← The AI Policy Podcast
The AI Policy Podcast
China's EUV Manhattan Project and Export Control Mythbusting with Chris McGuire
01:28:28
In this episode, we're joined by Chris McGuire for a conversation about AI and semiconductor export controls.
Loading summary
Transcript120 lines
- [00:00]
A
Foreign.
- [00:11]
B
Welcome back to the AI Policy Podcast. I'm Greg Allen and today we've got an episode that I am very excited about. It's something I've been thinking about doing for years but have been unable to do because today's featured guest has been in government service for so long and unable to do public facing interviews. But he recently escaped public service earlier this year, is now a senior fellow for China and Emerging Technologies at the Council on Foreign Relations and is allowed to speak his mind publicly and loudly, which is what we're going to benefit from on today's episode. So my friend and colleague, Chris McGuire, formerly of the National Security Council, formerly of the State Department, formerly of the National Security Commission on Artificial Intelligence, a genuine expert on all things at the intersection of the United States, China, AI, semiconductors, semiconductor manufacturing equipment, a ton of my favorite topics to talk about on this podcast. So, Chris, thank you so much for coming on.
- [01:08]
A
Thanks for being here or thanks for having me, Greg. I've been looking forward to coming on, so I'm glad we can finally do it.
- [01:13]
B
So today we've got a very kind of fun experimental format for the episode. We're going to do it in a mythbusters style format. So we're going to go through some of the most common arguments you hear against AI, semiconductor, semiconductor equipment export controls, and we're going to rate those arguments as either fact, fiction or somewhere kind of in the middle. And we're going to give you our best assessment, our most honest assessment of like where the actual truth lies in these enormously complicated issues. But before we get to all that, I want to give Chris an opportunity to introduce himself to all of you because now he's left his life of secrecy in the national security community. So, Chris, how did you get into technology, foreign policy, national security, and when did you first start working on AI and semiconductor stuff?
- [02:05]
A
Yeah, thanks. So I've been a civil servant until this year, as you noted, for about 10 years. I started in government. I was a State Department employee the whole time. I actually started working on nuclear weapons policy. I studied that in grad school, kind of came out and was working on initially multilateral nuclear policy, but then I had the US Russia nuclear portfolio at State Department, so doing arms control negotiations, negotiations, a new START treaty, INF treaty, and actively thinking about that. But also how does that impact broader US strategy? And I pivoted in 2018 into the kind of emerging tech space partly because I realized really, that is, while nuclear weapons are obviously super important to our deterrence posture, it really is last century strategic competition. And the strategic competition that's going to embody the coming century is emerging technology and particularly artificial intelligence. So I pivoted over from that, that into this, which in many ways actually has been a little more of a seamless transition than I would think. Obviously they're different topics, but regardless, my niche has sort of been being able to explain, understand and explain kind of highly technical topics to policymakers and then seeing the strategic implications of that. There are people who know more about the technology than me and there's people who know more about the kind of broad grand strategy than me. But sort of at the intersection of that is where in both the nuclear and kind of AI spaces I've been able to live. So, yeah, after that, as you said, National Security Commission on AI came over.
- [03:39]
B
And for folks who don't know what that is, can you just remind everybody what the commission was? Cause it was a huge deal at the time and you had kind of a fun job on it.
- [03:48]
A
Yeah. So it was pretty prescient. In the NDAA, the National Defense Authorization act in 2017. And then starting in 2018. No, sorry. In the 2018 NDA and then starting in, I think 1920, 21, the Congress basically set up a commission, kind of like a la. The 911 Commission. Right. To study. Hey, what are the implications of this? We think this is important. Go spend a couple years figuring it out. And we had kind of commissioners that were selected from both sides of the aisle in Congress. Eric Schmidt was the chair appointed by the Republicans. Bob Work was the vice chair appointed by the Democrats. And then we had a slew of industry luminaries, Andy Jassy, Safra Katz, a lot of people with national security experience, Chris Darby from IN Q Tel, Jason Matheny, many others.
- [04:41]
B
So.
- [04:44]
A
That was really Congress saying, we think that AI is going to be very important and kind of give us the forward looking recommendations. I think it was actually quite prescient. We developed a lot of the kind of groundwork for things that eventually turned into even the CHIPS act into some of the kind of broader kind of talent moves that the government has made. And specifically on the commission, I had the portfolio of kind of generally US China competition and then also semiconductor hardware policy writ large, which obviously has become kind of one of the biggest contestable spaces in the AI competition.
- [05:21]
B
Yeah. So I have to just interject here with a bit of connecting to my own personal story. So when the first chip export controls came out of the Biden administration, so there was stuff that was going on in the Trump administration. And I'd love to hear, you know, whether or not your NSAI work, you know, led you to interact with the first Trump White House. But in the Biden administration, when the October 7, 2022 export controls dropped, everybody's like, this is such a shock. How could this have ever possibly happened? And I'm like, did any of you read the NSCAI report? Like, this is what they told them to do. They're just. They're doing the thing that they said they were, you know, thinking about doing. So it was a big shift in US Policy, but I was kind of surprised the extent to which people surprised. And I felt like in the analytical community, I had this advantage, which should not have been so much of. Advantage of. An advantage of just paying attention to the nsca. I work while it was going on.
- [06:18]
A
Yeah, there were a small group of people that were. That were pushing for things along these lines. And look, actually, there was work that happened even before that. Actually, the NSCI released quarterly reports partly because. Because we realized that this was moving very fast, and if we just waited two years to release a final report, then we'd miss a lot of opportunities for engagement. But in some of our initial reports, we were saying we should control EUV.
- [06:37]
B
Prevention, which Trump administration did.
- [06:39]
A
And we were talking with the Trump administration. Trump administration heard that. I'm sure they were doing their own thinking as well at the time. So I don't want to take all the credit by any means. They also had the foresight. But, yeah, closely liaising with them, which they did. And I think that then coupled with then the Biden moves in 2022 to significantly expand those controls on equipment and then also start to control the chips itself themselves was. Was sort of the logical extension of both where the Trump administration started. And then, yes, as you say, like the recommendation, the nsci, which really saw that hardware is the area of differentiation between the United States and China. It's where we have the most opportunity to make moves that have a big strategic difference.
- [07:18]
B
So the State Department sends you to NSCAI during the second half of the first Trump administration and then during the B administration. Am I correct in saying the State Department details you to the National Security Council?
- [07:33]
A
That's right. I really pushed the State Department record for consecutive details, I think.
- [07:37]
B
But they're like, do you work here anymore?
- [07:39]
A
My gosh, you know, there's own challenges there. But no, I mean, look, we're all. We're on the same team here, right? You're a government Employee, like it's, you're working at, you know, the commission to kind of advance government priorities, also advanced State Department priorities. There's logic in like detailing people into these bodies and then similarly into the White House. So I started in the Office of Science Technology Policy Policy for almost a year, and then I was at the National Security Council for two and a half years where I broadly had the US China portfolio and by then was the deputy Senior Director for Technology and National Security. So broadly oversaw all of the work we did on technology protection, so export controls, but also investment screening, cfius, icts, anything in tech protection space. And then broadly had the semiconductors portfolio for a while too. And then broadly AI and kind of oversaw the broader work we were doing on emerging technologies, including biotech, quantum, et cetera.
- [08:40]
B
So a really, really important portfolio. And you continued doing that category of work into the second Trump administration, am I right?
- [08:47]
A
That's right, yeah. So then I went back at State Department, I was a senior advisor in the technology Envoy's office and continued to work closely with the White House as they kind of grappled with, with the same policies and debating kind of what do we do with the things that the Biden administration handed us over. Obviously, this is a very fast moving technology with huge implications. And, you know, the, the importance of this to all diplomacy rapidly increased over every year, over the last, I'd say three years. And it just has become so central to all of our relationships, but then kind of also kind of broader strategy. So was very helpful to be able to kind of work with the team. And, you know, these are complicated things where there are also a lot of, a lot of new stuff that had a lot of deep thinking behind them that, you know, there were a lot of people at the White House and the agencies that were able to, you know, we were able to kind of get up to speed on. And then they're obviously now running with it, you know, their own direction.
- [09:43]
B
And I'm, I'm kind of amazed and delighted that through all the various, you know, stages of your career that we just talked about, you and I somehow managed to be professionally relevant to each other. So I remember when you were first and foremost a nuclear weapons guy and you and I were having conversations about the extent to which Russia was introducing autonomous, robotic and AI capabilities to their nuclear forces and thinking about U.S. policy. Back when I was at the DOD, then at the NSCI, I was still at the Joint AI center at the DOD and you know, obviously we were very intimately paying attention to what was going on nsci because in many ways you guys were like grading our homework on behalf of Congress. And then when you were in the Biden White House and doing export control stuff. First I was involved in that in a minor way while I was at dod, but then that became a huge part of my work here at csis. So I benefited from knowing you and paying attention to your work through all those stages. And that's why I'm so glad that now the audience gets to benefit from what you're up to. So you just testified before the House Foreign Affairs Committee on semiconductor manufacturing equipment, I believe was the primary focus of your testimony. Before we get into the sort of Mythbusters part of the segment, I want to connect, you know, your testimony and your prior government service work on this topic to a very hot story that I'm sure a lot of our audience is looking for folks to unpack, which is on December 17, Reuters reported that a Shenzhen Shenzhen team composed of formal ASML engineers, mostly Chinese nationals, built a working prototype of an EUV machine in early 2025. You reacted to the news on X saying that parts of this story are likely propaganda. Can you explain your reasoning? What do you think is real in the story? What do you think is meaningful in the story? What do you think is not meaningful?
- [11:34]
A
Yeah, and look, this is a nuanced situation that involves more in depth conversations. I'm glad we're doing this. I think first of all we should just to the basic premise of my testimony and the ideas behind this, why this matters is obviously the cemetery. Manufacturing equipment controls are really the core of our ability to, I would say, constrain China's technological progress, be it in AI or anything else. The fact that China, China can't make the most advanced chips themselves, which it can't because they can't get the equipment to do it is the lever that we have. So if the, the foundation fails and the equipment controls fail and China gets at par with TSMC and others, then it doesn't, you know, we can't, we wouldn't even control the chips because they could just make them themselves. Like the whole foundation kind of falls apart. Now that said, they are lagging significantly and that way I testified in Congress, is that I think the equipment controls are, and they're working better than some people give them credit for, I think. And I think you can see that in some, the performance of their fabs and the gap in quality between our AI chips and theirs, which we can talk about in a minute, which is good news. But Also because this is the foundation and because if this regime, the equipment regime fails, then our ability to regulate technology globally constrain China at all in anything becomes zero. So our risk tolerance should also be zero on that. So I think that there are holes in the regime. I think the title of the hearing was export control loopholes. And look, the Biden administration took a lot of steps. I would agree with many things that the Biden administration did. I would disagree with some things that the Biden administration did. You've been very public on this. Personally, I think that there are things that we should have gone farther and Trump administration should take those farther. So those were things I wanted to highlight. Now that's bringing this back to the story that you mentioned today. So there was a story in Reuters that said China's making progress on euv. And obviously EUV extreme ultraviolet lithography machines are the most complicated of all of the manufacturing equipment.
- [13:47]
B
So of all the machines that planet Earth makes, probably.
- [13:50]
A
Right, that's correct. It's the most complicated machine that humans have ever made by a pretty significant margin. So if, if you know, China is able to, the theory behind the controls is actually this and some of the other machines behind it because they are the most complicated machines humans have ever made. They are also logically the hardest thing for China to indigenize. So it's not to say they can never do it, but it is the most difficult thing. And we can probably hold them back for a significant margin of time by, by forcing them to develop the entire supply chain behind these themselves. If they can do that overnight, they can do anything overnight. Maybe that's possible. But there's also, it's logically consistent that they can actually be very good at indigenizing but still struggle with this.
- [14:32]
B
Yeah. And you, you said, you said a really important line there, which was indigenize the entire supply chain. Yes. And that's, that's kind of where I find fault with this Reuters story, where there's, there's legitimately good reporting in this Reuters story that introduces new information, at least to the public conversation. And so I don't want to diminish it too much, but it says China built an EUV machine. And I think a more accurate characterization would be assembled an EUV machine because what it's pointing out is that most of the parts are actually parts from the same supply chain that ASML uses, same components supplier as asml. So really this is not a story of indigenous innovation so much, although that, that, that is present in this story. But it's mostly a story of export control evasion and getting around our existing export controls by identifying component suppliers. And that's not so different from where the Soviet Union was during the Cold War. There's a lovely set of declassified CIA reports on the Soviet Union's efforts to steal semiconductor manufacturing equipment. And don't get me wrong, when you're trying to reverse engineer machines, having a copy really helps. It is. It is good to have a copy if your goal is to reverse engineer this stuff. But there's a very big difference between having a machine, being able to make that machine domestically, being able to make all the components of which that machine is comprised domestically, and then between that, there's still another huge gap between being able to make a machine that will actually do what it does in the lab, do the same thing in the factory a million times. And then there's another huge gap between being able to do it once and being able to build a thousand of these machines. And so it's not like China is making no progress in this story, but I think it's difficult to convey in an article like this just how far what they're doing is from where they would like to be.
- [16:42]
A
That's right. And these are not only the most complicated machines in human history, but they actually are the most complicated machines in human history that produce the most complicated products in human history. So they themselves are hard to operate that also rely on the most complicated supply chains in human history. Right. Every element of this is at the max. There are hundreds of thousands of parts in these machines. It is very, very difficult to do. And even if you can do it, it is very, very difficult to do it well. And then to do it well at scale is even a level beyond that. So these things are very hard. And no one thinks that China's never going to have an EUV machine. But the question is, are we talking about the 2000s, the mid-2030s, 20, 30, or like the late 2000s? There are huge differences for US policy, US strategy, and the future of AI to the answer to that question in each of those segments.
- [17:28]
B
And personally, and let's just connect.
- [17:30]
A
If we're in the mid-2030s, we're probably in a decent spot. Right. If we're in, like, the late 2000s, then that's going to pose pretty significant challenges.
- [17:39]
B
Yeah. And so just to connect it to what the Reuters story is saying, and they're citing sources who actually are these former ASML employees who are working on this EUV machine. They're saying that it, quote, has not produced working chips. Sources say government is targeting 2028 for working chips, but sources say 2030 is more likely. So there we go. They have a working EUV machine. You know, that they've salvaged all these components to reassemble them. It's not working all the way because it can't make chips yet. But as I said before, ASML first made a working euv machine in 2001 and it was not commercially functional for many years after that. I want to say 2010 was the first, like prototype EUV Mach working in an intel fab. And so, you know, when we say 2030 is when they might start making working chips, that's still a long time from having a commercial market impact or a national security market impact in terms of making competitive chips. Don't get me wrong, I think Huawei will be faster than ASML at every stage of this story because they get to copy ASML's homework and ASML had to do it for the first time. But the point is like, like this is really, really hard. And just to take another point of comparison, China has a ton of copies of DUV immersion machines. They have thousands of those. They do not have a domestic replica of a DUV immersion machine. They have thousands of copies of advanced deposition and etching machines. And SMIC is trying to use the local Chinese replicas of those. And it's really causing problems in their production process. The yields at SMIC for AI chips, for 7 nanometer chips General, remain dreadful. And that's because the local copies of foreign equipment, the ones that are much easier to copy and much older than ASML's current state of the art, are still really, really hard to do in a good way.
- [19:43]
A
That's right. And look, we've long known that China has had an EV program that is not surprising to anyone. It's certainly not surprising it was made in China.
- [19:51]
B
25.
- [19:53]
A
They would be foolish and it would be a dereliction of duty for them not to have an indigenous EUV program, just like they have an indigenization for every other element of the supply chain. But whether or not those are actually achieving great success is a bigger story. So the fact that they have this secret program surprises me. Zero. I don't see anything in the story that really surprises me in terms of their capabilities that shows any kind of great leap forward. I think that there's still a lot of work to do like they said. But the last thing I'd say is we have long known that there's a really significant Chinese propaganda campaign around their capabilities. And actually, Jake Sullivan said this publicly in July 2024, has said generally there is a Chinese propaganda campaign around the idea of export controls and trying to push the narrative that export controls don't work. This is not a coincidence that every morning in the South China Morning Post, you're going to see new news about some Chinese breakthrough that actually is a chip that's 10,000 times better than Nvidia's chips. How many of these stories have we seen? But how many of those chips have we actually seen come to market? Zero. Right. Because they don't exist.
- [21:02]
B
Yeah.
- [21:02]
A
So this. And there's a logic there, right. The strategy is let's convince policymakers that export controls don't work so that they pull them down so that we can get the equipment that we really need or the chips that we really need and we're currently banned from. That is a very logical and viable strategy. We're seeing it on the chip side and I think the kind of report I put out was trying to explicitly push back on that, saying that actually there's this narrative that Huawei is really good, they're making all these chips, they're really great, they're a competitor to Nvidia, but you actually look at the data, including data that Huawei is. Has said themselves, and it's clear that actually they're not a viable competitor to Nvidia, that there's no way the math actually adds up. And you know that that is part of the campaign they're trying to push in order to get us to losing controls and get chips. What I see on this side, on the story, while, look, at the same time, we obviously should be concerned about China having an EUV program, we obviously should close all the loopholes and gaps in controls and we should accept zero risk on this front. We also have to acknowledge that this story exists in the, in the kind of situance of this propaganda campaign. And it is, you know, this looks like the same playbook to me on equipment that we know China is running on chips saying, hey, the controls are futile, you shouldn't even bother with them because we're going to do it ourselves. And then if we, and then industry lobbies here to say release the controls, government does it, then they get the machines and actually we're significantly accelerating our progress. Like that is what I would do if I was a Chinese, particularly if I was faced with a very hard problem domestically.
- [22:32]
B
Right. So when you say, as you, as you posted on X that this is likely propaganda. The part that is propaganda is not that like every single fact in the, in the story is false. Like there is a Manhattan Project style EUV.
- [22:48]
A
Absolutely.
- [22:48]
B
Program 100 part. That's. The part that's propaganda is that this is easy or that they're close.
- [22:53]
A
That's right.
- [22:54]
B
And in reality that much farther. And I think, I think a lovely data point, right, suggesting like, what is the true Chinese preference? What are true Chinese intentions behind all of this comes from Commerce Secretary Howard Lutnick, who in June was sharing some inside baseball about how negotiations were going with the Chinese in export controls. So this is a CNBC transcript I'm going to read for. Here's the question. What, what took 14 hours a day? What did. Was the. What, what was the back and forth? What did the Chinese side want? And Lutnick said, well, the Chinese side always wants to remove our export controls. You know, they always try to remove our export controls. And then he go later to say, so of course they want them and they want to use them. And of course we said absolutely not. Did they ask, did they ask 20 times, 30 times? Of course they did. They do every time. And we always say no, I mean, not anymore, because now we're selling them H2 hundreds. But the point is like, you're not asking 30 times, please get rid of these export controls. If you're like, domestic EUV breakthrough is like five minutes away. It is, you know, according to the Chinese internal projections, as reported in Reuters, they are five years away from making their first chip in a lab, which could mean they're like 15 years away from doing anything remotely comparable to what ASML is doing. And like, this is one EUV machine that is made of smuggled components of EUV machines. Tsmc, I forget what the number is, but I want to say they buy like a hundred EUV machines machines a year. Right. And it is really hard to make 100 EUV machines that can make millions of chips reliably.
- [24:43]
A
Probably. Probably. I don't think ASML makes quite that many, but I'm sure they have well over 100 total.
- [24:49]
B
It's probably, yeah, maybe it's 100 total and they make 25 a year. That might be right.
- [24:53]
A
It's a lot. It's a several orders of magnitude difference, which is, which is significant.
- [24:58]
B
Okay, so that was. Well, very well put. Thanks for unpacking that really important bit of, of detail. And I do want to encourage folks in the audience to read Chris's testimony that he gave before the House Foreign affairs committee back on November 20, 2025. It's up on the CFR website and goes into more detail about the whole state of semiconductor manufacturing equipment export controls. Now let's turn to another thing that you wrote recently, which is up on cfr, which is an analysis of Huawei and overall the Chinese chip making ecosystem. So just take it away. Chris, what did you say in this report?
- [25:38]
A
Yeah, basically I tried to look at what the companies are actually saying in their own words about both chip quality and chip quantity to assess what is the actual capacity of Huawei and Nvidia to make AI compute. Nvidia is out there saying Huawei is a significant competitor and we have to be very concerned and actually we have to export to the China market because Huawei is just going to backfill and they have as good chips as us in many ways and also a ton of capacity. There's throwing out numbers of like, well, they're going to make millions of chips. But then we have to situate that in the context of the quality of the chips and the overall competition. So basically what I did was just look at the. Both Huawei and Nvidia have published roadmaps for the bet for the quality of the chip that they're going to make from now out through 2028 or 20 through 2028 and then availability into, into 29. And what you see when you look at that is that the gap in the quality of the chip is actually widening substantially. And it's a little bit simplistic that obviously there's additional measures of quality. When you actually look at performance, when you account for like FP4 versus FP8 versus FP16.
- [26:46]
B
Okay, you gotta explain to the audience what FP4 is.
- [26:50]
A
Yeah, the precision at which you can process data. Basically more advanced chips actually have capabilities to have like a multiplier on top of what they could do before. So you actually can get kind of a doubling of performance by, by moving to a lower, lower precision. So it's kind of like a advancement in technology that's separate from the performance itself, allows you to get twice as much potentially. Yeah, specifically inference.
- [27:16]
B
The reason why we care about this is to get to apples to apples comparisons.
- [27:20]
A
Yeah, exactly.
- [27:21]
B
What chip can do what?
- [27:22]
A
Yeah, exactly. So even just not accounting for all that stuff, because all that stuff actually would even further benefit the US because the chips are more reliable and actually have better internal capabilities. But just looking at the rating performance spec of the, of the chips themselves, you're going to move from basically a 5x gap right now to a 17x gap in two years. And not only that, but actually the chip that Huawei is going to release next year, the Ascend 950, has lower processing power and lower memory bandwidth than the best chip that they make today. So they're going backwards. So this is very unusual. So when we're talking saying, you know, Huawei is a significant threat to convenience Nvidia, the differential in what the companies themselves say the chips they're going to make are, is radically increasing, then the second thing is then Huawei says, okay, well, look, Huawei acknowledges we're not going to catch up to Nvidia in quality, but we're going to save this with quantity. We're going to make massive quantities of worse chips, and that's how we're going to compete. Because correctly, aggregate computing power is actually what matters. You can have very large numbers of mediocre chips or smaller numbers of very good chips, and as long as the total amount added together, all equal, is the same, you get relatively similar performance.
- [28:32]
B
With a huge cost, not just in the energy bill, but also in the programming complexity. You're making life much harder for your developers.
- [28:38]
A
Those costs are real. But for the sake of analysis, let's just assume, okay, let's give the Chinese the, the assumption that actually, that we can get perfect efficiency on kind of transmission and also on energy. I probably actually genuinely do believe that they're not going to be constrained. That's more of a constraint for us than for them. Their constraint is chips. Ours is some combination of chips permitting an energy. But anyway, anyway, so let's just take that at face value, now that we have the quality metric, we can actually evaluate AI compute much more easily. And you can say, let's just give the Chinese the assumption that they actually are making millions of chips, which actually even the public reporting says they're making hundreds of thousands. But even if they are making millions of chips next year, given how bad their chips are relative to Nvidia's, and frankly relative to any other AI company, the aggregate AI computing power they're making is going to be next year. Even under the most generous possible presumption assumptions, 4% of what Nvidia is making, and it's likely lower, it's likely more likely.
- [29:37]
B
Let me just put that in context. So I think I've seen numbers, maybe it was the IFP report that says something like 80% of all the AI computing capacity on planet Earth is in America or under the control of American companies right now. And then you're saying next year, just, just talking about next year's production. Huawei under the most optimistic assumptions. Basically like what they're telling, you know, people when they want to brag is that they will produce 4% the chips that Nvidia will produce. And by the way, if you haven't been paying attention, Google tpus are having a great year. I mean there's other chips in the story on Team America besides just Nvidia. So the idea that like Huawei is this sort of near term apocalypse loomy not borne out by the data you're using, which is Huawei's most optimistic data. I think that's a really remarkable finding.
- [30:32]
A
Yeah, exactly. And even, you know, there's been, I think debates kind of publicly and I think Jensen and others have been saying, well, they're going to make, you know, millions of chips. So just plug those numbers in. And you see even with relatively conservative estimates on Nvidia's production, I think, which, you know, Nvidia would know better than anyone, but the assumptions here were 50% growth relative to the baseline this year and last year they were at 100% growth. You get to 4%. If you are pushing the Huawei numbers down a bit and you're not giving them the most generous possible assumptions. And then particularly if you start to account for some of these other factors, all of which I just kind of granted to the Chinese here, you quickly get to 2%, 1%. So it's very low. Now look, obviously there's other compute in China, smuggled chips, stockpiled chips, other things like that. But if we're just talking about is Huawei a competitor? Can Huawei compete with Nvidia? The answer is very clearly no. And actually in the future it is getting harder for them to compete, not easier. Even if they radically scale up production quantity, they're not going to be able to make up the quality gap. It's impossible.
- [31:35]
B
Amazing. Now you said that according to the published performance specifications, Huawei is actually going to make a worse chip in the future than they currently, and I'll use quotes here than they currently make. What do you think accounts for that difference?
- [31:52]
A
Yeah, so that's pretty remarkable, right? Like chips don't get worse. There aren't, I can't, I can't think of a single instance where a leading company released a worse chip the next year. And that just doesn't, it doesn't happen. So yeah, the question is why? I mean, no matter what, what that says to me is that they're struggling with yield and process. And the only question is, is it because their yields on the 910Cs they made last year were so were so bad that they have had to make smaller chips and worse chips in order to have something that is even viable. That's option one or option two is actually they never made any 910Cs or 910Bs in China at all. And actually every single one was made at tsmc. We know that there was a huge violation with TSMC where they.
- [32:38]
B
Yeah, the reporting out there is that TSMC made 2.8 million 910b 910c dies for Huawei shell companies. So it's like, wow, there's millions of Huawei chips on the, the market. Yes, made by tsmc. So if you look at Huawei's chips on the market, you are not necessarily learning about Huawei's capabilities, you are learning about TSMC's capabilities. So the question is, what does Huawei's capabilities look like if they lose access to tsmc? We might be getting that answer in the form of this lousier chip that they're going to release next year.
- [33:13]
A
I think it is a clear indication that at the very least the vast, vast majority of those ascends were made at tsmc and it's very possible at Aldermore. And I think it is important to highlight, highlight how significant that violation was in the context of the US strategic posture. If you, if you look at the, the data of just generally like the, the, the, the AI computing power that China has produced over the last, in kind of 20, 22, 3 and 4 compared to the U.S. even though the export controls were in effect, actually the, the ratio was more or less the same, which is shocking. Right. That shouldn't be the case. And the reason for that was because China got a massive plus up from the TSMC virus violation. If you take the TSMC violation out of that data, Huawei or China's AI compute really starts to plummet in 23, 24 and then completely bottoms out in 25. But because of that, it basically artificially propped China up for two whole years. 2025 is the first year where China's AI compute is actually at a significantly lower ratio than the United States. So obviously that's hugely unfortunate. I think that that single violation probably delayed the impact of the export controls by two years themselves. But the good news is there have been moves to stop that. I think it is substantially less likely and very unlikely that we're seeing any violations at that scale occurring now. What that means is that the impact of the controls is ramping up more now it's exactly the time to keep our foot on the gas, not to take it off.
- [34:46]
B
They're really starting to bite because of that dang TSMC violation. I think there's one other thing which is not a story of a violation per se. It's just a story of an error, I'll say by the Biden administration and I hope you don't take offense, but HBM controls, right, Because TSMC makes the logic die which goes in the chip. But the. The whole chip module is sur. Is that logic die surrounded by a ton of hbm. And it was not illegal to sell advanced HBM into China. China Until December of 2024, I want to say. And Huawei was stockpiling that whole time. So in the same way Huawei got a massive boost from TSMC because of violating the export controls, they got a massive boost from Samsung, sk, Hynix and others. Not in violation of export controls, just in a gap in our export controls. And so like what does Huawei really look like? How competitive are these they. When we remove those various legs of the stool that has been propping them up, it's as you said, I think they should be. We should expect them to get weaker over the next few years, not stronger.
- [35:55]
A
That's right. And. And I assume we get the word.
- [35:58]
B
Chain, I should say.
- [35:58]
A
Yeah, that's true. And I think we're fortunate that also China's packaging capabilities are not nearly as good as those of tsmc. So yes, they, they received these dies, they received the hbm, but they probably lost a significant number of those chips in packaging them together than it rather than if they'd actually had them at tsmc. Now they. A significant number of ascends that. That is problematic. But it actually could have been more completely. You're completely right on the HBM controls. I think one big mistake the Biden administration made on that front was we identified this as a, as something that needs to be cut off back in early 2024. February, March. And there were internal consensus on doing that and we waited until December 2024.
- [36:39]
B
And Reuters reported on it in July, so.
- [36:42]
A
Exactly. So it was, it was, it was very frustrating. We were seeing. We knew this was fl. And look, I think that the Biden administration was generally trying not to be constantly firing the semiconductor export control gun at the Chinese and instead kind of get things into a package and then go with that. And there's pros and cons of that, obviously. Look, we see now that maintaining the kind of relationship with the Chinese and situating these kind of actions in the context of a relationship is important because it's important to kind of manage. You don't want that over the top Chinese retaliation like I think we saw in response to the very large terror threats, which then makes it much harder to take fewer actions or competitive actions. But at the same time, we had to move faster than we did because the Chinese were not waiting until once a year in order to decide how to circumvent our export controls. Right. They were doing that every single day and we were updating them once a year. That inequity resulted in some pretty big exploitations.
- [37:45]
B
Yeah, I mean, I feel like our schedule for updating export controls was based on diplomatic, political and bureaucratic convenience. And their schedule for circumventing export controls was immediately, at all costs, at any way we can possibly, you know, make progress. And that's just, that's a, that's a not good asymmetry to be working with.
- [38:07]
A
Okay, but the, the good news on hbm, I agree with that completely and I will, I will own that the good news on HBM is that it is controlled now that they, we have, you know, the Chinese have probably worked through their stockpile this year. I think even semianalysis says they think they're basically exhausted it. So any chips they make now going forward are going to have to use domestic HBM, which they don't make. And it's interesting because the Ascend 950, the one that's being released the early, early next year, does not include HBM, which, wow.
- [38:38]
B
I mean, HBM became standard in AI accelerators in 20, 20, 15, maybe, maybe even before that. It stinks if your chips are happening to be used in hbm.
- [38:50]
A
So there's a litany of production problems that they have from the logic node fabrication yield. I mentioned the packaging issues they have before and then also HBM constraints. And Morgan Stanley put out a report yesterday, I think that says that the Chinese aren't going to be able to produce hbm. And CXMT obviously is moving as fast as they can, but they have their own problems and the equipment regulations restrictions that they face. They're not going be able to produce HBM next year. So that means the. No Chinese accelerator next year is going to, is going to include hbm. We'll see when they can actually start making it. But then again, they have to ramp up production. They have a lot of constraints at making this. And the overall picture, when you look at it is the gap between our chips in quality and quantity, and their chips is growing. The gap in AI compute between the US And China is growing at an exponential rate, which is very, very bad for the Chinese, when also the amount of compute you need to do any kind of frontier AI training or inference is also increasing at an exponential rate. They can't afford to increase at a linear rate. Even if they increase, even if they scale their. Their production quantity at a linear rate and they don't get the commensurate improvements in quality that we're seeing and that they need to catch up, they will fall farther and farther behind. And we should do as much as possible to close gaps. And we should not underestimate the Chinese. I don't. I don't mean to say that that everything for hunky dory. No, like, yeah, they're smart, they work hard. Absolutely. And this is a top priority. But the situation right now is also better than a lot of people who have a lot of financial interest in tearing down the controls are painting it as amazing.
- [40:26]
B
Okay, so I've promised the audience that we're gonna use this Mythbusters format. We're now 40 minutes into this podcast. I'm legitimately loving this conversation, so hopefully you can stay on a bit longer than maybe I originally promised, and hopefully our audience can listen in a lot.
- [40:43]
A
There's a lot of myths to bust, so let's do it.
- [40:46]
B
Yeah. Okay, so let's start. This is a favorite of opponents of export controls, and I want you to rate it fact, fiction, somewhere in the middle, and then explain your reasoning. So here we go.
- [41:00]
A
Okay. And I haven't seen these in advance either. So this is off the cuff.
- [41:04]
B
Well, that's my fault. I meant to email it to you.
- [41:06]
A
No, it's fine.
- [41:07]
B
I like that. We're going to get the. Off the cuff.
- [41:09]
A
Yeah. Yeah.
- [41:11]
B
Okay, so forced indigenization.
- [41:13]
A
Sorry, it's fact, fiction, and what, somewhere in the middle? Okay, got it. Yeah.
- [41:18]
B
Forced indigenization. Export controls accelerate indigenization efforts, reducing China's reliance on the US and allies hardware and software. What do you think?
- [41:29]
A
Fiction. Now, look, the Chinese have a very serious indigenization program. Program. Partly because they don't want to be reliant on Western technology for anything. Maybe because of the theory of export controls, but this is a very abstract concept. The fact is, the Chinese determined in like, 2014, 2015, a long, long time ago, that with respect to advanced chips and advanced chip making equipment, they need to be at the frontier, and they need to do all of this in China. So the idea that anything that has happened between 2014, 2015 on the US policy side has moved the needle for the Chinese on that is fiction. When we came into the White House 2021, we were thinking about this and these were real debates that we had. And the conclusion we had was that the indigenization pedal is already on the floor. They're doing as much as they can because they see this as a critical national need and emergency. And what that also means is even if they maybe they throw another 10, 20, $50 billion at the problem, but the absorption rate of that is also going to be low or the marginal utility of that is going to be low because they've already spent so much they can keep dumping money at it, but they've already picked all the low hanging fruit and the median hanging fruit. Right. We're talking about super high hanging fruit here. And they know that this is hard. So the degree to which any policy moves that we make change things is very marginal also, I would say, especially now it became clear whether you agreed with the controls in both the first Trump administration controls on EUV and on SMIC and then the Biden administration's ramp up of this on equipment and on chips. Whether you agreed with that or not, the fact is that what the signal that sent to China was, the US is very serious about holding you back on advanced chips and they have internalized that and there's no putting that toothpaste back in the tube. There's nothing that we could possibly do now to convince them that actually we're doing a legitimate change and you don't need to actually work on this themselves. And that all this also is consistent with, you know, how the Chinese you every other technology, and there's never been a single technology where the Chinese government has been, has been, you know, comfortable relying on the west, particularly one of strategic importance. So I would say fiction.
- [43:49]
B
Okay, I love a lot of what you said there, but I am going to disagree and I'm going to say, I'm going to say somewhere in, in the middle. So here's, here's how I kind of tell the story. So the line again, as I read it, was export controls accelerate indigenization efforts, reducing China's reliance on the US and allies hardware and software. The reason why I say somewhere in the middle is I think you are 100% correct that China already had an indigenization program before we adopted any kind of strategy of export controls on strategic semiconductor technologies. That comes from, from the first government guidance plan, which, sorry, first government guidance fund, which I think was $100 billion in 2014 in China, comes from Made in China 2025, which set explicit goals for replacing shares of foreign technology and also said like, we need a local euv. Right. So it's clear that they had all of these goals. It's clear that they were spending crazy amounts of money on all of these goals. But I would say that actually it was the Trump export control controls on ZTE in 2018 that did meaningfully shift the conversation in China from a industrial policy priority with like a little bit of national security focus to an overwhelming national security priority, you know, against which, like, we need to mobilize the entire private sector. So I think like what it did is it escalated the issue from the industrial policy folks in China to the national security folks in China and also made the more commercially oriented private sector of China more willing to do this. So why do I say in the middle then? Because it can also go the other direction. Like when China was having its best indigenization work going, it was because there was no restrictions on other segments of the supply chain. Right. Huawei would have already indigenized AI chips if we had given them unrestricted access to Euro uv, unrestricted access to tsmc. So I think like a dumb export control policy can unambiguously accelerate a strategically astute and well targeted export control policy can slow indigenization efforts depending on like what market share and what choke points are and are not controlled. And so that's why I still love everything you said. But I would, I would sort of come down like in the middle.
- [46:15]
A
I agree. I think actually I, that is entirely consistent with what I said. Maybe the question that I answered was not in theory, do export controls have an impact on indigenization? To which I think your, your answer is exactly correct. Yeah, I think my response is the question of have any of the semiconductor related controls over the last eight years had any impact on, on China's semiconductor and digitization efforts? And I think the answer on that is no.
- [46:45]
B
Yeah, yeah. And then there's something that you said that I think is super important, which is there's nothing we can do now to turn it off. And this I think is almost like the most important point from what do we do going forward perspective.
- [46:56]
A
Well, eight years, five years, two years, like we're here where we are now, so it doesn't matter.
- [46:59]
B
Like, like China reads congressional testimony, China reads congressional debates. And what is crystal clear to China is that like the H200 reversal decision is super unpopular on a bipartisan basis in Congress. So are they going to look at and be like, oh, let's shut down our indigenization program because Donald Trump made One decision in 2025 that is absolutely going to be the guiding light for U.S. policy for the next decade? No. Right. That they have to. They're going to continue, as you said, pedal to the floor because the, maybe the next administration will reverse this, maybe Congress will reverse this. That alone is enough to justify the continued acceleration of the indigenous efforts. So for me, it's not about like, how do we apologize, how do we persuade them to stop, it's like, about how do we force them to fail by having the sort of correct architecture of export controls that makes it maximally expensive and complicated.
- [47:58]
A
Completely agree. And I think the legislative efforts on this side are really significant. There's. They. We keep seeing more bills come out from both sides of the aisle. I think in the last 24 hours alone, the, the H vac, the house foreign Affairs Committee ranking member and chairman both put out separate bills that would regulate AI chip sales to China and express significant concerns about each two hundreds. The one from the chairman in particular. Oh, the one from the ranking member, Representative Meeks would straight up ban the sales. The one from the, the, the chairman would basically. Representative Mast would basically say we should regulate GPU shipments to China like we do weapons to allied countries where basically Congress has to have the ability to provide oversight and potentially vote on licenses and vote them down. And it also requires the administration to submit a export control strategy to Congress and until that happens, bans the issuance of any licenses. And not only that goes a step further, that it revokes all existing licenses. So the fact that, that, that this is, this is the Republican chairman of the House Foreign Affairs Committee, that saying this really indicates that there's significant concern about the issuance of these licenses on the Republican side. And I think we're in a very unstable moment in this policy of where it's going to go. And then to the point that you're making, if you're the Chinese and you're seeing this, there's no way that you could ever kind of change your indigenization stance given how chaotic US Policy is on this and will continue to be.
- [49:34]
B
Yeah, very well put. Okay. The next myth is about damage to American companies and allied companies. So export controls hurt U.S. and allied companies with national security significance by increasing uncertainty, making them forego sources of revenue and diminishing their market share. And obviously I, I didn't Put it in this quote, but we're talking specifically about the semiconductors and semiconductor equipment part of the story. And this. So what do you think? Fact, fiction or somewhere in the middle?
- [50:06]
A
I think somewhere in the middle there's a few components of this that all of which are different and also different companies and industries are a little different. I think you could actually argue that well crafted export controls increase certainty. They, they say very clearly here are the rules of the road, you go forth after that. But like this is where you are not going to be allowed to say. And that was actually part of the idea behind the Biden administration's position on the ban AI chips. And then also the one the Trump administration put forward is basically saying here's a line, you cannot cross that. Now obviously that line has moved down several times and the reason for that was because of advancements in AI that enabled people to get equivalent levels of performance by kind of enabling technical workarounds. And candidly, it was also that we were learning as we were doing right. This was hard. We were regulating on the frontier. And the first control, we made a mistake. Interconnect bandwidth was a parameter that was in there and chips that didn't exceed it were all okay. And then we realized afterward that actually you could train AI models with chips with low interconnect bandwidth and therefore the control had a catastrophic technical problem. We had to fix that. So there are going to be. The changes that were made, I think are more to just adapt to the technical reality. But anyway, you know, I think they can increase certainty in some instances. Yes, of course they are going to impact the revenue of companies. On Nvidia's case, I think it has not radically because, or if anything at all since export controls went into effect. Nvidia's capex is up over, over 10x. They're selling all the chips that they're making. Obviously there's voracious demand globally, so I'm not terribly concerned about the impact there. But, but it's unfair to say that, oh, export controls never have impact. Of course they do. But the, the, the question is these are done for national security reasons. Right? And if the national security value of restricting the sale to China is such that it's worth having our companies forego revenue, then that's a decision that has to be made. And there are substitute effects that you have to account for. Right? I mean Nvidia, there's 100% substitution effects. But the equipment companies, if the equipment companies were completely blacked out the at China, they, they Undoubted, like let's say you said no SME to China. Yes, they undoubtedly would lose revenue. Would it be survivable? Probably would. Would it, would some of that then be absorbed from other countries? I think probably too. Right. I mean, every year the equipment companies come in and they say, well, this year we think China's going to buy less. May this is a year where we're going to see 20%, maybe in the teens percent from our revenue to China. And then at the end of the year, every year it's like last like three or four years. Every year they're like, ah, turned out the Chinese bought more than we thought actually of our revenues from China. And even though there are restrictions, China's still buying a lot of lower grade equipment. So I think what that shows is even if it was at a 20% level or even a lower level, the company's forecasts are fine and could survive. But look, we have to acknowledge that export controls can have an economic impact. These are serious tools with serious consequences. And you have to very carefully weigh the competitiveness of your company and your industry against the impact. If the United States were to impose controls that were to decimate its own SME industry or its own AI chip industry, that would be very bad. The thinking behind every single control that certainly we did and I think that the team has taken forward is that that is not going to occur with the way that they are crafted. I don't think there's been a single control that has been put forward on China that has resulted in massive financial losses for us AI chip companies or SME companies that has strategically impacted the United States. And that's good. And that's how we should design policy going forward. Where to draw that line exactly is the challenge. And I think we probably can err on the more aggressive side, particularly given the immense national security implications here.
- [54:09]
B
Yeah, I think you got it right. And I'll say that what you characterized accurately is the sort of general traditional trade off of export controls. I mean it's, it's revenues that, that would have been legal and now they're illegal. But now I'm going to go, I'm going to go farther than you and I'm going to say in the specific story of export controls on semiconductors and semiconductor manufacturing equipment that the United States has imposed over the past five, seven years or so, it's actually fiction. And here's the three arguments that I'll make why I'm a little bit willing to go go farther. The first is a Short term versus long term distinction, right? If Nvidia was going to sell, you know, $50 billion worth of chips to China next year and we say like now that is banned. Yes, right. Like they lose that money in the short term. But think about it from a long term perspective, China is committed to indigenizing as fast as they possibly can. So if you are Alibaba, Tencent, Baidu, you know, all these big cloud and AI hyperscalers in China, they need a bridge to be built for them, you know, for the next five years or maybe the next decade until Huawei doesn't stink, until the indigenization efforts start reaping fruit. So like in the global cloud and AI data center market share right now, Alibaba outside of China is pretty small. Tencent Baidu outside of China, they're pretty small. And if Nvidia sells them all these chips, well then they're actually going to gain market share a lot. That's why you saw Microsoft coming out in support of the Gain AI act, right. They were supporting chip export controls. And here's the thing. This year, Alibaba, Baidu, 10 cent, they want to buy Nvidia chips, but in 10 years, years they're going to be buying Huawei chips. Right? And so the point is, what share of the global market do you want Alibaba, Tencent and Baidu to have in 10 years? If Nvidia builds them this bridge, maybe they won't have 5% of the global market, maybe they'll have 30% of the global market. And so my point here is that Nvidia in the long term, they are destroying their own addressable market by helping Chinese companies grow, grow so that when those Chinese companies are grown up, they won't be buying Nvidia anymore. Whereas Microsoft, they're buying Nvidia now and in 10 years they're probably still going to be buying Nvidia. And that's why it's way better to help them grow than to help Alibaba, Baidu and Tencent grow.
- [56:50]
A
That's super. That's super, right? And important, right? That like it, the impact of a control on a particular market actually has positive impacts on the upstream parts of that supply chain chain. If you're going to constrain the Chinese, right? So like the fact that Huawei can't make AI chips is very good for every single US cloud provider. It's also good for Nvidia that doesn't face a competitor, right? The fact that the Chinese can't make the equipment is good for US equipment companies like there is a degree to which US controls can be used and should be used to maintain not just our lead but also the global competitiveness of our companies in the industries that we see as super important and strategic. But when you give the Chinese the capability to compete in those industries like we would in the global cloud market, if we send a bunch of AI chips to China, then that actually can have a negative impact on, on U.S. companies. So yes, sending AI chips to China obviously might be a positive from Nvidia's financial standpoint. Short term. Yeah, I mean I actually think it probably has almost no impact on their sales, but it allows them to diversify their customer base, which probably is good if, if for them. If you know, US cloud companies were to move more to Asics. But doesn't really, that's really, that's, that's.
- [58:12]
B
A really good point. So let's, let's double click on that. Nvidia. Since ChatGPT's release, every chip that TSMC can make for Nvidia, there is like 5 to 10 customers who would love to buy that chip. And so like losing access to Chinese customers, which you know, they didn't lose access because they could still sell H8 hundreds and H twenties, but let's just say it's like a total loss. The result is not a chip that doesn't get purchased. The result is there are fewer customers bidding up the price of that chip. And so to the extent that there has been any short term impact on Nvidia, at least during the period of GPU chip shortage, and keep in mind that like the head of Google's AI data center infrastructure structure, he said that seven and eight year old TPUs are still 100% utilized. He said that a few months ago. So the chip shortage is still with us in big ways. With the caveat that sometimes energy build out is more of the bottleneck than the chip shortage. But the point here is like to the extent that Nvidia has suffered, it's not because they haven't been able to sell every single chip. It's just that instead of like 10 people wanting to buy every single chip they have, have, it's not 15 people who want to buy every single chip that they can make. Okay, that's right. There's also the sort of changing composition of demand which I think in the equipment part of the story really eliminates almost any financial impact that the equipment companies have suffered from these export controls. And don't take my word for it, here comes a quote from the chief financial officer of as in an investor relations call from July 2024, quote, the way we look at demand for our tools is not from a specific geography. In this case, China we look at and that's the way we model our sales medium and long term. We look at what is the global demand for wafers and whether those wafers are being produced in country X or country Y. At the end of the day it doesn't, doesn't matter. And I think that is very important to recognize in looking at your model. We don't have a specific China element in our models. It is the global demand for wafers that drives our modeling. And we know that's true because when those export controls started hurting companies like sk, Hynix, TSMC and Samsung for their production facilities in China, they were planning on building more facilities in China. So did they cancel building those facilities? No, they canceled building them in China. So the same amount of equipment was still sold. It just went to a fab that was built in Taiwan or Korea or America or Japan. It didn't go into a fab that was built in China because is the ultimate demand for wafers that drives demand for equipment. And then in the Chinese case, there's another changing the composition of demand, which is not just geographic, but it's the one that you mentioned which is sophistication. So Chinese purchases of EUV machines, machines went to zero, but Chinese purchases of duv machines went through the roof. And so ASML is still making all of that revenue from a Chinese customer base. They're still making all that revenue from the advanced chip segment of the market. It's just that less fewer chips are being made in China and the chips that are being made in China are being made with less advanced equipment, which is probably more profitable on a per machine basis for asml. So that's why I come down and sort of of say in this specific instance, I think, I think you can call it fiction. There's one final argument which folks listening to this podcast have heard me make before. But go look at other industries where there were no export controls. Look at Tesla, which literally like, like in the, in that Reuter story you heard them saying, oh my gosh, they've got former ASML employees working on their secret EUV project. Well, the New York Times reported earlier this year that almost every single single executive in Tesla China who was there for the first two years is gone and now works at Chinese competitors to Tesla and Tesla. According to the Global Times, which was bragging about it when they make their, their cars in China. 95% of that supply chain is Chinese companies that Tesla helped grow to maturity. And now that's the same supply chain that is feeding all these Chinese competitors that are really ruining Tesla's market share and fortunes in China. So my point is, if you look at the specifics of like, what are the industries where we didn't have export controls that are high tech, it's going way worse than the industries where we do have export controls. And then there's this changing composition, short term, long term phenomenon.
- [62:47]
A
That's right. And look, I'd say, I'd go even farther to say there's basically two industries where we have had pretty significant successes holding back Chinese indigenization efforts. One is advanced chip manufacturing and the other is jet engines. And both of those are industries where we have imposed export controls. The other industries where we don't impose export controls, we have had a much harder time. Now, obviously you have to be judicious on where you do you impose export controls and you have to make sure you do it in a place that's logical and actually hard to indigenize. But there is not an instance of an industry where we've imposed substantial export controls. And then it's like hugely, hugely backfired on this, at least not in recent times.
- [63:27]
B
Okay, our next, our next factor fiction comes from Jensen Huang, who said this actually at a recent event at CSIS where he was effectively saying like, chip smuggling is a myth because the servers that these are built into weigh 2 tons. And how are you going to sneak something that weighs 2 tons across the border? So he's effectively saying, you know, he doesn't believe any of the reports of chips smuggling and he thinks overall it's a myth and that Nvidia understands where its chips are going. So fact or fiction? AI chip smuggling to China does not happen.
- [64:03]
A
Well, that's clearly fiction. I mean, the idea that there are no chips that are being smuggled to China is clearly false. I mean, the Department of Justice, the same day that the administration announced they were sending H200 chips to China, busted an AI chip smuggling ring and went talked at length about how important these chips are to US national security and how damaging it would be if they go to China. And that's just in the United States now. Clearly there are efforts in other countries. I mean, Malaysia is like the number two importer of AI chips in the world behind the United States. That's silly, right? Obviously Malaysia is not the number two AI power in the world. So therefore Something must be going on there now. And we have no ability to kind of track these chips once they go somewhere. There are certain countries to which we have export restrictions or license requirements even, but there are many to which we don't. You know, yes, some of these racks are big, but the chips are small. Obvious. You look at the markets in China, you know, there are. There is chip smuggling in China. Now, there is a question on this of how significant is the impact of that. And it's something that obviously we should be concerned about. But the idea of the controls is that this is really a quantity game. I don't care if someone has one H100 or an H200 or a GB300 or even if they had an early Rubin, it doesn't matter. Right. You need large numbers of these chips to aggregate the control computing power. If the smuggling numbers are sufficiently low that it's, you know, they're getting 100,000 chips a year even, which is. I think there's a CNAS report that estimated their median was 140,000 chips a year. You know, particularly as our production numbers go up and maybe their smuggling doesn't scale at the same rate. Like that becomes. It's acceptable. Right. Like we can live in a world where we have very big compute advantages. They're smuggling some chips, but it's not half the chips. The other concern, though, that this doesn't address that actually, I think is the bigger issue is compute smuggling. And what I mean by that is I think what the Chinese are doing more now, and I think they're, you know, there's all sorts of indications in the public of this is rather than the chips go to Malaysia and then smuggle racks across the border. I do think that's happening. To say that it's not happening is a little bit silly given all the reports of it. But the thing that takes easier is to just create a shell company, have that, if you're, you know, Huawei or Alibaba or whatever, have that shell company make a data center in Malaysia that uses all the chips. We have no insight into what's into where chips are going in Malaysia or Thailand or, you know, the vast majority of, of of countries in the world. And then that data center is then used to train Chinese models, perform inference on Chinese models. And you don't need to ship the chips to China because all you do is you ship the data to China. You're just going to send the files over. There's a Wal Journal article a few months ago about how actually they're just smuggling hard drives back and forth instead of the chips. So compute smuggling is a really big problem, I think. And because there's no restrictions on China's remote access of chips, I mean, actually China could access, they have unlimited access to US Cloud. China could access as many H2 hundreds or GB3 hundreds as they want on AWS or Microsoft Azure and, and not be in violation of U.S. law. So the amount of computing power that China theoretically has access to is unlimited. But also there is a lot of data centers. At least US Companies are subject to US Law and are going to be responsive to US Law. If we were to say, hey, don't help the Chinese or something, if we were to make that policy change in the future. But these cutout data centers in Malaysia and Thailand and Indonesia and especially Southeast Asia, but it could be anywhere. I think the compute smuggling from them is a massive, massive problem. And that is where I would say China is getting the vast majority of their compute from these. Even if you look at the Chinese production numbers and you integrate smuggling and stockpiling and you kind of get a total number, it doesn't seem like it lines up with Chinese demand. If you assume Chinese demand is kind of roughly US Demand, maybe a little bit less, it's probably only a quarter to a third of what their demand should be, which means that they might be. This is a little rough, but the remainder of that they must be getting from remote access. So there is an enormous amount of compute smuggling going on. And if we're serious about constraining China's access to compute, that's something that we really have to get our hands around both via US Cloud and also foreign cloud.
- [68:26]
B
Yeah, I like your point, but I'm not sure I like your terminology. Just because, like, that's fine. I feel like compute to me screams sort of the hardware side of the smuggling of the compute thing. What you're talking about, about is the results of the compute. I don't know. Yeah, it's okay.
- [68:45]
A
I'm open to new, I'm open to new terms on it.
- [68:47]
B
Yeah, well, you know, just workshop it for a minute.
- [68:49]
A
But.
- [68:50]
B
So let me ask though, you know, if you were king for a day, what we have tried to be doing for a long time now is having these restrictions on selling the chips, the chip making equipment, the data center stuff, et cetera, into China. But as you point pointed out, you know, that's, that's restricting China's ability to buy and own on Chinese soil, but there's still options available to them. To rent abroad. And so if you were king for a day, what would you do?
- [69:22]
A
So the way I think about it is if, if we, if, you know, if I was staying for a day, I would, I would say COMPUTE is, is one of the most important strategic resources on earth. It is only increasing importance. And our strategy has to be to collect as much US dominance in COMPUTE as possible, US and allied, which is very consistent with the Trump administration's AI strategy actually, and also with the Biden administration's general policy approach. Maintain as large of a lead as possible. I think the way you get at that is you constrain China from making chips, buying chips and renting chips. Make, buy, rent, and then what are the things you have to do in each space space? I mean the making chips, it's completely closed loopholes in semiconductor manufacturing equipment controls. It's equalized with allies. But also I think you probably want to cut off 300 millimeter tools is what I said in my testimony. Just be clear with the fdpr, if allies won't get there, just be clear that anything kind of in the moderately advanced category is going to get swept in and make sure that we're not going to have the, these kind of, you know, whack a mole games that, that we've been playing, given the importance of that. And also cut off servicing to make sure that the maintenance of existing machines, US and allied machines, can happen. In the buy category, it's prevent China from buying our chips, both directly from us, obviously, but then also prevent them from buying them via cutouts abroad. So make sure you have idea where chips are going. Right. Like chips should be, you know, especially large quantities of chips. Like someone buys one chip, I don't really care again, but like large ships, shipments of chips, we should have an idea of where they're going, particularly to countries that we don't trust. Maybe there's various ways you can modulate this policy. Maybe we don't care if you're generally shipping to the UK but if you're shipping to, I don't know, Rwanda or something, then the US government might want to have information about that shipment before it goes. So generally having insight there and then rent, the last category is kind of, I think there should be restrictions on access to ships virtually that generally mirror those that we have access to them physically. I mean, it doesn't make sense that right now the Chinese can not only just rent access from the cloud, but actually they could probably even install their racks into someone else's data center and then say Technically we're not owning and operating these chips, this other company is. But it's all constructed pursuant to exactly how Alibaba creates a data center. So it's effectively an Alibaba data center. This is silly. I think we need to.
- [71:55]
B
Correct me if I'm wrong. I think the only substantive action that had been taken against this was the 50% ownership subsidiaries rule, which would have, which would have roped in a lot more restrictions for a lot more Chinese companies, but that has now even been reversed. So to me it seems like the restrictions on renting chips is a real gap in the overall architecture.
- [72:16]
A
Yeah, so I generally. That's right. The, the Biden administration's AI diffusion rule, which by no means perfect, but one element that it had was a restriction on using chip to train models over a certain capability. So that may or may not be the best.
- [72:32]
B
You know who you were doing that for.
- [72:33]
A
But it is the start of saying, yeah, this is, there is, we're now putting end use restrictions on the use of US chips anywhere. And what you have to do to make that effective, like an end use control, it's only as effective as the person who is owning and operating it. So the, I think what generally, you know, this administration, any administration should do is try to push as much computing power as possible into trusted providers. People that you think will follow, follow US regulations and then impose regulations that restrict the end use of those chips to support China. If you have that framework, I think you can get at the rent problem at least to like massively cut down. Obviously you will probably still see some, some kind of smuggling, but it will be much more on the margins rather than supporting like you know, 50, 75% of China's AI computing power, which is where we are now.
- [73:19]
B
Yeah. So I, I gotta, coming back to the, to the smuggling factor. Fiction. Totally agree fiction. And, and, and I'm going to go even farther. I'm just gonna say like, you know, we talked about fact fiction somewhere in the middle. It's like embarrassingly fiction. I would be embarrassed to have my name associated with the argument that like chip smuggling does not occur. And it's not just because that like New York Times reporters, Wall Street Journal reporters, the Information reporters, Financial Times reporters, Reuters reporters, voters have all like interviewed smugglers, looked at facilities, seen receipts for purchases of hundreds of millions of dollars worth of chips that have been smuggled into China. And in the case of the information reporting really got sophisticated descriptions of the tactics being used and how they avoid investigations. It's also that like the the, the, the chief of the Bureau of Industry and Security testified under a oath before Congress saying that smuggling was a fact that they knew about it. And then more recently, as you said, there was a Department of justice investigation, right, where like they have like wiretaps, they're like intercepting shipments. The fact that anybody at this day would claim that like chip smuggling is not a serious problem where we need to clamp down on enforcement of all of this, I would just be embarrassed professionally to be associated with, with that argument in any way. Okay, let's go to our next fact or fiction. China's progress in AI models and especially open source models proves that the restrictions are not holding Chinese companies back. And here like think about like the deep seek debate we were having back in January. So fact, fiction, somewhere in the middle.
- [75:10]
A
I mean, I think, I'd say fiction. I think the causality there is, is too, too definitive. And I think Chinese companies have been pretty clear that they are holding them back. I mean Deep seat CEO has said explicitly the biggest thing holding us back is the ban on AI chips. The Chinese government leadership, the premier has said the same thing. So it is holding them back. I think, you know, we can have valid debates about the degree and what works and what doesn't. And I think it probably is, I think it is significant per their own words. The whole thinking here also is that it is going to get more significant as long as we don't do anything. As long as we don't, you know, start to, to move to a sliding scale approach and start shipping chips to China. The impact of these controls is going to ramp up because the amount of compute that's necessary every year is going up exponentially. And China's chip production at best is going up linearly. So that gap is widening. And there are various things I think over the last couple years that have caused the controls to have less initial impact. Some of them will was Nvidia was able to sell like a 800-8-800s to China that actually were almost as effective. Then you had the TSMC violation that significantly bust up, bumped up China's capacity. But we're building this regime and there is a general consensus around the importance of strategic computing power. And as we close off these loopholes, the next one becomes a little bit harder to exploit in a massive way. And I think that, that the impact in the past has been real. I think if we hadn't had these controls, China probably would be at the frontier. I think they probably would be making data centers around the world Absolutely. There'd be nothing to constrain them and it would be difficult in every other area they have an advantage. And I also think that the impact of the controls in the future is going to be greater than it has been in the past. Just because if for no other reason than the increasing rise in computing power that's necessary, but also that we have, have closed loopholes over the last couple years and it's getting increasingly harder for, for the Chinese to kind of maneuver in the space, the limited space that we're, that they have.
- [77:20]
B
Yeah, I totally agree. And I think that this is an example of the opponents of export controls moving the goalposts in a way that was not what the, either the first Trump administration or the Biden administration or the second Trump administration set out out to do. Right? So it's like we are, we are going to put all these obstacles in China's way. But yes, they're still making forward progress or actually in the case of the Huawei chips, as you pointed out, they're actually making backwards progress, but they're still making forwards progress. In AI models, you know, Deep Seek is still doing interesting things. But like imagine if, you know, we're actually running a three foot race, right, and we have an American racer and we have a Chinese racer and I make the Chinese racer run with a hundred pound weight vest on and instead of having like Nike super shoes, you know, they have to run barefoot. Okay, well they're still able to walk forward. Does that mean that the controls don't work? Like. No, like the, the, the, the Biden administration, the first Trump administration administration, that they did not promise that like China was going to be sent to the Stone Age and be forced to live there for a thousand years. They said that these were useful measures to take in the race that we're having with China. It's going to slow China's progress. And, and there the key question to ask, to determine whether this is fact or fiction is not the, the moving goalpost, which is is China still making forward progress? Progress? The, the key question to ask is where would they have been in the absence of these controls? It's the counterfactual question. And you said it right, like it is super plausible, perhaps even likely that China would already be ahead of us in AI had we not had these controls. And I hear that from AI researchers at the top labs who read these Deep Seq papers, they're like, my God, where would these guys be if they had? If they were, you know, building a gigawatt a Month the way China has demonstrated the ability to do. Okay.
- [79:30]
A
Yeah.
- [79:31]
B
So Chris, I've just, you know, spent my four myths, but I wanted to give you the opportunity to, you know, debate your favorite arguments on this topic. So are there any myths out there that you want to talk about?
- [79:43]
A
Yeah, the one that I think I'd flag that people are also throwing out is that our use of our chips is sticky and they could get addicted to them. And that if we get China addicted to our technology, technology that therefore it'll be really hard for them to move off of it. And this actually will disincentivize the indigenization that we talked about before and also kind of accrue us strategic benefits in the future. I think that's one of the other big arguments that is being put forward on why we should export chips in addition to kind of Huawei can do it, which I tried to dispel in the report. But yeah, Greg, curious what you think about that.
- [80:20]
B
Okay, so this is great. I should have included this. And I'm glad that you included this. And I'm going to say if it's a future facing argument, I think it's fiction. Right. It is the case that Nvidia has ecosystem effects and so that the more customers use Nvidia technology, especially Cuda, where the customers are sort of co investing alongside Nvidia in making Cuda more valuable as the they use it, it is definitely the case that that is sort of a, you know, to use the preferred term, an addictive kind of ecosystem. And that is why Nvidia is very difficult to displace as top dog in this market. If you are AMD or Intel and you are competing against Nvidia in America or Europe or Latin America or Asia outside of China. The reason why I call this argument fiction is that China, China is just different. China is just different. They are willing to do things that Europe, America and others are not. Like if America said everybody has to switch from Nvidia to AMD within the next three years, like what would the Cuda moat look like? It would look shot full of holes within three years. And so the point here is that what we can do for China by selling them a lot of Nvidia chips today is building them a bridge to a Huawei future. But we already know that they are unwilling to be addicted to Nvidia chips because they told us in 2020 with document number eight and with the three, five, two strategy. That was an era in which most Chinese private sector companies were already addicted to Nvidia in this sort of like ecosystem entrapment kind of phenomenon. But Chinese policy was everybody has to get off Nvidia and three years and we are willing to spend unlimited amounts of money to, to do what it takes to do it. And they were. That was looking pretty good for China in 2020 because Huawei was fabbing 7 nanometer chips at TSMC even before Nvidia was. So the reason why I call it fiction is that as as we talked about earlier, China's preferences, China's goals, China's strategy, China's sort of limitless willingness to spend, all of that is already set in stone. And whereas most countries, most regions, most customers are not willing to do what it takes to get unaddicted to Nvidia, China unambiguously is. So in this specific instance, it's fiction.
- [82:55]
A
I would generally agree. I would make a one caveat though that I don't think it's just China that's this way. I agree with all of your analysis. I think the thing that that differentiates China is it would potentially have a very high willingness to do this and also would have a lot of capital and resources behind it, which most actors do not, but some other actors do, namely USAI Labs. Right. Anthropic has actually also moved away from Cuda and Nvidia chips. Google similarly, OpenAI just announced a deal with Amazon to use Trainium chips. So these are also. Obviously these companies are in no way similar to the Chinese government in basically every way, but they do have a lot of resources, a lot of capacity and are able to break through the moat by throwing that at the problem. I think certainly the Chinese government would throw unlimited resources and unlimited capacity at the problem as well, which I think would render the moat relatively ineffective, particularly compared to the hardware moat. Where can they actually get the physical supply and. And that you can't really solve with resourcing the same way.
- [83:59]
B
Yeah, no, your argument is better than mine. So you can persuaded me. Yeah, we have this precedent. Google has successfully moved out of the Nvidia ecosystem even after being super dependent upon the Nvidia ecosystem for most of its early AI journey. And they have less money and less resources than China. So it is clear, right, that China could replicate Google's achievement even though we think it would be hard. It was hard for Google, right?
- [84:29]
A
That's right.
- [84:31]
B
Okay, that was a really fun Mythbusters part of the show. Now we're going to go to what is just my last question, which I'm super interested to hear from you. Because you have exposure to so much of this story over so many years of it being a thing. So what is the single biggest mistake we could make on export control policy and what is the single best policy that we should have that we don't? So sort of the, the biggest mistake we could make and what is the biggest opportunity we're currently missing?
- [85:01]
A
Yeah, I think the biggest mistake we could make is going back to a sliding scale approach. The, the innovation of the kind of Biden administration's approach on controls, especially on, on computing, was moving from an absolute approach really, to a relative approach. Now, keep in mind, we have always had controls on like supercomputers, right? Like, we've never allowed, you know, the best supercomputers or something go to the Russians or anything like this. The idea of controlling large amounts of computing power is, is not, not without precedent. But the Biden administration said obviously now that it's particularly commercial items, dual use is we're going to draw a line in the sand and that line is never going to move. And the idea is, because if you win, if you had a slide sliding scale approach there, and you said, okay, you can have chips that are two years behind or two generations behind or something, China or Russia or others could aggregate very large numbers of those chips and achieve similar effects as if they had, you know, slightly smaller numbers of the most advanced chips. So the, the idea of the sliding scale approach doesn't work in this, in this aggregation age, particularly given the challenges and the dynamic dynamics that we face with respect to AI compute. Moving to an idea where, okay, we're going to keep our best chips ourselves, but we'll allow China to have basically as many of the second or third best chips that they want, will dramatically decrease the US Advantages in compute and have significant negative consequences. I think we have to maintain the line in the sand and from that the advantages will grow over time. We go back to the sliding scale approach and actually, I think we barely see any benefits, benefit at all on the things that we should have and we don't. I think, look, kind of as I went through before, there's the make, buy and rent elements and there's things we should do in each of those buckets that we've talked through. I mean, of the two of them, I think that something on remote access is really big. The fact that there's nothing there is a big problem and there's multiple elements of this as well, I almost don't want to. It's hard to choose between those three buckets. But just given that we have, we have some equipment controls, we have some chips controls, those loopholes have to be tight, they have to be closed. But on remote access, we have nothing. It's completely unregulated. China can just use, you know, US cloud to power anything and everything they want with basically no restrictions. Unless it's like, literally for like WMD and users or something. That seems like a pretty, pretty egregious gap that I think, I think we really need to. Need to close.
- [87:27]
B
Makes total sense to me. Okay. Chris, I've been wanting to do this for years. I'm so glad we finally got to do this. You bring a ton of expertise and a real ability to explain things in ways, as you said, you know, has been sort of the, the, the magic of your career, you know, bringing bridging these two communities together and making them able to talk to each other, even though both of their domains are quite complex. So I appreciate you sharing and with the audience. Thanks for coming on the AI Policy podcast.
- [87:55]
A
Thanks for having me, Greg. This is great. Let's do it again sometime.
- [87:58]
B
All right, take care, YouTube. Thanks for listening to this episode of the AI Policy podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word.
- [88:18]
A
Word.
- [88:19]
B
This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mann. See you next.