Login
EMERGENCY POD: H20 Drama
We’re bringing back Lennart Heim of RAND and author of Chip Wars and newly on substack, to discuss the new H20 drama, when exports were banned in April, and now selling it with a 15% export fee.
Today our conversation covers….
What’s at stake and the strongest arguments in favor and against selling AI chips to China
Will cutting off chips really make China more likely to invade Taiwan?
Where Trump goes from here on Blackwell exports, HBM, semiconductor manufacturing equipment, and what could change the current conciliatory direction of travel for the broader US-China relationship.
Listen now on iTunes, Spotify, or your favorite podcast app.
Why Care About AI Chip Exports to China
Jordan Schneider: Lennart, what is the H20? Why should people care about it? What were the first few months of the Trump administration doing when it came to this chip?
Lennart Heim: The H20 is a chip that NVIDIA designed as a response to export controls in 2023. It’s the typical game: you draw some lines, and then new chips get created right below those lines. The H20 is exactly such an example, but it did a neat trick.
It maxed out the specifications that are not controlled — memory bandwidth. They put the best high-bandwidth memory the world currently has on this chip and created an export control-compliant chip that was introduced at the beginning of 2024, a couple of months after the updated controls. The chip was sold throughout 2024 with lots of interest.
When the Trump administration started in January, the Biden administration didn’t get around to addressing this problem. Many officials spoke out in favor of taking action, but they never got to banning it because of many stakeholders, different opinions, and running out of time.
Trump then banned this chip, as reported in April 2025. Not through the normal regulatory process, but by using a tool called “is-informed” letters, which are pretty fast. You can send a letter to the companies that produce these chips telling them they can’t sell these chips anymore because you suspect an export control violation is going on. The administration argued this chip was simply too good.
From my personal point of view, banning the chip was a big success. This chip should not be sold. We need to reduce our thresholds — this is simply too good of a chip. That was the latest status. Then over the last few weeks, we saw some flip-flopping back and forth, with more information revealed every day. While we talk, probably more things will come out.
Jordan Schneider: Here was President Trump on Monday:
[Trump Audio Clip]:
Let me ask you two questions — one about China, one about Russia, if I could. On China, your administration agreed to send the most advanced or advanced NVIDIA and AMD chips...
No, obsolete chips. The 20s? No, this is an old chip that China already has, and I deal with Jensen, who’s a great guy, and NVIDIA. The chip we’re talking about, the H20, is an old chip. China already has it in a different form, different name, but they have it, or they have a combination of two that will make up for it, and even then some.
Now Jensen — Jensen’s a very brilliant guy — also has a new chip: the Blackwell. Do you know what the Blackwell is? The Blackwell is super-duper advanced. I wouldn’t make a deal with that, although it’s possible I’d make a deal with a somewhat enhanced — in a negative way — Blackwell. In other words, take 30% to 50% off of it. But that’s the latest and greatest in the world. Nobody has it. They won’t have it for five years.
But the H20 is obsolete. It still has a market. I said, “Listen, I want 20%. If I’m going to approve this for you, for the country, for our country, for the US — I don’t want it myself. You know, every time I say something, it’s for the Air Force.” When I say I want 20%, I want it for the country. I only care about the country, not about myself.
He said, “Would you make it 15%?” We negotiated a little deal. He’s selling an essentially old chip that Huawei has a similar chip for — a chip that does the same thing. I said, “Good, if I’m going to give it to you” — because they have what we call a stopper, not allowed to do it, a restrictive covenant — “if I’m going to do that, I want you to pay us as a country something, because I’m giving you a release.”
I released him only from the H20. Now on the Blackwell, he’s coming to see me again about that. But that will be an unenhanced version of the big one. We will sometimes sell fighter jets to a country and give them 20% less than what we have. Do you know what I mean?
Jordan Schneider: This is a good moment to take a step back and look at the arguments for and against selling China AI chips.
There are arguments against selling AI chips because selling helps upgrade the Chinese AI ecosystem that’s going to compete with America’s. There are specific applications of the chips that we would be selling to China that we would be very uncomfortable with — military ones, intelligence ones, or broad human rights violations that you wouldn’t want American technology to be helping to further.
There are also arguments in favor of selling. These include the idea that selling NVIDIA chips would retard domestic chip development, making it harder for SMIC and Huawei, and whoever else wants to try to build domestic AI chips to find a marketplace. There’s also the idea that selling chips into China would maintain Chinese dependency on the US stack, keeping Chinese developers using CUDA, building infrastructure around US technology. There’s some broad soft power and agenda-setting advantage that China's use of NVIDIA hardware will give to the US going forward.
Maybe we should run through those systematically. Let’s start with the biggest one, which is that you shouldn’t sell these chips to China because upgrading the Chinese AI ecosystem is a strategic threat to the US. Chris, this is almost a grand strategic question of how much of China’s rise is okay and how much isn’t, because the military intelligence and human rights applications are almost secondary to how scary you see a richer, more flourishing, more powerful China to be.
Chris Miller: I would segment out the “richer and more flourishing” side and just talk about technological capabilities. They’re interlinked, but the US strategy hasn’t been to try to make China poorer or less flourishing. The question is just who’s going to lead in AI.
The trend over the last five years, and the last 50 years, has been that if you want advanced AI, you need lots of advanced computing, and there’s a small number of companies that produce the chips in question. If you think that advanced technology has mattered in the past in geostrategic competition — which is pretty hard to argue with — it’s probably going to matter in the future. Therefore, who wins in AI matters.
Just as we would be less happy if we were all using Huawei phones and relying on Alibaba Cloud 阿里云 because there would be pretty significant political ramifications downstream of that, if we find ourselves in a future where either the US or third countries are relying on Chinese AI providers — whether for models, applications, or AI cloud — that implies less political influence for the US, a weaker US, and a stronger China. Those are the stakes.
Both sides of this argument agree on that basic framing. The question is, how best do you get there? One argument is that you restrict compute access and thereby hobble the growth of Chinese AI firms. A second argument is that you try to, as Secretary Lutnick has said, get China addicted to the AI stack. The question to ask is: how addicted are they willing to become? How addicted could you make them? Can you leverage that addiction in the future, or not? These are where the empirical questions are focused.
Jordan Schneider: One more argument in favor of selling: the idea that keeping China dependent on TSMC-fab chips lowers the risk of a Taiwan war, which I have some questions about. This is something that Ben Thompson has been pushing, which has percolated into the administration and Congress.
Will selling Beijing TSMC chips make them less likely to invade?
Lennart Heim: What do you think? What do you make of it, Jordan? I’m curious. For me, it doesn’t seem to be the main calculus behind it. I buy it on the margin.
Jordan Schneider: Maybe on the margin a little bit. There are two levels to the question. First, the political calculus to go to war or not to go to war — this would be an extremely weighty decision where the fates of nations would be at stake to do a serious blockade, strike, or actual D-Day style invasion.
Whether or not the chips are there, whether China is gaining or losing relatively in AI hardware, strikes me as about the 12th thing you would be thinking about if you were a Chinese premier. Domestic political developments in Beijing, how much you trust the PLA to not be corrupt and actually work as intended, political developments in Taipei, how willing the US seems to fight for Taiwan, how excited Japan is to let the US fight from its territory — all of those strike me as much more germane decisions.
There’s some real technological myopia among tech analysts thinking that the chips are the one thing — the Silicon Shield stopping war. As cool and important and potentially world-shaking as advanced semiconductors and artificial intelligence may be over 50 years, if you are a head of state making the biggest decision of your life, it’s not going to come down to “Well, Huawei tells me they can only make 750,000 chips in 2028, so it’s not going to work out, might as well bomb Taiwan because if we can’t have toys no-one can.”
Ben concludes his latest piece arguing that selling chips reduces invasion risk by saying that “Far too many people in this debate seem to operate as if the U.S. is the only actor in the world, with every other country, including China, operating as mere props. That’s simply not true, and accepting that is the first step to a cogent policy that preserves what leverage we still have, while minimizing the risks that too many are too unwilling to contemplate.” Ben Thompson more than anyone should know that technological progress does not reach a static end point and China has lots of case studies to point to of making it up value curves under adverse conditions. Thinking that their only out is to invade does what Ben says he wants to avoid, painting China as a prop. A more likely future where you price in agency for the government and their firms will see attempts to strive commercially under a set of geopolitical constraints, just like engineers at Deepseek did. The idea that a Chinese leader would think that “we’re missing out on AI so I guess we’ll have to start WWIII” strikes me as a bizarre conclusion.
One level down from that, there is this very open question, which we debated on Sunday’s edition of our new defense tech podcast Second Breakfast, about to what extent the chips and technology are going to be enabling ends up reshaping the military balance of power. That is still very much an open question that smart people can disagree on — whether what you can do with putting chips in your autonomous drones so they can target without interference, or whatever. You can imagine a lot of different crazy futures where AI matters.
By the way, it could work in the other direction, lowering the risk of a Taiwan war if America has a big lead when it comes to semiconductors. Then a leader in Beijing would look at the military balance of power and the advantage that US and Taiwanese forces get from being more AI-applied, and think, “There’s no chance of us winning. Why even try to play this game in the first place?”
Did you make it this far into the this newsletter? Well, maybe that means you want to work on this stuff full time!
The Horizon Fellowship provides a direct path into the world of emerging tech policy: a fully-funded placement at a federal agency, congressional office, or think tank in DC for up to two years. A deep understanding of China’s technology ecosystem and governance approaches is essential for effective US policy on these issues.
For its 2026 cohort, Horizon is actively seeking candidates with expertise on China’s technology ecosystem, policy landscape, and strategic goals. Prior technical or policy experience is not required. Join a community of 80+ fellows working on some of today’s most pressing issues.
The application deadline is August 28. Learn more and apply here.
Chris Miller: The other key facet here is that if you look at sales of advanced chips from Taiwan and its ecosystem to China, most of them are not AI chips. It’s mostly smartphone chips and PC processors. AI chips are a portion, but a small portion. This gets back to the question you’ve raised on a lot of shows, Jordan: how AI-pilled is Xi Jinping?
The answer doesn’t seem very AI-pilled. The best evidence for this is that SMIC and its seven-nanometer production are still producing a whole lot of smartphone chips, which you would not do if you thought we were in a race for AGI that will define the future. Both of those facets again point against the Silicon Shield as it relates to AI chips being central here.
Lennart Heim: Just to clarify, they’re not allowed to produce AI chips at TSMC. They can produce everything else there. Why not? Because Ante — they did some bad stuff — but almost every other Chinese company can just go to TSMC and produce chips there. There’s a significant flow of chips from Taiwan to China as we’re speaking right now, just ideally not AI chips. We had some hiccups in the past where there were also AI chips.
Chris Miller: A key question, Lennart, is how obsolete is the H20 relative to what Blackwell can do, but probably more importantly, what Huawei can do? Want to walk us through the numbers?
Lennart Heim: I don’t think it’s fair to describe the H20 as an obsolete chip.
Chips have many specifications. Let me break it down to two simple ones. We should care about computational power — how many FLOPS it has, how many operations per second it can crunch. But then also memory bandwidth, which means you need to read and write memory. The memory capacity and bandwidth — how fast you can read and write this memory — is key.
One of the key inventions we’ve seen over the last few years, which AMD did first, is so-called high bandwidth memory, which is a complex technology. We’ve got three companies in the world doing it right now: SK Hynix, Samsung, and Micron, building this HBM.
The H20 is bad on the FLOPS — seven times worse than the H100, even worse, 14 times worse than upcoming chips like B100s or more. It’s not a competitive chip there. But on the memory bandwidth side, which is again key for deploying chips, it’s pretty good. It’s even better than H100 because the H100 uses five units of HBM, whereas this one has six units of HBM. It gets a mind-boggling four terabytes per second of high-bandwidth memory.
No Chinese chip has such good high-bandwidth memory. More importantly, even if they have the right now, the Ascend 910C, which has some HBM at 3.2 terabytes per second, they’re not allowed to buy it anymore. It’s been banned since December 2024. Right now, China is struggling to get its hands on this HBM. They’re trying to produce it domestically, but this will take time, and even if they produce it domestically, it would initially be worse.
I don’t think the H20 is an obsolete chip. It’s a pretty competitive chip. It’s fair to say it’s a worse chip than many others, but if you look at this other dimension — the dimension of deployment — it’s pretty good. It’s really good.
Chris Miller: That is one of the key axes of debate. Some people say the goal is to stop China from training high-end models, and therefore, you focus on the FLOPS. If your goal is to constrain inference, you focus more on memory bandwidth. Walk us through the way these different chips are used.
Lennart Heim: That’s a fair debate we should be having here. We should think about export controls: what do we want them to achieve? Right now, it’s fair to say that the H20 is not an amazing chip for training AI systems. There are some things that numbers don’t always capture. You still build on top of the NVIDIA software stack. If your company used NVIDIA before, there’s a pain in switching. There are a bunch of problems with Huawei chips that you don’t see in the specifications — they overheat, you need more of them, the software stack isn’t great yet, and you can’t even get enough.
All of these things just mean that the H20 is not a great training chip, but beyond the numbers, you’re still stuck on the software ecosystem.
On the training goal, that’s still being achieved here. Where the debate begins is what we think about deployment. What I’ve learned over time is that if you want to be precise — if your goal is to only stop them from training, but everything else is below, or you only stop them from training big systems — it’s really hard to be precise on all of this. AI is ever-changing.
The biggest thing we’ve seen over the last few months is this rise of test-time compute of AI models thinking — how they think, how they produce tokens. That’s what the H20 is amazing at. One could say the usability and importance of the H20 only went up since we got models that do more thinking, generating more tokens, and also generating tokens to then train the next generation of AI systems. These are the arguments that Paul would say: “Well, actually, this is a pretty good chip for producing these new things that are more important in the AI development lifecycle.”
Chris Miller: The other argument that the president made is that Huawei already makes these chips, which is true to an extent, but walk us through the numbers there as you see them. There are questions both about the quality of Huawei’s chips as well as the numbers that can be produced. Secretary Lutnick said they can produce 200,000 a year, and I suppose that’s right. How does that compare with what we’re going to see with H20s?
Quality or Quantity?
Lennart Heim: The key dimensions here are quality and quantity. Many always talk about the quality argument here. I think the quantity argument is way more important. You already mentioned a number — Lutnick said Tesla also testified that 200,000 Ascend chips are being produced in 2025. How does this compare to the US? We’re churning out around 10 million chips this year — significantly more. This means if we’re selling — and there have been projections about NVIDIA selling a million H20s — we sell them five times more than what they can produce.
This is where the debate starts. The quantity argument is key here. If you would only sell them a couple thousand or 200,000 something, that’s a vastly different debate than selling a million or potentially even more. Just the sign that China wants to buy them speaks to their problems producing domestic chips.
On the quantity side, China’s simply not there yet. They’re getting better and producing more chips as we speak, but they have many difficulties along the way to produce more chips. Do they have enough high-bandwidth memory? How good is the smuggling operation to get this memory? How good is the packaging yield? All of these things just add up so that you eventually really can’t produce competitive chips.
The chips they get out of it — if you compare the Ascend 910C to NVIDIA’s best chip right now, which is being sold, the B200 — it’s way worse. It’s way worse on the high-bandwidth memory part, and it’s also way worse on the computational performance. It’s also worse than the H20, which you’re selling, at least on the memory part.
The point is, if you’re selling the H20 — and what many missed, there’s a chip, at least there were rumors around it, and pretty good rumors — there’s a chip called the H20E. What does it do? It doesn’t use HBM3. It uses HBM3E. I previously said it has four terabytes per second. If you use HBM3E, you can probably go up to five terabytes per second or even more.
What indications do we have that this chip is not getting sold? The FLOPS are still being kept, but the memory just continues going higher and higher and higher. That’s another thing to be tracking here. As long as we don’t have updated regulations for it, we just don’t know where the line is going to be drawn here in terms of quality of memory bandwidth, but also most importantly in terms of quantity.
If I could ask for one thing, please reduce the quantity. That’s the key thing we should pay attention to here.
Jordan Schneider: One of NVIDIA’s lines that Jensen has been saying they used to have 95% market share before the restrictions, and now it’s down to 50%. First off, they’ve never actually given numbers for that. But second, I’d guess that they were the only people making accelerators that people wanted. Even if it did go down to 50%, it’s not like it was the same pie — the pie went down such that the 5% it used to be now turns into 50% of the whole pie. The idea that Huawei — that number does not tell you that Huawei necessarily can fill it up.
As Lennart said, Jensen cares about this because lots of Chinese companies are willing to spend — his projection is what, $15 billion a year in sales? To think that Huawei and Baidu, and Tencent — they are not dumb. They are going to spend billions and billions of dollars in CapEx. By the way, this CapEx number seems small if you’re talking about Google and Meta, but is pretty large relative to the total CapEx that you’re seeing from the Chinese hyperscalers. They’re doing this because they think it is useful and important and relevant to their AI ambitions going forward, not to do Jensen a favor or anything.
Existential Priorities, Moral Values and AI Chips
Chris Miller: Could we talk about what we know in terms of who in China will be the large-scale buyers of these chips? Jordan, you mentioned Tencent, Alibaba. There’s AI firms like DeepSeek. There’s ByteDance, a huge player in China’s ecosystem. Lennart, if you have a sense of numbers, if any of those are public, or at least talk about who are the buyers of these chips inside of China?
Lennart Heim: I don’t think we have public reporting of it exactly. There’s definitely been some reporting that big hyperscalers, the big cloud companies — Tencent, ByteDance and others — are definitely interested in this. I’m not sure how interested ByteDance is because they’re building tons of clusters in Malaysia, which by the way, can buy whatever chips they want there and just continue building.
The normal hyperscalers will continue buying these kinds of chips, but they’re all hedging. They all also get Ascend chips. They’re not stupid. We just see with the policy flip-flopping, they don’t know when they’re going to get cut off. They’re all just hedging with Huawei Ascend chips while they’re getting better, because something we would just subsidize the transition while we do this.
That’s the thing I’m worried about here. It’s just a case that Huawei will get better, they will produce better chips. The chips will be significantly worse and significantly less quality than the US, but they will get better. That’s the thing we all need to acknowledge. There was a policy at some point which was made, which was telling Huawei they will need to produce their own chips. That’s just the path we’re going down here. There’s no going back here. The question is: what do we do in the meanwhile? How big will the gap potentially be? I’m a firm believer that this will be quite a massive gap, which will have big impact on the AI competition.
Chris Miller: That is one of the key lines of debate, but also empirical questions that’s hard to research or get hard data on, which are the decisions of the private tech firms in China, the Alibabas, the Tencents, and others. Because to the extent that you’re right, that there’s a meaningful quality difference between NVIDIA and Huawei GPUs, for example, they got a strong incentive to build as much as possible on NVIDIA.
You can see an argument that says, well, they’re going to buy Ascends, but put them in the closet or not really take them seriously because they want to build their products. But you’re saying no, that’s probably not the case because even those firms that don’t have a strong incentive on their own to help out Huawei, do in the context of potential future export controls and loss of access to NVIDIA chips. The argument that controls align the incentives of Tencent and Alibaba with Huawei and the Chinese state — you think those incentives are already fully aligned.
Lennart Heim: More importantly, we should always work through the arguments for it. There are arguments in favor of selling H20s, and that’s the same debate to be had here. On the other side, it sometimes lacks some technical details here.
The market share argument is a fair argument — you want to maintain NVIDIA’s bigger market share, and reduce demand for Huawei. I just don’t think that’s the case.
It’s an existential priority for China to develop the semiconductor industry.
Importantly, it’s not like the semiconductor industry only gets better because of AI chips. The majority of chips the world produces are not AI chips. Who’s producing at the most advanced node at SMIC, but also TSMC? It’s Apple. Usually we produce mobile phones first there, so they’re pushing it forward anyway for the newest Huawei smartphone that will probably soon produce something like a six-nanometer node, which will then be leveraged to produce better AI chips.
Even if you reduce the market demand right now, semiconductors will get better and these will lead to better AI chips eventually. If they then just transition to this, then also what is the tech stack argument here? Sure, we keep them hooked on CUDA, and it’s a pain to go from CUDA to PyTorch to MindSpore to the Huawei ecosystem. We can model this as a one-time transition cost. Many American companies have done this. Google switched to TPUs at some point. OpenAI right now is using Trainium chips on AWS. They pay a significant amount of cost to switch and run these different hardware stacks. But eventually they’re doing it, and they’ll also eventually do it with Huawei.
It’s not like if you use CUDA, your systems are more aligned. If you sell China AI systems that don’t spit out CCP propaganda, I’m in favor of that. That’s spreading American values, liberal values. That seems fine. But if you were just selling them chips, there are no values, no constraints that come with selling chips. You can just do whatever you want on it.
That’s again where we’re just missing this tech component. We kind of got it right in the UAE: sell them the cloud, let Microsoft build here, versus here we just sell the underlying component. They can build whatever they want on top of it. That’s just missing in the debate.
Chris Miller: This is a key aspect of the export control debate that’s fascinating. A lot of people don’t get this: if you restrict sales of tools, then you hurt the tool makers, but you help the users of those tools. In the chip industry, if you sell fewer lithography tools, it’s bad news for ASML, but it’s probably good in the long run for TSMC and other companies that face less Chinese competition. Similarly, if you sell GPUs to China, it’s bad news for GPU sellers. Or sorry, it’s good news for GPU sellers, but bad news for US AI firms who face stronger competition.
One of the strategic questions is at which level do you try to cut off? The US has, until recently, cut off at multiple levels and is now shifting. Well, we’ll see where we are next week, but this week it seems like it’s shifted towards a policy of sell the GPUs but keep the controls on the chip-making tools.
Lennart Heim: Which makes sense. If we would reverse, selling them extreme ultraviolet lithography machines from ASML, I would be way more on a rampage than selling them AI chips. I also complain more if we start selling Blackwells over H20s. That’s a fair debate we should be having here. People can fall into different types of positions here. We can disagree on some arguments here. You have these different types of controls, which stack with each other, and the AI chips are the first ones to fall. That makes sense.
Chris Miller: One of the arguments is that if you make China addicted to AI chips, you gain long-term leverage. The mental model that people think of here is: if you get them using EUV lithography tools, they don’t have their ecosystem, and it takes a decade to try to replicate your tools. So maybe this is a good one for Lennart. Does the same dynamic hold here, or if not, why? What are the differences?
Lennart Heim: There are many different facets of being addicted to something. In the ideal case, it just means all Chinese firms are really reluctant to adopt Chinese chips, and therefore, they have less revenue. SMIC is wondering, nobody wants to buy their chips, and instead, all the Chinese just buy US chips.
I already talked about how SMIC and semiconductors get better anyway, independent of AI. But it’s a fair thing to say: the less people use Huawei’s AI software ecosystem, the worse it is. That’s a fair argument to be made. I just think they know they want to produce it anyway. They just know we need our AI chips at some point. They’re not full steam on this. Maybe they could go stronger if they wanted to. Maybe they’re full steam on it, but they just don’t do better for many reasons.
China is using the US tech right now, maybe delays it to some degree, and even subsidizes it. Let’s just think about Volkswagen — you know my German heritage — and its love affair with China. How’s this going right now? Did this stop BYD? Not really. I expect the share of Volkswagen being sold to China in the future will be low. The argument to be made here: they made a ton of money in the meantime. That’s a fair argument to be made.
The reason I feel nervous about AI chips is that they increase the total compute deployment training capacity in the interim. If AGI is a singular point, AI’s just not going to materialize in five years, then all we discuss here doesn’t matter that much, because the good thing about AI chips is they get exponentially better. We’re not going to talk about H20s in five, six years from now because we have exponentially better chips already here.
That’s an argument. We can just say: don’t worry, we just sell them, we make some money, they get a little bit better AI, but AI’s not going to be decisive in the next four to five years. But then later, ideally, we stop it. We don’t sell them. We have better chips that are exponentially better. Again, it goes back to where we draw the threshold, and when and how AI matters. Which is a diffuse question.
I have a pretty uncertain view here. I’m just like, man, AI could be a really big deal in the next three to four years. It seems likely it’s going to be a big deal — bigger or less big, depending on how it goes from just transformative economic growth being determined, to the future of the military, up to just going to fizzle out. We should address this uncertainty here. I just work on national security ris,k and I’m trying to minimize downside risks. I don’t see the benefits here in the long run, that why we should sell them. Fair argument. There are some good arguments here, but overall, it doesn’t cut it, at least for me.
National Security and Politics
Jordan Schneider: When you look at some companies, it’s a really big deal having Chinese market access. Intel — 35% of their revenue is from selling CPUs into China. This was a big deal for the tool manufacturers. In some years, it was 30, 40% over the past few years. NVIDIA’s a $4 trillion company — they will be just fine and still be able to deliver you that exponential curve of rapidly improving AI chips even without the extra $10 billion of sales.
There’s the maximalist version of this question: if you are 100% sure that AI does not matter and is not a strategic technology, then yeah, sell it. Go crazy. Do whatever you want with it. But it’s a tricky line of thought where we’re writing an AI action plan where we want to make AI dominance, we think this is going to usher in a new golden age, but we’re willing to take some of this downside risk that we’re making it easier on China, which we’ve identified as a major strategic threat.
There is a broader context of the relationship that you can try to trade things in. Say, we wanted them to scuttle some submarines or stop messing with the Philippines — there are lots of other asks you can make from a balance of power regional dynamics perspective that you could have put on this. It’s wild that it didn’t even seem to be in the context of the debate or discussion between the US-China trade deal, but was just a decision that Trump made independently because Jensen got to him, and he wanted to have good vibes in the relationship, and the 15% tax we’re putting on it.
What if it went to buy drones for Taiwan or to shore up funding for BIS so they could do a better job of tracking down all the chips that are getting leaked out into China? There are some lines where if you are going to follow the premise that China is a strategic threat and we’ve got to watch and hopefully shape how much they’re going to gain on the US from a relative technological competition perspective, there are other moves you can do to use this card more in your favor than letting the other side pocket it.
Lennart Heim: What we’ve seen so far is that the H20 got sold again. Then some said it was part of the trade talks, and others denied it. Then the Chinese came out denying it was part of the trade talks. Eventually, what I don’t like about it: export control was a national security consideration. When the October framework came out, and there were certain companies in countries like Poland, Switzerland, in this tier two, many were complaining, “you’re dividing a European trade union,” but it’s a national security thing. It’s not a trade deal we’re doing here. This is at least where export controls originally came from.
Now we are mixing them with trade things, and now we get 15% of the revenue share, and amazing, let’s pay off the debt, let’s do other great things. I don’t think national security is for sale here. If we could get other national security concessions here in return, that’d be amazing. It would be nice to hear more and communicate about this. There are people like me who are willing to walk back. Hell yeah, let’s sell the H20 because we got a beautiful deal out of it.
I just don’t think 15% of the sales cuts it here. It’s just money. Money doesn’t help you.
Jordan Schneider: It’s really interesting, the analogies that Trump used in his talk, where on the one hand, he talked about selling fighter jets to allies. This is something we do — we sell F-35s to Saudi Arabia, ostensibly an ally, and we cut off a few miles per hour off its top speed or what have you. And then the other word he said: restrictive covenant. This is a real estate word. That’s the only time I’ve ever heard it used before. It’s like, okay, I am a landlord, and we’re cutting you a deal, doing some sort of deal, which is a straightforward commercial transaction, not having anything to do with national security.
I remember on Logan Paul’s podcast, he was like, “This is the most important thing and it’s going to shape the future” — to go from that to “oh, this is just another real estate deal. Yeah, I started at 20, Jensen got me down to 15.” Not without any of the grand strategic import that this decision again may not, but also may end up having for the future of this technology in the world.
Lennart Heim: Can I make a point about real estate? What you do with real estate is often you don’t sell it, you rent it out. If you want to give the Chinese computing power, rent these chips — it’s the best of both worlds. They get the computing power, you make money, you might even make more money because there is NVIDIA making money, and maybe Microsoft or your favorite hyperscaler in between. You still have more control and more leverage.
You don’t need chips in your basement to run them, you can access them remotely.
They could literally dial in. They could dial into our beautiful new UAE five-gigawatt cluster or dial in to the US and existing cloud providers. Then, in the future, if they go rogue, or you want to make sure it doesn’t go to certain military-linked entities, you usually have more leverage.
If we do the concessions we talked about — the different things we want to walk back before you sell chips — just tell them you can use the cloud, which is by the way, perfectly legal as we’re speaking right now. If they want the computing power, use our cloud. It’s all legal, you can go for it. We still make money.
Jordan Schneider: FYI Trump White House, NVIDIA employees gave to Kamala over Trump in 2024, 10 to one…
Chris Miller: There’s an interesting political economy dynamic here, which Lennart, you’re referencing, which is getting back to: if you sell the tools, you enable the chip maker; if you enable chip makers, that type of competitive dynamic.
What we’ve seen is GPU sellers, NVIDIA most prominently, being very vocal on this issue. We haven’t seen hyperscalers be vocal at all, even though one should conclude this implies more competition for them. Then we’ve seen mixed responses from AI model companies. Anthropic has been pretty vocal in opposition. I haven’t seen OpenAI. It strikes me that companies that have a lot at stake have been taking very different strategies — some being vocal, some not. I don’t know what exactly explains that.
Lennart Heim: You know which GPU they’re using? NVIDIA, and if you speak out against them, Jensen’s going to get you. If you look at Anthropic, who is slowly migrating to more Google TPUs and Amazon Trainium, you can see the deals, they can speak out against it where everybody else is reliant on Jensen.
I can at least confirm from many conversations with many people in these companies, this is part of the calculus they do here — you would rather not come out against Jensen. It’s clearly in NVIDIA’s interest. That’s why they’ve been pushing sovereign AI, selling chips as their thing. That’s beautiful. This helps them. Nobody else is doing it. This is not where Google’s coming in. The only competitor here is AMD.
NVIDIA’s market share is only going to go downhill from here. The total market will go up — AI is a big deal but AMD is getting better. Google GPUs are getting better, Microsoft chips are getting better, and Amazon chips are getting better. We have more and more startups getting better. We just have more AI chip competition. NVIDIA also feels slightly nervous about all of these issues.
I would love to live in a world where NVIDIA had a smaller market share and see what the hyperscalers and AI companies would say. Many of them would come out. OpenAI at least came out in favor of export controls historically when they talk about energy dominance and more. Right now, they’re all quiet because somebody else might then knock on the door.
Chris Miller: I’ve gotten lots of questions about what does industry think? Of course, what you’re saying is, well, which part of industry are you looking at? Which segment, which specific companies?
Jordan Schneider: Why don’t you do the HBM political economy? This has been reported that the Chinese government is asking for high-bandwidth memory as part of concession number two. What does that tell you, Lennart, that ask?
Lennart Heim: If I were running China, I would ask for high bandwidth memory over asking for H20s personally, because I’ve got my sovereign drive anyway. I want to build better and better AI chips. If I look at my current AI chip industry, I would want EUV, but maybe this is too much to ask for because we did this early on. Trump did it back in the day. But what is the thing we’ve only recently done is banning high-bandwidth memory units.
We got our chip, and next to the chip, we put the memory, and these memory units are being produced by Samsung and SK Hynix, and Micron. They’re not allowed to go to China anymore. We’ve seen reporting that at least the Chinese, again, the Chinese put forward: could HBM be traded? Is there something we can do here? I hope the US government will draw a clear red line here.
We talked about how you would walk back things. There are arguments in favor of selling chips. We talked about them. What we do here is not sell them our chips. What we do here is enable them to build better chips. The best way how the 910C or the 910D, whatever the next best chip they produce, will get better is by having higher bandwidth memory. Right now, China does not have the capacity to produce even HBM3.
There’s reporting about the first trial production of HBM3. In contrast, NVIDIA is starting to equip HBM4 and using HBM3E right now. Again, don’t get me wrong, China will get better. They will eventually produce high-bandwidth memory. There’s a lot more to be done, which could stop them from producing better memory. But in the meantime, while they’re scaling up this production and trying to get better, at least we should probably not sell them our high-bandwidth memory to make their AI chips more competitive. Because we might regret this in many years when we’re then competing in emerging markets and Huawei has a better chip, which can better compete with ours.
Chris Miller: The interesting dynamic in the memory space is that two of the three producers are not US, but Korean.
Lennart Heim: That’s also why we see probably tons of smuggling here, because it’s pretty close to China, and there are certain tricks to get more HBM. Don’t get me wrong, China is smuggling HBM right now, which is sand in the gears, but again, I’m in favor of throwing sand in the gears, and ideally, we get better enforcement, and they will get less HBM eventually.
[h2] AI Chips and Chinese Political Economy
Jordan Schneider: On July 15th, we got the news that the Trump administration is letting Nvidia start to sell H20 chips. A week later, the MSS published a notice to the public, saying to beware of digital spying via foreign-produced chips. Ten days later, the CAC — the Cyberspace Administration of China 国家互联网信息办公室— summoned Nvidia representatives over risks of being able to control AI systems in China remotely and accused them of having planted a kill switch in them.
Then we have a private leading cybersecurity research firm in China hat published a report which went viral, talking about all the ways that there could be backdoors. Ten days after that, on August 9th, state television did a whole report about how there might already be backdoors in these H20s, and they cite former ChinaTalk guest Tim Fist from CNAS and his report on this topic.
Why Beijing is pretending they hate the H20
Chris, what’s your read on this interesting brushback pitch we’ve gotten from the central organs about H20s in China?
Chris Miller: There are three potential explanations, not mutually exclusive. One is that the Chinese security services are paranoid. The discussion in Washington of the Chip Security Act, which would mandate geolocation verification, has been happening simultaneously with the H20 debate and has intensified those concerns. That’s explanation one.
Explanation two is that it’s part of an effort to discourage private Chinese tech firms from using H20s. There are people around Huawei or in the government who are afraid that H20s will take market share, and this is a way to say “buy more Huawei chips” as well.
The third explanation is that this is pressure on US firms like Nvidia to say, “We need you to do more, or else we’re not going to let you back in the market.” We’ve seen this in other segments of the tech sector, where China will ramp up pressure on a private US firm to have that firm then try to use its resources to shift the debate in Washington. You could maybe envision the HBM debate being part of what China’s looking for in the broader trade negotiations that are underway.
But it certainly wouldn’t be a very attractive endpoint for Nvidia if they got approval from the US and then didn’t get approval from the Chinese side to sell. Perhaps China thinks it has some leverage there. How exactly to attribute these three causes? I’m not exactly sure what shares I would put on each of them, but all three seem potentially relevant.
Lennart Heim: China also put out guidance a while ago on energy efficiency. This was actually in April or May when the H20 was sold before it got banned initially. They put out guidance that the H20 is famously energy inefficient if you look at FLOPS because of the export control bandwidth limitations. I don’t know exactly what this guidance means, but it discourages companies from using it.
Nobody’s been following it because now they’re buying it up in the single millions of chips. But it feeds into the same narrative here. You try to push certain companies or create artificial demand for Huawei chips and slowly tell them, “Hey guys, at some point we want to do our own AI chips.” As Chris was saying, I think all of these stories are simultaneously true. It all just makes sense, and there’s no big downside for them to do these kinds of things.
Chris Miller: Actually, there was a state media source — I don’t know if this is the one you’re referencing, Jordan — but one of its criticisms of the H20 was that it wasn’t environmentally friendly.
Jordan Schneider: They cite this exact NDRC line that Lennart talked about, where the goal is 5 teraflops or half a teraflop per watt, and the H20 can only give you 0.37.
Lennart Heim: It’s pretty bad — pretty environmentally unfriendly for training, but pretty damn environmentally friendly for deployment of AI chips. Way better than any Huawei chip, I can tell you that.
Jordan Schneider: Here’s a moment where some mirroring might be in order. We’ve just had an hour-long conversation about how messy and convoluted American policy towards artificial intelligence is, with many conflicting priorities. The same thing is happening in all these different ministries in China.
This is big news — a change in the landscape where people want to have their say and make their stamp on it. You don’t necessarily need to attribute some four-dimensional chess move. I’m sure the people in the MSS read Tim’s report and thought, “It would be stupid if we bought all these chips only for them to turn into bricks or spy on us or have bombs in them that are going to blow up like beepers in Lebanon.”
I’m sure folks in CAC feel the same way. Then there’s the same debate that we’ve been having for the past hour: is it net positive or net negative for domestic self-sufficiency to have a competitor to Huawei potentially take a big chunk of the market domestically? This is being played out in China.
At a broad level, now is the right time to ask for more from Nvidia. Now that they’ve gotten the green light and there’s $10-15 billion of demand for these chips sitting somewhere in Taiwan that they’re excited to ship out, they can say, “You better step it up or cut the price or do an extra screen to make sure there aren’t any kill switches.”
The way this is playing out on Twitter is, “Oh, China’s saying they don’t want them. That means we should sell them.” Reading that Chinese state media or state organs are saying something doesn’t necessarily mean it’s true. It’s not that hard to play — let’s not even give this credit for four-dimensional chess. This is just two-dimensional chess of saying, “Oh no, we’re worried about the chips. We don’t even want these chips.” That changes the political economy of the debate in Washington, where it makes selling these chips potentially easier.
That’s something to watch out for as we see the Chinese government saying, “Ah, no, we didn’t want these all that much. This isn’t a big concession. We’re worried about the second-order effects of this.”
But the fact is, the demand is not going anywhere. It’s not as if Alibaba’s not going to buy these chips because of these warnings.
Lennart Heim: Alibaba would be pretty sad if they suddenly only needed to rely on other inferior chips, where they can’t produce enough of them. Ideally, if I were running the Chinese government, I would put out regulations that I can sell all of the Huawei chips I can produce, and then fill the rest with some nice Nvidia chips.
But what’s interesting is that there’s some misunderstanding of what the Chip Security Act is supposed to do, and location verification. The idea is not to check if a chip is in China and then have a problem. The idea is to check if a chip is in Malaysia, Singapore, wherever you think chips are being smuggled, and then verify they don’t end up in China. This was never supposed to go on chips that go to China, because ideally, we don’t have any chips going to China, at least not the advanced ones.
This is an interesting confusion. This whole debate of hardware-enabled mechanisms and location verification was big in the UAE, Saudi Arabia, Singapore, and Malaysia, all of these smuggling hotspots, that people were worried about. Some people have been pushing — if we now stop selling chips, I’m arguing we should sell them cloud — but people could say, “We can sell them chips, but put something on the chip.”
But just knowing a chip is now in China and we know its location and city — how does this help us? Everyone can dial in remotely. Even if it sits at Tencent, who says that the PLA isn’t using it? You can dial in remotely. I don’t know what’s going on there. If there’s some misinterpretation of documents, it’s a confusing situation.
Jordan, you previously made a point about Intel, which is interesting. Intel made a lot of money in China, and Intel is still allowed to sell its CPUs, but Intel’s CPU share in China is going down. We will see the same with Nvidia and AI chips. Even if you’re allowed to sell in China, your share will potentially go down. Why is this the case? There is similar guidance, for example, for all government computers to go to homegrown, domestically produced chips. “We can’t trust Intel anymore on this.” We will see the same on AI chips.
China is pushing self-reliance to produce its own AI chips. They also named security concerns here — that’s why the government is coming first. I don’t know the exact numbers of Intel sales right now in China and how much money they’re making there, but I’m pretty confident it’s been going down, and the government is not buying any more Intel chips because they just put out this guidance.
We’ve seen this playbook before. The only difference is now we have this confusion about which chips are allowed to be sold, which ones are not to be sold, and how good they are. But the story’s nothing new.
Chris Miller: Could we talk about what we know in terms of the big buyers of AI chips in China and their relationship with the state? You’ve got the private tech firms — Alibaba and Tencent. You’ve got the AI labs, DeepSeek most prominently. One of the key questions seems to be: what is the relationship with the state today, and how is it changing?
To what extent should we see them as arms of the state? That’s certainly not accurate. Totally independent is certainly not accurate either. There’s a spectrum. To what extent are these political priorities shaping their procurement decisions?
Jordan Schneider: There was some reporting which was clearly sourced by the intelligence community over the past few years that after the Chinese Ministry of State Security 国家安全部 hack of the SF-86 — that’s the form you submit to the US government when you want a security clearance, which basically is your confession of sorts to the Catholic Church where you talk about all your divorces and all your debt and everything that a foreign intelligence community might want to know about you — that data was tapped by the MSS through Alibaba and ByteDance engineers to put into a more useful format.
We’ve seen over the past few weeks reporting from Business Insider about a public tender from some corner of the PLA that wanted H20s to do whatever they wanted to do with it. I used to be more sanguine on this type of thing, but this is the most dual-use technology to beat out other dual-use technologies. It seems preposterous that, insofar as this is a strategic resource, the Chinese government would not be able to leverage data centers that are located in China — that the US does not have any kill switches or on-chip governance on — to do whatever they want with it, whether that’s building a surveillance system or helping with weapons manufacturing.
The Pentagon has now signed what I think is a $200 million contract with OpenAI, and this is just the beginning. This stuff is useful — we’re willing to pay a lot of money to get it into the Pentagon in one form or another. If selling a lot of H20s materially raises the amount of usable functional compute that can be put into anything in China — it would be really surprising if you didn’t have the Chinese government wanting to take these new tools out for a spin, if you didn’t have the Chinese military-police complex wanting to take these new tools out for a spin.
Chris Miller: There are two points you can analyze. One is: if AI tools exist, will the military use them? Obviously the answer is yes. But on the procurement side, if you’re a data center procurement official or executive at Alibaba Cloud, to what extent is your decision-making shaped by what you read in state media versus what your boss tells you to build an effective cloud, in which case maybe H20s are your best option versus a sense of — how do we think about this? Because those are the people who are going to decide how many Ascend chips to buy, unless they’re getting a dictate from the top, which maybe they are.
The counter-example I’m thinking of is there was a time when parts of the US military were using Chinese drones — not because there was a policy to use Chinese drones, but because they didn’t have any US drones. Is there a scenario in which your procurement executive at Alibaba is just going to try to ignore Ascend chips because they were told to build a good data center?
Jordan Schneider: At some level, yes. These are companies that report quarterly earnings and pay their employees based on how well the company performs. People get stock options. By and large, the incentives of the people who are buying these chips are to drive the most revenue for the money you’re spending on your CapEx.
But it only goes so far. There is this broader strategic realization, which you don’t even need Beijing to tell you — this door could be closed at any time.
Lennart Heim: We closed it.
What could change the current conciliatory White House dynamic towards China?
Jordan Schneider: Maybe now’s an interesting moment to talk about the sorts of things that could change the dynamic we’re on now on chips and the broader US-China relationship. We have Congress as a variable. There have been several senators and congresspeople who’ve been like, “Wait, what are we doing selling these chips to China? I thought we banned and said this was our golden ticket to the 21st century.”
Because Trump is doing this at such a personal level — we’ve seen him turn on Putin, right? We’ve seen him go from all-in on Putin to “we’re going to ask some questions about this guy.” We’ll see what happens in Alaska. But there is the possibility of Jensen saying the wrong thing, taking too much of a victory lap, or Xi Jinping doing something obnoxious. There are a lot of personal interpersonal dynamics that could change what the Trump administration ends up doing, which is probably the more relevant variable than whether or not Lennart can convince you that Huawei can only make X amount of chips.
Lennart Heim: It’s an interesting moment in time because we just have all of the trade negotiations, right? Everything is volatile, and certain things are just on the table, and they’d be willing to discuss them. We see the Chinese bringing forward, at least according to reporting, the idea of HBM.
It will be interesting to see what the government is going to say. It’s going to draw a red line. We had statements before the trade negotiations in London that H20 is above the red line — they wouldn’t negotiate it. We can all try to put together the story of what happened, but we won’t know for sure. But there will be more discussions about these kinds of things. The Chinese can bring it up.
But I’m also more interested in the semiconductor manufacturing equipment companies. If Nvidia got this beautiful deal, I know what we’re doing — they’re all trying to give the president a call. It seems like it’s a handful of people who are making these decisions, and I hope they’re well-informed about which things are more important. If I see any news about EUV machines being sold to China, I’m probably going to get a heart attack because I don’t want this to happen.
Jordan Schneider: From a personal transaction perspective, there isn’t someone in the semiconductor capital equipment ecosystem that Trump is going to give the time of day to. He felt like he had to deal with Jensen because this is America’s most important CEO. I don’t think any of those folks have the panache and skill to make it work.
Even Ben Thompson, who I gave a hard time earlier in this podcast, understands very clearly that there’s a lot of risk in selling more tools to China than we already have.
Lennart Heim: Going even further, it wouldn’t be good for Jensen if Huawei is not good at producing AI chips. It wouldn’t be in their interest to say, “Hey, yeah, let’s make sure we sell ASML chips. Let’s make sure to hit them on every single dimension we can to make sure Huawei is just less competitive.” I would love to see that this would be at least a good part of the story here.
Chris Miller: Congress will be interesting to watch on this issue. The trend in Congress has been vocally pushing for tougher controls, both in the first Trump administration and under Biden — not universally, but that’s been the predominant push. We need to watch Senator Cotton, for example, and what he does or does not say publicly on this issue.
Jordan Schneider: Chris, do you want to tease out the Russia comparison a little bit? Congress was really not happy. They ended up putting some sanctions on the table. What have the dynamics been there over the past six months?
Chris Miller: The last six months in Russia have seen Congress officially not play much role at all. They put sanctions legislation on the table and then pulled it back actually after Trump requested it. But there have been a number of Republican senators who have been influential in shaping Trump’s thinking. Lindsey Graham, for example, seems to have played a role in shaping Trump’s thinking on Putin over the last six months and the way that Putin is stringing along.
We’re going to Alaska later this week, and maybe all that will prove irrelevant if Trump changes his mind. But it does seem like you could argue that even though Congress has done nothing on Russia, in fact, it has helped change thinking in the White House. I wonder if this would be true here, but this seems like a place where Trump’s going to make more of his own decisions, especially insofar as it intersects with the China trade negotiations, which it seems like it may.
Jordan Schneider: It’s less salient than a land war.
Chris Miller: There’s no domestic constituency.
Jordan Schneider: Just weirdos with tech national security podcasts.
Nvidia Chips Past the H20
Chris Miller: Before this week, it was reported that Nvidia is coming out with a downgraded version of some new downgraded chip post-H20, the B40 or B30. That’s now irrelevant because of H20.
Lennart Heim: It’s unclear. We flip-flopped the decision on the H20, but notably there is still a license requirement. Nvidia had a license granted, so if they wanted to go all the way back, they could have removed the license requirement. From October 2023 to April 2025, there was no license requirement. Then they introduced the license requirement, which is still intact. The only thing which happened as of last Friday is they granted the licenses according to reporting.
If they still want to sell a chip which is not subject to export controls, they would produce a new chip called B30 or B40. It needs to be below the computational power threshold, so the same as H20, and also have lower memory bandwidth.
According to the reporting, I think the FT leaked what is in the formulation — it needs to be less than 1.4 terabyte per second memory bandwidth. The H20 is at four terabyte per second, so the B40 would probably not use HBM anymore. It would probably use an inferior memory technology, but significantly cheaper because why use HBM if you can’t have that much memory bandwidth anyway? It’s so-called GDDR technology, which you usually use for graphics GPUs.
If people talk about this being only the fourth best chip, I don’t think H20 is the fourth best chip. The B30, B40 — that’s a more fair description of a fourth best chip, and I would still not call it an obsolete chip, but it’s definitely a worse chip. It’s only a chip where the US government at least decided, “Here’s where we draw the new lines. This chip is fine to be exported without a license,” so it could still be coming. I have not heard they’re stopping production yet. I guess Nvidia’s making a calculus right now on how much demand there is, but it’s clearly the case that H20 is better. The question is, will all the licenses be granted going forward?
Chris Miller: Trump said at the press conference a couple days ago that he’ll consider a downgraded Blackwell. Are there ways we should think about what that might look like, if in fact it materializes? Of course, with huge questions over whether or not that’s actually real.
Lennart Heim: One thing which stood out — he said ~ 30% or 15% to 50% less performance. What many people are missing about AI chips and computing chips is they get exponentially better. If your chip is 15% less, that’s nothing. That’s still the same generation.
If you really want to sell worse chips, you need to go back a few generations and then the chip needs to be like seven times worse, not only 50% or 15%.
There’s an argument to be made that you want to sell worse chips, but it’s not a little bit of a downgrade. We really need to take the exponentials into account. If we trim down a Blackwell chip, for example, a B200 by 15% to 50%, it’s still roughly twice or three times as good as the Huawei chip. We can produce millions of them while Huawei struggles, according to reporting, to produce 200,000 this year.
That’s a key thing to get right here. People need to keep in mind the exponentials — chips get exponentially better. Fifteen to 50% trim is nothing in the grand scheme of things. I would make my voice heard to say this is probably not a good idea of what we should be doing here. The government drew lines before, and the lines are way lower, and that’s where they should be.
ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
Mood Music
