: The pilot going down — I’m very glad, well, pilots, I should say — I’m glad they both got rescued. That unfortunately did not break through to normal social media feeds the way that the “civilization is going to end” broke through. From my normal friends that are not attached to national security — and I’ve heard this from multiple people — they didn’t know there was a pilot shot down. They saw the tweet though. And if there’s anything we learned from the last couple years, the message that gets through is sometimes more important than the actual statistics. Even if we did achieve all of our operational-level military objectives, that’s not what everyone’s hearing.

Secretary of Defense Rock: At least we found the off ramp. It’s a strange and horrifying way to get there, but we did. I do wonder — over the last couple of weeks, it seemed like Iranian munitions were getting through more easily and hitting targets more frequently. I wonder how much pressure there was on the administration externally — if you go all in, we’re screwed. If we really went through and started hitting their infrastructure the way Trump was saying, they would have gone scorched earth.

Eric: I think the Qataris, the Emiratis, and even the Saudis knew that the Iranian targeting procedures were increasing in sophistication. There was a bit of a dead hand — even if you conducted retaliatory operations, there was some element of the Iranian state that was going to hit you back.

Justin: I think Tony pointed out a really interesting thing about the messaging. What’s disheartening about the ending-of-civilization and the rest of the genocidal talk is that the military actually had a really awesome story to tell — we were willing to spend multiple aircraft, put people into harm’s way deep inside Iran to retrieve one human and bring them back. People are making up crazy stories about what actually happened because they can’t believe the U.S. military would care that much about one person.

Eric: We were retrieving a Stargate.

Justin: There’s been that, there’s been “they actually tried to go to Esfahan to extract the highly enriched uranium.” The truth is there’s actually a really good message there that goes way beyond the warrior ethos crap that’s been talked about — it talks about some of the core things about never leave a fallen comrade, about who we are as a nation and who we are as a military. And that just got completely washed over by everything that happened right after it. It should have been a triumph. It should have been something we held up and said, this is who we are. And we preceded that with the messages we did. It’s not even a loss at that point. It’s tragic.

Jordan: It was four hours and then we got “praise be to Allah.” Was it even four hours?

Tony: In a normal administration, you would have that — we sent in a whole JSOC team to go recover one guy deep inside enemy territory — and then you would contrast every story of the Russians shooting their own wounded because they don’t bother to recover them. That should be played every day. That would be the highlight reel at the NATO summit. But not in this admin.

Secretary of Defense Rock: Even just the zero hesitation — it was interesting to follow as it was happening. You see these videos, in broad daylight, very low, getting shot at by random policemen. Pretty remarkable. But it’ll probably be a footnote in the grand scheme of the public consciousness.

Justin: The only good thing is that SEAL Team 6 is the one who went, so there will be at least three books written about it. Of that, I’m positive.

The Toll Booth Is a Wasting Asset

Jordan: The toll booth is a wasting asset. There will be ways to go around the Strait of Hormuz, and they might not be super economical. But the one thing the Gulf has is money. Everyone’s talking about how Iran has this incredible trump card — yeah, they played it. And now because they played it, the global economy is going to end up adjusting.

Eric: It’s very similar to PRC using their rare earth schedule.

Tony: Can we pivot then to the May meeting, the alleged meeting between Xi and Trump? Whatever was on the docket, I assume it has to have changed.

Jordan: I think there was never anything on the docket, Tony. They’re just going to take some photos.

Justin: Biden-Xi San Francisco 2.0. Nothing of substance will come from it because Biden couldn’t remember his time. Take that.

Jordan: They banned AI from nukes, okay Justin? Come on, give them some credit.

Tony: Let’s be clear — we talked about agreeing to talking about banning AI from nukes.

Justin: Can we just point out that putting AI on nukes would be the most idiotic thing anyone could do? There’s a reason we still use eight-inch floppy disks to run those computers — it’s because we don’t want them on the internet. We don’t want them thinking at all. That’s the last thing anybody wants, the doomsday weapons to have thoughts of their own.

Are We Prussia in 1806?

Jordan: Before we go to China — I want to get Mr. Secretary’s “Are we Prussia 1806?” take. Can we run that one back?

Secretary of Defense Rock: This was a piece I had written in August of 2025. Clausewitz is most known for On War, but he was a ferocious writer — hundreds of pages of letters, essays. One of his most interesting essays, which was the basis for my piece, is called “From Observations on Prussia and Her Great Catastrophe,” which he wrote around 1823–1825. He was trying to figure out how this state — an army that Frederick the Great had barnstormed around central Europe — just collapsed like a house of cards in its third battle against Napoleon.

There were a lot of parallels. It was eerie the way he talked about how nihilism captured Prussian society, how the elite culture had atrophied and was only thinking about themselves, how the Prussian military became obsessed with the way their rifles were cleaned, the way they marched in order during a parade, how many awards went on a jacket. The quote that stood out most to me was he wrote that “vain and moderate faith in these institutions made it possible to overlook the fact that their vitality was gone. The machine could still be heard clattering along, so no one asked if it was still doing its job.”

It feels like that’s the way our government and civic culture functions today — this big machine that moves along while the cords get ripped out one by one out of these agencies. And at some point it gives. I re-circulated it today kind of randomly — is this conflict our Jena-Auerstedt?

History Does You

America as Prussia in 1806

A vain, immoderate faith in these institutions made it possible to overlook the fact that their vitality was gone. The machine could still be heard clattering along, so no one asked if it was still doing its jo…

Read more

9 months ago · 281 likes · 23 comments · Secretary of Defense Rock

Eric: Prussia gives us an interesting lens. Clausewitz is in the shadow of Frederick the Great — arguably the most capable battlefield commander in Central Europe during the War of Austrian Succession. Using a small but nimble and exceptionally well-trained force, he beat Austrians, beat Russians, and took Prussia from an economic backwater into a major European power. The Prussian elite sort of missed the ripples of Europe that came out of the French Revolution. The gambit that all European states faced in 1770 was substantially different by 1795. The ability of the French to mobilize en masse — the visage of a Napoleon who assembled a force of 100,000 men on the Channel in 1805, marched them halfway across Europe, defeated the Russians and the Austrians at Austerlitz, achieved a favorable peace, all while the capable Prussians continued to elect not to react. There is a lesson here — the Prussians were looking favorably back upon a golden age of military excellence. And that golden age, maybe similar to our 1991 war, was very real and authentically stunning. But at a certain point, shutting up and playing the hits doesn’t get the crowd out in front of the stage anymore.

Tony: What stood out to me was the analysis that Napoleon didn’t necessarily innovate in warfare — he mastered that which was available to him in a way that others had not. If you look at our slow reaction to counter-UAS, our slower reaction to triple mass, I would say those are warning signs that that’s where we are.

Justin: Napoleon basically said, you can’t beat me. I spend 30,000 lives a month as a matter of principle. You’re not willing to do what I’m willing to do to win. Where the parallel really comes in is when you think about an enemy willing to take those losses. We talked about this in the last show — the way our risk tolerance changed with our technological dominance. Do we have the stomach for when those things change to still be a great military power? It’s a frightening thought.

Tony: I’m going back to conversations I had this week with some senior folks who remarked that looking at the Navy today, they don’t have the tolerance to own the littorals. I don’t think that’s entirely true from an operational perspective — there’s a very good reason why you don’t want a carrier group to do a show of force through the strait and eat a Silkworm or 10. But that is a perception that is out there. A lot of the reactions we as Americans, as leadership, have had in the last four to six weeks feeds into propagandistic narratives among our rivals. For people who live in the system, who run the system in Beijing, this validates a lot of their perceptions of American culture and approach to casualties in the modern era.

Justin: Does the fact that we were unwilling in this instance to risk the capital ships to force the strait raise the risk of the assumption from the enemy becoming, “well, they’re not willing to risk their ships, so they’re not going to send them”?

Tony: I think it’s a little different. There’d be one case — if we pulled all our boats out of WestPac, but we’re not going to do that. There’s a question of whether the Americans are willing to take initial casualties and then say, we’re going to fight on. In their minds, it validates the idea that if they hit a carrier or level Guam, that tolerance is too damn high.

Secretary of Defense Rock: We didn’t take Kharg Island. We deployed the troops, but we didn’t actually follow through. If you were Iran or China, that would be my takeaway — they’re willing to take the steps but not actually follow through.

Tony: I want to make clear I’m not making the case that we should have done a thunder run to Tehran. But it is how our enemies think. And I would point you to Ukraine for the last four years and ask whether our enemies are 100% brilliant. But they do adapt. They do read signals. From the reporting in the New York Times, Washington Post, and elsewhere — the Russians were helping with Iranian targeting. The Russians learned that overhead ISR goes a long way to making people feel pain. There were plenty of cases in the Ukraine War where we thought we had the trump card technologically, and then the Russians survived just long enough to change their tactics.

From Jena to Waterloo

Eric: In 1806, Prussia is decisively defeated at Jena and Auerstedt, the state is shocked. But by 1814, Prussian troops are reassembling, contributing to an allied move against Paris, helping compel Napoleon into his first exile. Marshal Blücher takes an army of mostly militiamen and helps finish off Napoleon at Waterloo. A couple decades later, 1866, the Prussians beat up the Austrians at Königgrätz. And several years after that, the Prussians returned to Paris in grand fashion. The darkest depth of Prussian experience is the moment you cited in your essay. But they came out of it quite quickly and aggressively. What are the lessons for us? Is it truly darkest before the dawn?

Secretary of Defense Rock: The Prussian military didn’t really reform as much as they copied. They copied Napoleon, they changed to a corps system. They had started coming up with their famous staff system but hadn’t fully developed it. They were able to harness the revolutionary, nationalistic fervor that they rode all the way to Paris. One of my favorite quotes — after the Battle of Lützen, which was basically a draw, Napoleon remarked to the Prussians: “These animals have learned something.” But they hadn’t really learned something as much as they just copied it and were willing to stand their ground and fight.

I sort of remarked at the end of that essay — maybe it just doesn’t happen. The United States has so much economic power, maybe there’s a way to ride this out the way the Prussians did. Clausewitz always said you can learn from history, but that isn’t necessarily a prediction of what’s going to happen. There is always a bit of hope. But the alarm signals — the longer this goes on, the harder it is to pull yourself out. It happened so quickly for the Prussians that they were able to ride that wave. I do wonder how much longer this can go on, this sort of rumbling along, before it bites you and knocks you out forever.

Jordan: I have another “On War” take. I was reading Machiavelli’s On War the past week or two. Kind of trash. Sorry, Machiavelli. The one piece I liked — he was like, when you’re training your troops, don’t say “turn,” say “left” or “right,” because otherwise everyone’s going to turn in different directions.

Tony: Machiavelli is definitely the lesser of all the theorists.

Justin: Jomini is in front of him, which is saying a lot.

Eric: I liked Sign of the Times and the Batman soundtrack.

Tony: On a side note, there’s this perpetual narrative around Sun Tzu being too simple. I’ve always been on his side because I’ve watched so many officers forget the things they say are so simple. Do you really need complex theory?

Eric: And fight a battle in the swamp?

Jordan: Kharg Island was dry. Eric, sand — it was right there.

Eric: Don’t gobble proffered baits. I remember John Lewis Gaddis not quite shouting that at me, but saying, what do you remember? If you see something obvious that you’re supposed to go and take down, don’t do it, you foolish princelings.

Justin: I can never find the quote, but Napoleon wrote something to the gist of: don’t do what your enemy wants you to do. It’s enough to know they want you to do it. That’s a good, well-reasoned way to approach life.

Jordan: Justin, I found your actual quote. The 30,000 one is apocryphal, but he had a conversation with Metternich in 1813 where he says, “a man such as I am does not concern himself much with the lives of a million.”

Justin: That’s a guy who’s going to change the way the European feudal states wage war if he’s willing to just march the entirety of Europe.

Eric: Levée en masse. People like to dress up Austerlitz and a couple of other engagements because there are these magnificent maneuvers. But a lot of it was just the steady application of grapeshot and illiterate infantry. His infantry attacked in column, not in line. The British infantry were known for being a little more sophisticated — troops operating a steady rate of fire, with a large frontage, able to hold a lot of territory with a smaller number of men. The French would just have a bunch of oafs charging in a column. If the oafs in front fell, you walked over them. It was simplistic, but it was elegant. And it worked. That’s why Napoleon seized Moscow — an extraordinary accomplishment based off some really basic geometries of war.

Justin: Last thing on Napoleon — not ever, but at least on this little bit. The Napoleon movie was terrible. I turned it off 30 minutes in.

Eric: What happened, Ridley Scott? How do you do Gladiator II and Napoleon back to back? You lose your fastball at some point — you’ve got to hang it up if you’re sending that out.

Eric: The Waterloo movie with Rod Steiger and Christopher Plummer is on YouTube. It’s not perfect cinema, but it’s epic filmmaking. They got something like 65,000 Red Army regulars to appear in it. Rod Steiger chews the scenery. It is pretty good.

The Colby Catechism

Jordan: We’ve got to close on the Undersecretary of the Papacy.

Eric: That is probably the most unexpected diplomatic feud of the week — between someone who should not be necessarily involved in diplomacy, the Undersecretary of Defense for Policy, one Elbridge Colby, and the Pope, the Church of Rome.

Jordan: You think you’d be excited — it’s the first American! It’s cool! Come on, can we just leave it at that?

Eric: Somebody who — the Pentagon and Pentagon officialdom do tons of point-to-point encounters, typically with foreign militaries. They do international travel, go to the Shangri-La Dialogue, go to the VFW jamboree. Some Pentagon officials are better at it than others. But Elbridge Colby’s got the reputation of a guy who got pantsed by a bunch of cartoon dogs online and started swearing that everybody hating on him on Twitter were Russian agents. He had a really epic crash-out at one point, but now he’s in a position of real responsibility. And he, according to reports, got into an ecclesiastical knife fight with representatives of Pope Leo.

Jordan: So the reporting is that he basically said, you’re nothing, you don’t have an army. By the way, in the 1400s, there was an anti-Pope in Avignon, so don’t get ahead of your skis on this one. By the way, I went to that Pope zone on my honeymoon. It’s all right — it’s no Vatican.

[fwiw FT is now reporting that it wasn’t Colby but someone else…]

Eric: It’s fascinating that Colby effectively paraphrases a quote attributed to Stalin. In some moment of frustration toward Yalta or the Potsdam Conference, there was some whisper that the Pope was worried about the plight of Catholics in Eastern Europe. And Stalin scoffed and said, the Pope — how many divisions does he have? That’s what the Undersecretary of Defense for Policy intimated.

Tony: In China policy circles, there’s been a gripe for years that the Catholic Church bent the knee a little too much with the Chinese Communist Party in terms of the CCP getting veto over which priests could serve. Given the legacy of the Catholic Church under Pope John Paul II fighting communism and going to Poland, it left a bad taste in a lot of people’s mouths. If that had been the fight, I would have said, probably not the time, but I get it — that’s a long-standing issue. But this is something entirely different — basically hinting that you might Maduro the Pope.

Secretary of Defense Rock: I’m coincidentally reading David Kertzer’s The Pope at War, having previously read his The Pope and Mussolini. My one takeaway is the Catholic Church is very sensitive about its image globally and the way it interacts with nation states. The idea that you’re going to go in there and chastise representatives and they’re just going to roll over — what is the thinking?

Eric: What’s the goal? The most precious resource any senior official has is their ability to focus. They are heavily booked from 6 a.m. to 10 p.m. if they’re discharging their responsibilities ably. To set time on fire over the presence of a 14th-century anti-Pope is extraordinarily bizarre behavior.

Justin: The Catholic Church has done plenty of terrible things in its history. But in the modern era — let’s say since Pope John Paul II forward — it has at least attempted to care about its flock. Saying harsh things like “there shouldn’t be war” is not really a stance the Pope needs to be chastised for. That’s kind of his thing, man.

Jordan: Who are we more angry at, NATO or the Pope, for their Iran war takes?

Dereliction of Duty 2.0

Secretary of Defense Rock: It’s funny to think that Colby is supposed to be one of the serious ones. He got bipartisan support to be confirmed. And you’re fighting with representatives of the Pope?

Tony: Colby was supposed to be the serious China thinker — which was only the case if you didn’t know what you were actually talking about on China. If you thought being a China hawk meant pointing at China and only that, then yes, he was the serious guy. I know people high level in this administration at the staff level who thought what he wrote in his book — about maybe we just have to proliferate nukes to deter the Chinese — wasn’t serious. But they’ve now spent the last 10 years telling themselves that nothing is actually real. And that is how you get officials who do things like threaten an anti-Pope. In any other place, you wouldn’t have been there in the first place. But people just say crazy shit because that’s what gets you in that role, and turns out, yeah, you are going to do that crazy shit.

Jordan: From the Maggie Haberman article, the tick-tock on how the U.S. went to war — towards the end, she basically goes around the Situation Room getting everyone on the record. “Stephen Cheung gave neither a yes or a no, but said that whatever decision Mr. Trump made would be the right one.” That’s the world we’re in.

Justin: She has a really good point in that article about the difference between General Caine and General Milley: Milley saw his job as preventing the worst impulses of use of military force. General Caine sees his job as only providing best military advisement.

Jordan: And then there was that wild bit:

One person familiar with their interactions noted that Mr. Trump had a habit of confusing tactical advice from General Caine with strategic counsel. In practice, the general might warn in one breath about the difficulties of one aspect of the operation, then in the next note that the United States had an essentially unlimited supply of cheap precision-guided bombs and could strike Iran for weeks once it achieved air superiority. To the chairman, those were separate observations, but Mr. Trump appeared to think the second canceled out the first. At no point during deliberations did the chairman directly tell the president that war in Iran was a terrible idea.

Justin: I could not for the life of me figure out why they held a press conference saying we’ve achieved air superiority over Iran. When you have air superiority, everybody knows.

It shows they had a checklist. These are all things the general said, and they checked them off. When you hear it framed that way, you realize that’s not actually what he was doing. He was just giving you what the military is capable of and where we’ll struggle. At no point is that a values judgment on whether we should use it. The danger is if you’re not surrounded by people who are going to have that hard conversation, it makes that general’s position — being the guy who will push back — much more important. And as you just read, they’re not exactly surrounded by people who are going to push back.

Secretary of Defense Rock: This is Dereliction of Duty all over again — H.R. McMaster’s book exploring the lead-up to the Vietnam War between the JCS, McNamara, and Johnson. McMaster’s conclusion — dubbed “McMasterism” by civil-military experts — states that if the president and politically appointed officials are pursuing policies not in the interest of the United States, you have a professional duty to speak out directly to the public or Congress. That hasn’t happened. I think it shows how that’s just not a thing. That book was on reading lists for general officers for decades. It’s a perfect encapsulation of how apolitical professionalism — framed in this way where Caine is purely giving tactical and strategic advice and that’s his only job — becomes a dodge. Officers in the past knew you actually should give political judgments.

It’ll be fascinating to see, especially in the context of Randy George recently being fired, how this continues. CQ Brown hasn’t really said anything. The SOUTHCOM commander may have resigned because of tensions. Nobody says nothing. McMasterism is fraudulent, in my opinion.

Tony: You’re asking them to do something they never did for the first 40 years of their career.

Eric: Nobody makes GO by telling the principal that the principal is wrong. You can’t do it in public because there’s always going to be some hanger-on with a yes answer. You’ve got to have the self-confidence to say, I don’t need this job.

Eric: I know General Caine — worked for him. I often think he is vamping. He knows it’s ultimately pointless. He’s been thinking about war for a long time. He was an interesting pick. If he’s talking about MREs, if he’s talking about all of these statistics, I think he is trying to get to the end of the press conference so something else catastrophic doesn’t happen. Maybe I’m giving him too much credit. I think he is ideologically on board — I’m not saying he’s some secret defender of the Constitution. But I think he is also trying to exert a degree of professional discipline. And if he is speaking, he is preventing others from also speaking.

Seven Months to Election Day

Tony: We’ve got about six months and three weeks until Election Day. Does this conflict — look, we’re just going to ceasefire here, right? The war is not technically over.

Eric: The Israelis are still conducting a substantial air campaign in and around greater Beirut and southern Lebanon.

Tony: And what happens if there are mines in the strait and they hit a boat? Can we get another conflict in before the midterms? I don’t mean that to sound silly, but it’s very clear they were on a roll and had designs for other things. How does this impact their thinking going forward for other campaigns?

Jordan: There’s no oil in Cuba. Gas prices will be just fine.

Justin: Right after it was going well — “job accomplished, we didn’t need anybody, we told you guys we could do it” — we started seeing comments about Cuba. Then when things started turning and it became “hey, is anybody else going to come do something about this strait?” — we stopped seeing comments about Cuba. Maybe we’re seeing a recognition of the limits of military force during this form of the administration.

Secretary of Defense Rock: I thought the comment about wanting some type of NATO support for reopening the strait — it’s like, on one hand that, and on the other it’ll be some joint venture where Iran and the U.S. share crypto for tollways or something. For the long term, if this holds and Iran gets this tollway, maybe you work around it. But is a future administration going to have to expend political capital to reestablish freedom of the seas?

Tony: There are relics we can point to. During the Arab-Israeli wars, the Suez Canal was down for years at a time. There are periods of recent history where we’ve had to deal with things like this, and they don’t get resolved — they get forgotten by other conflicts. People learn to deal with them. For the administration, how do we get out of this as fast as possible in a way that doesn’t look like we ran away? From a strategic perspective, that might be the first sensible strategic thing they’ve done — the exit is more politically valuable than the actual negotiating results. The American voter does not care whether Iran still has enrichment. They care whether things like this keep happening. For the rest of the world, these are very serious consequences that are not just about a midterm election.

Justin: So we sent an email to Randy George — at least his old CFR account. It didn’t bounce back. So potentially General Randy George coming on the show?

Jordan: Trump earlier today labeled Tucker, Megyn, Candace, and Alex Jones as all third-rate podcasters. The question is: will he want to really start with the first-rate WarTalk? Or is he going to want to build up — enter the field with a third-rate run, then go to second rate, and then hit the stars?

Eric: We’re a notoriously tough room. He’s a long-time listener, first-time caller. We are easy to hunt down — WarTalk on Substack. “I assure you, I am the former chief of staff of the Army and I would like to come and riff.”

Justin: I appreciate having guests who have two first names because I never feel like I’m messing it up when we introduce them.

Tony: There’s always something satisfying about being able to call a senior officer by their first name. Like, hey Randy.

Justin: As a former warrant officer, I just did that routinely because nobody knew what to do with me.

Jordan: I think we got the Colby Catechism as our song. What genre is it?

Eric: It’ll be a Gregorian chant.

Lyrics:

Open source AI in China did have a moment. Reeling from the shock of ChatGPT in 2023 and 2024, Chinese model makers gained global mindshare and adoption by putting out models that were worse than America’s top labs but were varying degrees of open source. There are real idealists in the Chinese ecosystem, and Deepseek’s global impact had the whole Chinese ecosystem spending 2025 trying to match their impact with open models.

The Chinese government, who saw industry seemingly finding a way out from the pressure of trailing after the ChatGPT moment, started saying open source in policy documents. But money to support open source development hasn’t materialized. Chips aren’t free and you can’t train models and serve users off good vibes.

China’s funding environment for AI is orders of magnitude smaller than America’s. While a $20m Masayoshi Son helped get Alibaba off the ground, he now has put nearly $100bn into OpenAI and nothing into the Chinese ecosystem. Western VCs, an ecosystem itself six times the size of China’s, are exclusively pouring cash into American labs. Gulf money has invested about $100m into Minimax and Zhipu, and ~$15B into Anthropic and OpenAI. In contrast to their hundreds of billions invested into chip hardware, the Chinese state has only started to dip its toe into lab support. Underwhelming valuations and capex projections by China’s biggest AI players are coupled with rushed IPOs paired with headlines like ‘Chinese AI Unicorns are Running out of Money.’

China’s leading AI players do not have the funds to burn tens of billions like America’s leading labs, and the sheen of open source vibes in 2025 has worn off. So they’re now scrambling to make money and hook investors with closed source models. Smaller open models help with overseas go-to-market and robotics, but won’t cover the costs of 1GW clusters.2

Will Open Source Stick as Party Rhetoric?

Open source as a concept has broken containment. Party outlets are increasingly using the concept as a political metaphor about China’s civilization model and its take on geopolitics.

In a People’s Daily commentary, Zheng Yongnian 郑永年, dean of the School of Public Policy at The Chinese University of Hong Kong, argued that “Chinese-style modernization is a kind of ‘open-source’ modernization. It fully embodies the responsibility and value choice of “self-reliance and helping others, benefiting the world.” (“中国式现代化是一种“开源式”现代化。充分体现了“立己达人、兼济天下”的责任担当与价值选择”)

The piece uses ‘open source’ not in the narrow AI sense but to describe a geopolitical development model built on openness, experimentation, adaptation, and sharing. The Fifteenth Five-Year Plan outline, approved in March, also includes a more technical line calling to “promote open-source system construction and improve open-source operating mechanisms” (“推进开源体系建设，完善开源运行机制”) without any connection to AI. And an accompanying Xinhua commentary describes Chinese modernization as “an open modernization, a win-win modernization, and an ‘open-source’ civilizational practice.”

In other words, senior policymakers have started latching onto “open source” as a useful way to describe China’s approach to global development, especially in regard to South-South cooperation, often with no direct connection to AI at all.

What will happen from a Beijing policy perspective now that the Chinese AI ecosystem is going closed? Probably not much. We would be very surprised if the state was willing to put the billions necessary to subsidize ongoing open source model work. Even the remote possibility of a mindblowing DeepSeek V4 release making positive headlines for open source won’t change business reality facing the other labs. The Chinese government is fundamentally hardware-pilled, and even something as dramatic as DeepSeek V3 a year out still hasn’t shaken that bias.

Assuming China’s AI lab ecosystem doesn’t consolidate even further, there will still be players looking to make a splash with models that aren’t quite at the Chinese frontier, especially if Chinese AI companies must avoid subverting the party’s open source mantra. Perhaps the most optimistic story to tell for open source in China today is that with distillation and data hacks, it may not require that much money to make an also-ran model. And it’s easier to create buzz with an open model than a closed one no-one wants to pay for.