Media Briefing: The Future of U.S. Chip Controls in the Global Tech Race

CFR experts discuss President Donald Trump’s approval of advanced artificial intelligence (AI) chip sales to China and what that decision means for the future of AI, national security, and U.S.-China relations. 

CHANG: (In progress.) The contents of this discussion and Q&A will be on the record, and a recording of this will be posted online at CFR.org at the conclusion of the discussion. This conversation should go for about an hour, with about two-thirds to three-quarters discussion among the experts and then at the back end questions for those attending. 

This briefing is part of the Council’s ongoing mission to inform U.S. engagement with the world, work that also includes the analysis and resources posted across our channels including on ForeignAffairs.com and CFR.org, where you will find work on the broad scope of economic security and the tech competition with China. 

We have a robust lineup of Council exports—experts for you today—not Council exports; Council experts talking about exports today. We will start off with today’s moderator, Adam Segal, the Ira A. Lipman chair in emerging technologies and national security, and our director of the Digital and Cyberspace Policy Program at CFR. Over to you, Adam. 

SEGAL: Thanks, Ben. And thanks to everyone for joining us this afternoon. As Ben said, we’re going to have a conversation among ourselves and then we’ll open it up for questions, and hopefully we’ll have a robust discussion there. 

Let me just real quickly run through who we have on the—on the call. 

We have Rush Doshi, who’s the C.V. Starr for—and senior fellow for Asia Studies, and the director of the China Strategy Initiative here at the Council; Jessica Brandt, who’s a senior fellow for technology and national security at the—at the Council; Chris McGuire, who’s a senior fellow for China and emerging technologies; and Mike Horowitz, who is a senior fellow for technology and innovation as well as a professor at University of Pennsylvania and director of the Perry World House at University of Pennsylvania. 

So I’m going to ask Chris to start us off to kind of do a level-setting. Tell us what happened—what technologies are being affected and why it all matters. So, Chris, over to you. 

MCGUIRE: Thanks a lot, Adam. 

So yesterday the Trump administration announced that they were going to approve the export of NVIDIA H200 chips to China. I think my—hi, sorry. My apologies. Am I back? I had a technical issue. Back here? OK. Sorry about that. 

Yesterday the Trump administration announced that they were going to approve the export of NVIDIA H200 chips to China. So this is a sea change in U.S. technology policy, U.S. AI policy towards China, and in many ways the U.S.-China relationship writ large that others will talk through. The kind of core of U.S. technology protection policies that the—both administrations had been pursuing has been restrictions on advanced computing, whether it’s the kind of access to semiconductor tools, or access to TSMC, or access to the chips themselves. This has been kind of the number-one thing that the United States has used to kind of constrain China’s technical capabilities, and it’s also something the Chinese are very poor at doing themselves, that we have big advantages in. But the reversal of this is a pretty—is very significant. 

So the Biden administration’s policy was to maintain as large of a lead as possible over China in certain technologies, especially, you know, the most advanced chips and fundamentally in artificial intelligence. The reason that AI chips like the H200 are so good, are so important for AI is that fundamentally you need very large numbers of them to develop leading-edge AI capabilities. And the reason that the U.S. is focused on export controls is that the Chinese are not very good at making them themselves. Specifically, you know, there is the recognition in the previous couple administrations that these are dual-use items. They have significant military and civilian applications. And therefore they are not treated as commercial transactions. They’ve been subject to controls and evaluated—the transactions have been evaluated on a national security basis. 

The Trump administration announced that they are going to sell the more advanced chips. I’ll go through in a second and kind of contextualize that and how important they are, and then also announced that they would charge a 25 percent fee. There’s various ways that—there’s rumors on how they’re structuring this to do so, but the kind of principal here is that these are being much more viewed as commercial transactions, less as national security transactions, and kind of the overriding thesis is what’s good for American companies is inherently good for U.S. national security. Whereas I think that was generally not the view previously. 

So the H200, what is it? It’s important to kind of put it in context. It’s the second-best chip that NVIDIA makes, or it’s the best chip of the previous generation. It is not as good as the Blackwell chip that President Trump decided not to send to China, but it is also—it is still an extremely capable chip. It is six times better than the best chip available to China right now. It is nine times higher than export controls. China doesn’t plan to make a chip as good as the H200 for the next two years. And it’s not clear that China’s going to be able to make a chip that’s dramatically better than the H200 for the next decade. 

So this is important because the question is coming up a lot of are the Chinese going to be able to substitute for U.S. chips? If we don’t sell them to China, are the Chinese going to make them themselves? One thing I would note is actually just looking at the data, and just trying to actually quantify this, Huawei and NVIDIA both have public roadmaps where they both have said, here are the plans for our specifications of our chips out to 2028. Like I said, Huawei’s chip next year is actually going to get worse than their chip this year. Which is kind of unprecedented in the—you know, when it comes to making advanced semiconductors. And they’re not going to have a chip that’s as capable as the H200 until—that they’re really going to field until 2028. 

So this is giving China a capability that it cannot make itself, and that it probably will not be able to make itself, for several years. So then kind of thinking through, what are—what are the implications of this, specifically for the technology competition. And Mike, and Jessica, and Rush, and others will kind of walk through the other elements of it. But I wanted to talk through quickly what it means for AI companies, and hyperscalers, and really the U.S.-China relationship. I mean, fundamentally, the area where the United States has the biggest advantage over China in AI is computing power.  

If you think of the other elements of the AI stack you have, you know, data, talent, algorithms, electricity, applications. All of these things I think either they’re a wash or China’s actually better than the United States. But the U.S. advantage in hardware is really significant, and it’s very important. If this is now ceded, and now the U.S. and China have more or less equal access to advanced hardware, that is going to create a—that is going to be a sea-change in the AI competition. I think it’s going to result in Chinese companies being able to close the gap with U.S.—with U.S. models.  

Right now they’re kind of several months behind, but you could easily see DeepSeek and others go—get much closer to the frontier, if not potentially release frontier models with access to unlimited amounts of computing power. I think you also could start to see Chinese hyperscalers start to build data centers around the world. We’ve seen, you know, repeatedly the model of China having—you know, providing good-enough goods that are not as good as what the U.S. is offering, but competing on price rather than on quality, allowing China to kind of build global data centers with these, you know, second-best chips is exactly kind of what we’ll see. 

And then we’ll have to also monitor what it means for supply for U.S. companies, because this is a very supply-constrained environment. There are many inputs into the chips, whether it’s, you know, runtime at TSMC, or the high-bandwidth memory, or other things that are—for which there is much more demand than there is supply. So we’ll see if NVIDIA is selling, you know, millions of AI chips to China potentially, on top of helping the Chinese what effect does it have on U.S. and allied capacity? And could there be a decrease there? So very significant move. Big change in technology policy. Has big implications for the U.S.-China relationship and AI competition. But I’ll stop there and kick back to Adam. 

SEGAL: Thanks, Chris. 

So I’m going to go to Rush now. And maybe, Rush, you can pull back a little bit. Chris mentioned the kind of the larger relationship. And last week we had the issuance of the National Security Strategy, which seems to deemphasize great-power competition, really talks about the U.S.-China relationship in the economic context. So maybe you can help place this in the larger framework there. 

DOSHI: Well, thanks very much, Adam. And thanks to all of you for joining. Hope this will be useful to many of you.  

Let me just make maybe two broad points. The first is that this decision is not part of a quid pro quo. It wasn’t as if the Trump administration decided to release the H200 in exchange for something, which some people are suggesting. It was actually already a done deal that we were going to get the Chinese to buy our soybeans, and that maybe they’d sell us their critical minerals. And that was already taken care of. This decision was made weeks after President Trump and President Xi met, more or less unilaterally. Largely, I think, as a result of lobbying within and outside the administration, suggesting that this was something that’d be in the American interest. And I think perhaps there can be a case made for that, but as Chris has mentioned that case is not very strong—at least on the empirical side. 

So that’s kind of to set the scene. This is not something that is fundamentally related to a transaction with China. It is an independent U.S. action. And it’s the second time this has happened. Earlier in the administration, the Trump administration, quickly release export controls the H20, the chip that Chris mentioned is much weaker than the H200. And that decision was also made without any request by China, unilateral. In fact, the Chinese government thanked the U.S. for its unprompted, you know, good spirit and good faith release of the chip at one point, but very much indicated that they had not asked for it. So that, I think, sets the scene a little bit. 

But it also comes at a moment where the administration’s approach, Adam, as you’ve sort of indicated, has sort of—has changed on China. You know, early in the administration we saw tariffs go up to 140 percent. That pushed China into a corner. They decided to reach for their rare earths tool. They went after rare earths, and magnets, and minerals. And the administration, alarmed by the possibility of big industrial disruption, decided to deescalate. Now, after that first deescalation in April, I think there was a sense from China that maybe they’d figured out the Trump administration’s number. And they called the number again in the fall, right ahead of the Busan summit. 

And so what we’ve seen since then is a softening of the administration’s approach across the board, whether that’s on Taiwan or that’s on technology. But what’s interesting is the NSS seems to ratify—the National Security Strategy seems to ratify that softening. I mean, China does not appear, I think, until page nineteen of the document, which is highly irregular and is quite a sharp contrast to the first Trump term National Security Strategy. And there isn’t really a focus on many of the verticals of competition with China, like, for example, securing our critical infrastructure from cybertech, or even enough on what the U.S. interests might be in Taiwan. There’s some discussion of the geographic nature of those interests, but nothing bigger. 

And so the NSS does kind of ratify that softening. And then we see this major change on technology policy. And I think, as Chris mentioned, it is the most significant change on technology policy probably in about seven years. And that, I think, is worth really unpacking. Why is it that the Trump administration has made this move? What exactly convinced them that now is the time to export these chips? Especially when the gap between American compute and Chinese compute is growing? Why throw away that advantage? 

And I’m not sure the administration yet has much of an answer. We saw President Trump’s Truth Social post suggesting at least a partial explanation, that this would lead to jobs, or others suggested it would lead to greater innovation. But the honest truth is all these chips are going to be made in the same fabs in Taiwan, by TSMC, that the chips that are made for America are made in. And so every chip sold to China is one less chip that’s sold to the United States, potentially.  

And the revenue that could have come from China was already available to these companies when they were purchasing the chips—when they were selling the chips to American companies. They didn’t lose any additional revenue. And they’re not making the chips themselves. They’re just designing them. So the argument that this is somehow going to boost American innovation or employment, it doesn’t seem like it’s borne out by the facts. But what it will do is probably give China a chance to pull ahead in this sector, because they have a number of advantages in AI. One area of disadvantage was compute. Their disadvantage may now diminish. 

SEGAL: Thanks, Rush.  

So, Jessica, Chris laid out some of the challenges on the—for U.S. firms. And the president, in announcing the deal, clearly was sensitive to national security concerns, because he says on, you know, the Truth Social posting, “national security is going to be taken care of.” So maybe you can walk us through what some of those concerns are. And, you know, are they going to be taken care of? 

BRANDT: Yeah, sure. Thanks so much. And thanks for having me. Yeah. I mean, I think we’re sort of engaged in sort of three simultaneous competitions at the same time. One, you know, the race to the frontier. The other is the race to integrate into national security systems—military and intelligence systems.  

And the third is the race to integrate these technologies into the broader economy, because they’ll have enormous economic impacts that, you know, reflect on national power. And, you know, I think the concern here is that this will, you know, enable to China—enable China to close gaps, you know, on all three. You know, pushing forward, you know, at the frontier. You know, if they’re able to build high-performing systems—you know, as Chris alluded to—you know, it could provide an even, you know, sort of more appealing offering to countries around the world that are looking for these solutions. 

You know, I worry about, you know, in the media term, like, you know, division through phone companies or other mechanisms to—you know, immediate diversion to the PLA. And, you know, over the longer run—you know, and I think Mike can speak to this—but, you know, ability to, you know, kind of shorten the timelines for advancing on, you know, integrating AI into broader military concerns. You know, so across the board not a good picture.  

You know, I think, as others have said, like, the stated goal here was, you know, to keep China dependent on U.S. technology and to boost revenue. And I think there’s, you know, potentially a better way to do that, which is, you know, to rent chips to China. You know, to provide China the compute it seeks, for a fee. Which I think, you know, potentially offers lots of what, you know, a chip deal—and I think, as Rush said, this actually wasn’t a deal. This was sort of a unilateral concession.  

But, you know, I think would have other advantages because, you know, access can be cut off at any stage. It’s sort of harder for customers to redirect access undetected. You know, kind of provides continuous leverage, because every time you have a renewal it’s a new negotiation opportunity. And I would say, you know, the economics seem better, because you can create kind of two revenue streams, you know, instead of one. One for cloud compute—you know, U.S. cloud firms, and one, you know, for chips.  

So I agree, you know, with my colleagues here who have sort of said I don’t think either—you know, neither renting chips or selling chips is going to stop China’s indigenization effort. I think, you know, Beijing committed to that path long ago, and it’s for strategic and not economic reasons. So, you know, I think if the administration, though, is committed to this path, there’s sort of a better way to achieve its stated goals. 

SEGAL: Thanks, Jessica. 

So, Mike, I guess one, broadly, do you have the same response that Rush did in the kind of larger import in the NSS and competition with China? And then, to pick up on a point that Jessica made, does this affect the timeline for, you know, PLA usage or, you know, the PLA can use the cheaper chips already, and—I mean, how does that affect our thinking about applications on the military side and in China? 

HOROWITZ: Thanks, Adam. And thanks for having me. I mean, I agree with my colleagues completely here. I think that this is almost the definition of an own-goal when it comes to a policy change. You know, the Trump administration in some ways taught us the first time around that treating China as a geopolitical competitor had massive upside, and that, in fact, it did not offend China. They said, yeah, we are, in fact, competing with you. And to see the Trump administration back off of that, in some ways macro as they did in the National Security Strategy, is troubling and reflects a more dovish approach. And this—in some ways, this decision about H200s is consistent with that more dovish approach toward China, where China is being viewed less as a competitor.  

I mean, and at the same time—to get to what you asked—you know, we live in a world where, you know, China’s military is attempting to be prepared to invade Taiwan by 2027, and views the integration of artificial intelligence into military systems as an important element of how they will achieve those goals. And there’s research by Georgetown’s CSET and others suggesting that the PLA, in fact, is eager to have access to NVIDIA chips and other American chips to fuel their military enterprise.  

And so while certainly they will use the chips that they have, and, you know, certainly they’re not going to exclusively rely on American chips in a world where, say, the Trump administration could reverse course at some point in the future, this undoubtably in the sort term both enhances—or, what’s the right way to say this—the decision by the Trump administration in this context both enhances China’s commercial AI capabilities in the short term, and enhances their ability to integrate AI for military purposes. And that’s really risky if you think that there’s a possibility of conflict that—you know, rising over the next few years. 

SEGAL: Thanks, Mike. 

So you’ve all been pretty skeptical of at least two of the administration’s arguments about why they sold the chips. Rush talked about the revenues that could be reinvested in innovation and that somehow, you know, would drive the U.S. forward, and Jessica was skeptical that this creates dependence on the Chinese side. But I think there’s a third argument that especially you hear from David Sacks, which is, you know, just they want to flood the zone, right? We want U.S. technology as broadly adopted and adapted as possible. I wonder how you respond to that argument. I’ll go to Chris first. Anybody else who wants to pipe in, go— 

HOROWITZ: Yeah, I definitely have some thoughts. 

SEGAL: Yeah. 

MCGUIRE: Yeah. I actually think that that is completely the right approach, that we absolutely should flood the zone with U.S. technology, and it is completely the right goal to make sure that U.S. tech and the U.S. tech stack dominates globally. I think where the difference of opinion comes in is, like—is the definition of what the U.S. tech stack in. 

In my opinion, the U.S. tech stack does not just mean U.S. chips; it means U.S. chips on U.S. cloud running U.S. models. That is the full—that is the full AI stack. And actually, if the Chinese were to export—they’re not exporting right now, but if they were to export the full AI stack it would be that, right? It would be Huawei chips on Alibaba cloud running DeepSeek models. So for us to export U.S. chips in order to actually power Chinese cloud running Chinese models, I don’t actually think that fully achieves the admittedly good goal of ensuring the U.S. tech stack dominates. If anything, I think it could—it could kind of undercut that by enabling competitor stacks to dominate and set. 

SEGAL: Mike, you want to come in? 

HOROWITZ: Yeah. I think that the—I think Chris is right. And I would just add that this underlying theory that others have mentioned that somehow through that American—that we could flood the zone in China’s market and somehow drive out their domestic investment and their own—in this technology has failed in every other economic product category for the last twenty to thirty years. There’s no reason to assume that AI technology is so different that things will change. And so it’s not—while we do wish to dominate the global market, the—it’s not clear that exports to China in this context will enable that dominance in a world where China has proven extremely capable of using a combination of IP theft and domestic subsidies to ensure that their champions stay viable and then undercut us later on. 

SEGAL: Let’s go Rush and then Jessica. 

DOSHI: Sure. I think it’s a little bit like saying you want the nuclear revolution to run on the American uranium stack, right? We wouldn’t just start proliferating highly-enriched uranium around the world. Now, obviously, AI is not nuclear weapons, but the idea is that it matters which part of the stack you want people to be dependent on. And actually, this is the enabling element of the stack or the actual application layer, the actual use of artificial intelligence around the world. China is playing in that space, and I think it’s important for people to understand the following. 

China dominates in energy. They dominate—they have far more engineers than the United States does. And they already manufacture the entire edge—that is, all the devices on which AI will run, from everyone’s computer to everyone’s phone. And so they have enormous structural advantages that if they can just get the compute and the models right, and the inference right, then the world is going to have to run on the Chinese AI stack. China has every part of the stack. The one thing that they’re really missing as substance is the ability to train their models and run inference at scale at the very best level, and that’s what we would be able to provide them with these exports. 

And I want to just make a second point very quickly, which is I think a lot of people think that the export controls are just about national security. And that’s a big part of it, right? It is about making sure that these chips don’t go to the PLA or don’t get used to train, you know, Chinese missile systems or do Chinese command and control. But that is not the only question at stake here. AI, I think the Chinese believe and we believe, is a transformative technology, not just in the space of LLMs but also in the space of, let’s say, manufacturing or robotics. And if you decide that you’re going to lose the compute advantage there, that others are going to get there first, and they’re the ones who make the robots and they’re the ones who make the cellphones, then at the end of the day what you have is a position where your competitor might be in a position to leapfrog you. They may be in a position to seize the next industrial revolution before you can. 

And if that happens we’re not just talking about a few hits to American GDP growth. We’re not talking about the possibility that the AI moment or the AI bubble, whatever you want to call it, could get damaged. You’re talking about the possibility that the United States simply will not be able to keep up in the next round of technological innovation. And China perceives there as having been four industrial revolutions, the first one being steam engines which Great Britain won, the second being electrification, the third being mass manufacturing which the U.S. won, the fourth one is upon us now. And they think if they win it they’re able to dominate and leapfrog. 

So I think talking about this just in the context of getting China to be dependent on our chips misses the larger historical stakes that are at play here, and they go far beyond the question of whether there’s an additional American chip in a Chinese missile that’s already quite good. It’s a bigger question than that. It’s a question of who dominates the next round of technology. 

BRANDT: Just a small point, but I, you know, worry about the piece of this that entails our government taking a 25 percent cut of these sales, especially in the context of us taking equity stakes in other companies. I think, you know, if we want to be able to export the full AI stack, like, part of that is going to have to be our companies, you know, being perceived as credibly independent actors, you know, of course, like, responding to market signals and regulation and all the rest. But I think we’re sort of not doing ourselves a service in that—in that realm. 

SEGAL: Great. 

Lilly, I’m going to ask at least one more question, but I don’t know if you want to remind people how to ask a question and then I’ll turn it over after I ask this next question. 

CHANG: I can jump in for Lilly or for our colleagues in AV. You can—yes, in the chat Lilly put the instructions. 

(Gives queuing instructions.) 

SEGAL: Great. 

So, broadly, what do we think happens next? Is there something that the U.S.—there seemed to be a fairly broad consensus that this was kind of a unilateral own goal. Is there something, you know, one would want for us to do to perhaps defend after that goal has been made or, you know, score the next goal? 

And do we think Congress is going to act? You know, there is a bill floating around, the GAIN AI Act. Would this mobilize greater support, or are we going to still see people unwilling to question the president right now? 

Chris, why don’t you start? 

MCGUIRE: Yeah, thanks. I think a couple things. 

When we’re thinking about what’s next, I mean, look, you’re right that Congress has kind of implemented—has stepped in here. There is already efforts to get several pieces of legislation into the NDAA, namely the GAIN AI Act which would basically have said chips can only go to China if they’ve already gone through American companies or American companies have already had, like, right of first refusal and refused them. That got very close to the finish line and I think showed that there is pretty significant bipartisan frustration here with the approach, but did not get across the line. 

I think that we’ll see, if and when that kind of boils over and you start to see a willingness of Congress to actually assert itself and dictate policy in this space. But I would be a little skeptical if that happens right away. I think that we’ll probably see the policy play out. 

I think there are a couple caveats that we’ve been very harsh, and I’ll say a couple things. One is implementation matters, right? Like, if—there’s a big difference—one single H200 chip doesn’t actually matter. The whole point here is aggregation. So if there are, you know, reasonable numbers of H200 sent, it probably—maybe it’s—maybe it’s, you know—maybe it’s not necessarily necessary, but it is probably a less strategically impactful move. Very large numbers of H200 make a big difference. 

The other thing that was reported as a potential shift to administration policy of releasing the most advanced chip on an eighteen-month delay, and this would be a really significant move because it would also mean that starting a few months from now we’d be talking about Blackwell chips, and you know, a few months after that we’re talking about the kind of Blackwell ultra chips, right? Like, we will just constantly be in this cycle. I do think that there needs to be some clarity, that this kind of constant, you know, changing of U.S. export control policy is probably not sustainable one way or another. And whether it just means the U.S. is just going to decide we’re just going to export chips to China and that’s how it’s going to be, and then Congress says OK that’s how it is, or whether there’s an actual policy that’s articulated, I kind of think at some point it might happen but I think it might take some time to shake out. 

I’ll also just say one last thing. There’s one other way that this move makes sense that we haven’t really talked about but, like, could be justifiable: It’s that AI actually doesn’t really matter, right? Like, what if AI is really overvalued? And I mean, in all seriousness, like, if these are overvalued assets, and AI is actually not going to amount to that strategic of a technology, then this would be a logical move. And in some ways I think it’s sort of the bet. The problem is that that’s really incongruous with the approach that the administration is taking on AI generally across the board. So just kind of putting that out there for how to think about it. 

SEGAL: Rush or Mike, I thought—or, maybe I saw you wanting to— 

HOROWITZ: Yeah. I mean—I mean, I would say two things. 

One, I mean, the other—the other defense of this, frankly, like, I would offer is, like, I’ve written elsewhere with Lauren Kahn that we should think about the AI competition as a marathon rather than a sprint. And if you imagine AI competition between the U.S. and China as a generational marathon, then the export of any one chip in particular, you know, isn’t necessarily—isn’t necessarily critical because no one is inherently going to be able to lock anybody else sort of out of these markets. In which case, like, let U.S. companies export, the—I think as like probably the, like, best argument that at least, like, I could muster—I could sort of muster on this decision. 

I mean, the other thing I would say is that the—given that we—you know, the dynamics that Chris talked about and the fact that we see lots of—you know, there’s a lot more change than continuity, why don’t we say, in Trump administration policy on sort of tariffs and things like export, you know, controlling these kinds of chips. It’s entirely possible that, say, if the Trump administration decides that China’s not following the broader trade deal that you could see the—you could see controls reimposed, or even more than what the Biden administration, you know, had—you know, like, had done. Like, I’m—I would be very surprised if this is the last that we have heard of this, especially because I would expect China to attempt to figure out how to use the leverage they have gained here from this unilateral concession to make a play for the Blackwell chip sooner rather than later. 

SEGAL: Thanks, Mike. 

So we have a couple questions. Why don’t—Will, why don’t we take a couple questions, and then I can also ask— 

OPERATOR: We will take our first raised hand from Trevor Hunnicutt. 

Q: Hi. Thanks for taking the questions. 

So I have two. First, I was just wondering, you know, the NVIDIA stock pared its gains after this announcement based on a report that China would still keep a pretty tight permitting process for domestic buyers of these chips. Just wondering if you had any reaction to that. 

And then also just more at a high level, you know, if this announcement from the president yesterday was kind of the appetizer, what does it tell us about what we should expect from the main dish of his state visit to Beijing in April? 

SEGAL: Go ahead, Rush. 

DOSHI: Thanks. Hey, Trevor. It’s a great question. And you know, it’s something certainly a lot of us have been thinking about. 

What I would say is that we have a precedent for this. When the Trump administration decided to export the H20 chip, China, you know, essentially also decided to basically say no thank you, to prevent that chip from being imported. And at the time that was a puzzle. People were asking, why did they do that? If it’s really true they want, you know, better American chips, why are they refusing the H20? And the reason why was in part they wanted—they wanted even better chips. They’re trying to indigenize as much as they can AI—you know, AI chip production. And so to some degree they’re worried that if they start importing some chips, you know, that maybe that goal will be damaged. 

But the other side of the ledger is that they also know there are some chips that they simply can’t make, so there’s no domestic competition for those chips anyway, and those are the ones that they want. And so it’s a very sensitive and sensible policy. What they’re trying to say is the chips that we need are the ones we want to buy from you, and anything less than that, that we’re not sure that we need as much, we’re going to make it harder for you to sell it. And when we say no to that other stuff, you’re going to come back to us and ask to sell the better stuff. That was the gamble that they made. They played a strategic game. They said, I don’t want your JV-tier product. And our companies were so desirous to sell them that they decided to go up higher and offer them even better. In retrospect, that play looks like it worked out. They went from a chip that wasn’t as good to now the second-best chip, all on the power of no. So that’s part of the story, I think, as to what happened here. 

And at the same time, they’re going to continue to squeeze NVIDIA a little bit because they want to get even better stuff out of NVIDIA. They know there are better chips coming, and if they can find a way to say, well, you want to sell us the H200 but we’re going to make it—you know, we’re going to buy a lot of it but we’re going to make it a little bit painful at times, they have more leverage to get even better chips later. That’s the playbook, and I think, again, they’ve shown that it’s working. 

There’s, of course, more to this. There’s more complexity here than I’m fully capturing in this short answer. But that’s my initial take. 

SEGAL: Mike? 

HOROWITZ: Yeah, my view is very similar. That’s kind of what I was hinting at in my previous answer when I suggested that, you know, this might be, you know, part of a Chinese play for the Blackwell. I think their—you know, that FT report about the—you know, that they might impose strict limits on H200 seems like tailor made to have—to try to encourage, you know, those that support exports to go back to the Trump administration and ask for—and ask for even sort of more advanced chips. 

I think it also suggests some of the dynamics that we were—you know, that some of us were talking about before where, you know—you know, some of the mechanisms through which China may continue to subsidize its domestic industry, which would prevent essentially flooding the markets with U.S. chips, from driving out Chinese competitors. 

SEGAL: Let’s take the next question, Will. 

OPERATOR: We will take our next question from Miriam Backsh (ph). 

Q: Thanks so much. 

So you mentioned that the GAIN AI Act almost made it into the NDAA. When that—when it was clear that that was out, it seemed like lawmakers wanted to include the SAFE Chips Act. How much do you think this announcement, this move by the Trump administration was a reaction to that, to get it in now while they can before the SAFE Chips Act would freeze current controls on compute? 

And the second question I have is I think related a little bit to what you guys were just talking about. I think the announcement—the recent announcement was for some H200 chips. Maybe Ms. Brandt can speak to the extent to which she’s concerned that any sort of controls on which entities would be allowed to buy might be affected by front companies and related affiliations rule or what have you. 

So two questions. Thanks so much. 

SEGAL: Chris, you want to take the first one? 

MCGUIRE: Yeah, I can take the first one. Jessica will speak to you a second. 

So it’s a fair question. I think in my opinion it’s probably the reverse. I think you’re right that once the GAIN Act disappeared that then the SAFE Act popped up, and the goal of the SAFE Act really was to codify the existing administration position on export controls, right? I think that there were many people in the—on both sides of the aisle who said, OK, well, we could have our differences on the H20 debate, but let’s all agree that this is where it should stay for a while. 

I kind of actually think that the SAFE Act was an effort by Congress to preempt concerns about the Trump administration doing this, but you could really cut it either way. It’s hard to tell for sure. I think that, look, the Trump administration has been deliberating, you know, changes to AI chip export control policy, you know, ad nauseum but certainly, you know, in detail for the last few months, so I don’t necessarily think this was a direct response to Congress. I think it probably was, you know, a result of kind of lobbying and global events, and the timing with the NDAA was a little bit coincidental. But hard to tell. 

BRANDT: And on the second piece, you know, I think front companies and opaque supply chains make it fairly plausible that chips, you know, could be diverted either to the PLA or defense contractors in ways we wouldn’t want. 

SEGAL: So that in some ways raises my—an additional question I have, which is, you know, considering we also had—at the same time that this announcement was made we had the attorney general in Texas bring this case of—you know, against someone who was illegally exporting H200 chips to China. So what does this say about the—you know, kind of the status or the coherence of our—of our export controls? You know, was this just a kind of—you know, an issue about one hand not talking to the other hand? Or does this, you know, speak to kind of a larger debate inside the administration about what should be happening? Or how do we kind of interpret this? 

HOROWITZ: I can drop in to start, I guess. 

Like, I—I think, like, the—I mean, there are frankly, like—actually, like, a pretty simplistic reading of that involves—that will not be, like, interesting at all, which is that—when they were illegal, like, somebody was being prosecuted and, like, the left hand didn’t talk to the right hand about, like, what policy changes were coming, which would be entirely understandable in the context of the United States government and it’s, like, size and shape. And sometimes agencies don’t talk to each other when they should. 

So, like, all that being said, I think what this does suggest is there’s continuing—with this probably—also the “yes, and” there is that there’s probably also continuing internal tension over what China policy should look like with export controls as a lesser included, in some ways, of that. You know, there were all these discussions, you know, prior to the release of the NSS about the extent to which there were, you know, very last-minute debates about what the China language should look like.  

And it wouldn’t surprise me if those were also—you know, essentially if some of those same debates were playing out in the context of, like, how these export control conversations were happening. And I think the Trump—and I think the president probably doesn’t mind the notion that they are going to allow for this sale, but there’s some uncertainties still kicking around in the back—in the back of people’s minds about it. Because, you know, I think the president believes in his ability to leverage uncertainty to get better deals for the American people. Whether he’s correct or not, like, we could debate about, but I think, like, that would be an arguable theory of the case. 

SEGAL: Rush. 

DOSHI: Yeah. Just a kind of quick point, building on what Mike said. There’s been this long view that President Trump, you know, is sort of a China hawk, or tough on China. But the reality has always been more complex, right? There’s been moment where President Trump, you know, had hawkish policy preferences, particularly on trade in the first term and then after COVID. But really, you know, his staff, and the president—the president and his staff have had sort of different approaches. The president’s staff has been much more competitive and confrontational on China policy, certainly in the first term, a little bit in the second term, than the president has been. And that delta has always been there. 

If you look back at the first time—and I’ll give you a few quick data points that explain how we got here now. In the first term we now know that President Trump offered to lift the export controls on Huawei and ZTE in exchange for Chinese—you know, basically to protect a trade deal with China, and partially as a favor to Xi Jinping. We also know that President Trump at that time was willing to tell President Xi that he should go ahead and build the camps in Xinjiang that were being used to inter Uyghurs. That that was sort of his own business. And that on Hong Kong later on that that’s essentially just a sovereign matter for China to worry about, and he wouldn’t have anything to say about it. At one point he was asked, do you stand with the people of Hong Kong or Xi Jinping? He actually said, I stand with President Xi. 

Now, he probably was speaking somewhat defiantly and loosely, and maybe I’m being uncharitable. But the point is that if you put all that together, you see a package of policy across the board where his first inclination has not always been to go hard at China. And so now you come to this moment on export control. And there were rumors before the Busan meeting that he would actually lift the controls on far better chips than the one that he just authorized to be sold to China. And it didn’t end up happening. No one’s entirely sure why not, but it's possible he was persuaded at the last minute that it wasn’t a good idea. And yet, here he is now several weeks later making a change in policy again. 

So it suggests that his instinct is always to look for ways in which he can stabilize relations with other great powers, because he believes that that’s what a strong leader does. And he believes that that means, or validates, I think, his overall strategic approach, which is to kind of pursue American interests from a position of stability on one front, so he can pursue a more competitive approach on the other. And when you look at this decision in that context, it makes a lot more sense. That probably this decision was taken, to answer your question directly, by the president, with a few aides telling him it’s a good idea. But I bet a significant amount of that administration is opposed to this policy, including many of the Cabinet officials who are currently not so vocal about it. 

SEGAL: So, of course, the other issue with export controls is allies and partners, and the controls that they have in place. And around chips, probably the most important is, you know, ASML and EUVs, and semiconductor manufacturing equipment. Do we think this is going to have—I mean, the Dutch have their own problems with China right now. But do we think this is going to have spillovers with the Dutch, the Japanese, others in some other critical areas of, you know, maintaining control around chips? 

HOROWITZ: I mean, I can jump in a bit on this. Like, I think the—I think, like, taking a step back, because I think we need that to understand the allies and partners piece, is, you know, export controls are intrinsically market-distorting. And in that—and we know from many other product categories. And AI is undoubtably the same—that export controls cause—you know, cause innovation, cause substitutions, cause, like, lots of different, like, kinds of things. The question—you know, they generally buy you time. They don’t prevent people from doing things.  

You know, what had happened over the last two administration was this effort, coordinated with allies and partners, to not just deny China access to the most advanced chips, but to the deny them access to the manufacturing equipment they would need to make it easier for them to figure it out. Essentially, those sort of actions by the Dutch and by the Japanese in particular are part of what would make the—made the export controls conceptually relatively more effective, by delaying the onset of some of that—some of the impact of the distortion inherently caused by export controls.  

In a world where the Trump administration is willing to export more and more advanced chips to China, and where it’s making clear to allies and partners in other areas that, you know, it’s all about the money, in general I would be nervous about the extent to which, you know, other countries might—the extent to which allies and partners over time, whether they will be still be willing to restrict the export of some of those advanced pieces of equipment to China. Because, you know, if they don’t feel like the Trump administration has their back on this issue in general, or potentially if there’s spillover from other issues, it could make them more likely then to export that technology to China as well, which would, in and of itself then, be a different level of game-changer for China. 

MCGUIRE: Yeah. I’ll just hop in here and say, like, I think this is—this is something to take very seriously. Just to highlight, the SME controls, the controls on semiconductor tooling, are the reason that we can do chip controls, the reason why we can prevent China from, you know, accessing larger amounts of compute, to prevent them from buying it from us, is because the core of it is we prevent them from making it themselves. And these have actually been very effective. You know, I think that there’s gaps in them, but fundamentally, like, the U.S.—the best U.S. chip is right now five times better. In two years, it’s going to be about twenty times better. And the reason is because Chinese chips have basically stalled and U.S. chips are getting better and better as they’re going on more advanced processing technology. 

Actually, when negotiating the controls with allies it was really important for the United States to first put on controls on advanced chips, because until we did that allies would go to us and say, why should we control the equipment to prevent China from making the best chips when they’re just going to then turn around and buy the chips from you? It doesn’t—it seems like you’re just kind of cutting off our market and then benefitting yourself. So what happened was in October 2022, the United States said, fine, we’re going to cut them off from our companies. Now it’s your turn. And then that resulted in the allied agreements that we have today. 

So if there is a—if we kind of walk back from where we are on the chips side of it, and particularly given now there’s, like, the kind of kickback arrangements, and there’s, like, direct benefit to the U.S. government, it makes it harder to take seriously, and also kind of undercuts one of the core arguments. Now, I will say the U.S.—the allied tools are so—and particularly the Dutch tools—are so reliant on U.S. technology that even if the Dutch were to say, we’re going to lift our controls, the administration would have the ability to say, sorry, we’re going to extraterritorially control the tool because it has U.S. tech in it that we’re not going to let go. That might be a very unpopular decision with allies. It might also be logically inconsistent. But it is not a move that would be inherently off the table. 

SEGAL: Thanks, Chris. So we don’t have any other questions, so I’m going to go lightning round. Which essentially is you can basically say whatever it is I haven’t asked you yet, or whatever last point you’d like to make, things we should be looking for, something about U.S.-China relations. So I haven’t given you a lot of time to think about it. So I will take a volunteer first, or otherwise I will just call somebody. Go ahead, Rush. 

DOSHI: Well, first, let me just say thanks to you, Adam. And thanks to all who have joined this call. Hopefully, again, it’s been useful to you. And we’ve tried to offer a variety of perspectives. I just wanted to say, as a kind of final thought, one of the reasons why this is so important is if you look at the U.S.-China competition you do see a China that is surging on every industrial front. A China that has two-time the American manufacturing value add, two times the American power generation, and three times the American core production, you know, ten or eleven times the American steel production, and twenty times the American cement production, that dominations in a lot of other industrial sectors, right? From electric vehicles, to drones, to pharmaceuticals, to solar panels. And of course, you know, the batteries upon which electrification around the world now relies.  

So you have a true industrial colossus, one thing that the U.S. used to be but isn’t at the moment. And that’s why I think it’s so interesting to see how the decisions that are being made on AI tie in with this reality, because you could argue that the one area of kind of singular American advantage, and when other areas are currently atrophying, is this incredible compute advantage that the country has. That it produces essentially the machines within its allied ecosystem, and the chips at those machines, upon which this most transformative technology is likely to rely. And so we’re lucky to have stumbled upon that as one area of singular advantage. 

And that is why it’s so interesting to see what the decision—what decisions are made about how to husband and steward that remarkable American asset. And so if that essentially is given away or sold, perhaps even at a discount, to a competitor, it does suggest then that one of the last places where America can seek economic and technological competitive edge is potentially going to be lost. Not right away, but down the road. And I think, to Chris’ earlier point, one thing that is very important for all of us to pay attention to, is this just about the H200? Is the Blackwell coming next? Or are we really adopting an eighteen-month lag, where China’s only going to be eighteen months behind us, for as long as we can keep them there? Because that latter approach is going to mean essentially that they don’t just stay eighteen months behind us, but actually catch up when they begin applying this technology at scale. Thanks, Adam. 

SEGAL: Thanks, Rush. Why don’t we go to Mike next? 

HOROWITZ: Sure. I mean, I would just say that I do think it’s important to understand this policy change in the context of broader Trump administration choices about China, which, you know, given the—which, you know, we have this NSS that backs off the focus on China as a strategic competitor, but, like, that could change. You know, change is more likely than continuity when it comes to, you know, many thing that the Trump administration has done, particularly given the, you know, what can be a very back-and-forth relationship between Trump and Xi. And so I think that even though, you know, industry may be sort of cheering today the possibility of greater exports, it’s entirely possible that there’ll be reversals tomorrow. But I think what won’t change is that China will very cannily and strategically attempt to use these concessions from the Trump administration to gain even greater access to U.S. semiconductor technology—and almost certainly sooner rather than later. 

SEGAL: Jessica, you want to go? 

BRANDT: Sure. I guess I would just say, you know, there are meaningful supply constraints. And my worry is that if NVIDIA does not find the excess capacity to sell to the Chinese market, then this could mean that we’ll be selling less chips to our own companies and to our friends. And that would be, you know, as Mike said, like, the definition of an own-goal. 

SEGAL: Thanks. And, Chris, you can take us home. 

MCGUIRE: Great. One thing I’d just say, to put this in context, is I think that we think about these AI chip controls as kind of unprecedented. And they get a lot of attention because NVIDIA is the largest company in the world. But I would actually argue that they actually have a pretty strong precedent in some ways, that they actually—the United States has always protected its compute advantages. And its computing advantages have always been essential to military applications. Ther have always been controls on supercomputers. We never let the Soviets have our most advanced supercomputers, right? We used our most advanced supercomputers for, you know, nuclear stockpile stewardship, and all sorts of advanced applications. 

The difference now, and the reason why these controls went into effect, is that NVIDIA basically has brought to market supercomputers in a box—NVIDIA, and AMD, and Intel, and, like, the entire semiconductor industry has done that. Which is an incredibly valuable product, right? That is why NVIDIA has become the most valuable company in the world. It’s unbelievably useful. It’s resulted in transformative change across the—across the kind of AI and tech industry. But that doesn’t change the fact that it has national security applications. It has always had national security applications. They U.S. government has always seen those.  

Just now it also has massive commercial applications. It’s creating tension there, but that fundamental core of the importance of computing power to kind of U.S. national power and defense intelligence applications has never changed. And this is the kind of logical extension of it. 

SEGAL: Thanks very much. Thank you all for a great discussion, and for all of you for attending. I hope you will take advantage of all the resources that CFR has to offer on this and other issues. And all of us can be reached online. 

Ben, I don’t know if you want to say more about follow up on this meeting. 

CHANG: Just to thank all of the panelists and to say that if you have any questions or would like follow up, please email [email protected]. Thanks, everybody. Bye-bye. 

(END) 

This is an uncorrected transcript.