How China Got 340 DUVi Lithography Machines and Why We Should Stop Selling Them with Saif Khan and Nick Brown

0:10 Welcome back to the AI Power Podcast. I'm Greg Allen, and today we've got an episode that I'm quite excited about. We've got Nick Brown, a senior fellow at the Center for Technology and Statecraft, along with Safe Khan, the director of the Center for Technology and Statecraft. Now, this is a relatively new institution, but very quickly after its formation, put out a paper that I think is a, if not a must-read, then at least a must-executive summary read.

0:39 This is more than 100 pages long, and it is a super deep dive into all the different elements of immersion DUV lithography equipment, its impact on China's ability to make progress in AI and in the semiconductor industry, and ultimately comes down with some very heavy-handed recommendations, but I think they're very well justified.

1:02 So, Nick, safe. Thanks so much for coming on the AI Power Podcast. Yeah, happy to be here. Thanks, Greg. So before we get into the meat of this paper, I want to get into you two. And I want to understand a little bit about your career arc and how you came to be interested in semiconductors, semiconductor manufacturing equipment, and ultimately, like what put you in a position to put together a paper like this.

1:25 Safe, you're also the founding director of this new think tank. And so maybe you could also talk a little bit about the center after you talk about yourself. But start with yourself, Safe.

1:34 Great. Yeah. So actually, I initially had a first career before coming into policy as an intellectual property and tech lawyer for almost a decade. And so it wasn't until later around 2019 that I entered the policy world. So for the last seven years, I've been working on AI policy, semiconductor policy.

1:51 The first job that I took in DC was as a research fellow at the Center for Security and Emerging Technology, which at that time was the first dedicated AI and national security think tank. You know, it's funny looking back on one element, which is that my very first report that I published in 2019 was actually on DUV immersion lithography.

2:09 And so in some ways it's kind of full circle to come back and now with Nick kind of work on this issue again. And obviously, lots has happened on the topic in the last few years, but I think it just remains one of the critical key choke points that is relevant to the long-term UX mana balance.

2:23 So I think it's worth just dwelling on something you said because you're a humble, reasonable fellow. And so I'm gonna have to brag here on your behalf. If you didn't hear that, national security think tank writing about immersion DUP lithography in 2017, way before it was cool as a topic to be interested in or obsessed with.

2:42 So credit to you for not just coming full circle and being a part of this report, but also anticipating so many of the different elements of AI and semiconductor policy that we're wrestling with right now. All right, sorry, I interrupted. Please continue, say.

2:55 Yeah. So after a couple of years at CSET, and I think that was really kind of a formative time in that, you know, we were able to, before AI policy was in the headlines every day, do some deep research that I think like kind of set the ground for policy thinking that was going to be coming over the next few years.

3:09 And so for me personally, that was doing a lot of work on semiconductor supply chain policy and the like. And then later served in the executive branch for four years, first at the National Security Council, covering semiconductor policy, export controls, the CHIPS Act, and then later as the Commerce Secretary's technology advisor, covering all aspects of AI policy.

3:28 And I think having that think tank experience for a couple of years before that was, I think, really helpful to develop a worldview and kind of develop ideas in advance because when you're in government, things move too quickly. Now, outside of government, as of last year, it's been a year at the Institute for Progress as a fellow where I'm still affiliated and CTS, our new institution, is actually an affiliate of IFP.

3:47 And then now more recently, we've started this center. And I think one of the key motivations for the center is to kind of, I guess, replicate that experience of creating a place where you can do highly technical and forward-looking research that when there's a policy window for it, we'll have all of the analysis completed while at the same time creating an environment where, you know, researchers can really become an expert on the topic and be seen as the person to turn to.

4:10 And I think in some sense, it's also then a way to do talent development as well, to really develop experts in the field. So I think that's kind of the core motivation of ETS. And so we have got a great team so far of six, including both of us. But yeah, I'm excited about the work we'll be doing over the next couple of years.

4:25 Very cool.

4:26 And now I want to turn to you, Nick. What's your story? Yeah, I guess um I'll sort of copy safe, insofar as I also had a prior career before policy. I was in software for a number of years. I was a software engineer, worked at Google for a number of years. Actually, both of us, I think, shared that we were both physics majors uh in undergrad.

4:44 So there's something about having worked through uh physics problems that brings people together, I guess. I got into policy in 2023. I shifted down to DC and worked at Rand doing uh all this technical analysis of semiconductors and semiconductor apply chains. What what uh what time frame are we talking here when you're at Rand?

5:01 Uh this was starting in 2023. So it was a considerably less sleepy era, although potentially still sleepy compared to the current era. But then after that, I also served in government. I was in BIS, which is of course the part of the commerce department that works on export controls and actually um develops them and enacts them.

5:18 So, you know, a lot of times uh, you know, the high-level policy ideas then get translated into specifics. So I was at BIS for a year, and then since being out, I've also, you know, with SAFE and our third co-founder, Constantine, started up CTS, and we're um small but mighty, I think, and growing.

5:35 Very cool. Okay, so now let's get into this paper. You've published a paper. Its title is DUV Immersion Lithography: The Foundation of the USAI Advantage, Closing Gaps in Export Controls to China, the rare double subtitle kind of a paper. So now I think in the common parlance of policy discussions about semiconductor equipment, 90% of the oxygen is usually taken up by ASML's EUV machines.

6:10 You're here talking about DUV and immersion DUV machines. So explain specifically what is one of these machines? What does it look like in the global supply chain? And then why are you saying that it is the foundation of US AI advantage when what we hear so much about is EUV machines?

6:32 Yeah, I'm happy to take that one. So, yes, EUV is they're the most advanced kind of lithography. Lithography are just all the machines that are responsible for defining the patterns that become circuits on a chip. EUV is the most advanced of this technology and arguably the most advanced technology on the planet, although potentially our frontier AI models in their own way might give it a run for their money.

6:56 Um they cost hundreds of millions of dollars and they do get all of the press. DUV is one generation back. So EUV stands for extreme ultraviolet, whereas DUV is merely deep ultraviolet. Um, and within DUV, there is what's called DUV immersion, and then what we call dry DUV. And you can think of it as DUV immersion, it uses water, and so it immerses a part of the projection of the light in water, and that affects the properties, whereas dry DUV uses air.

7:26 So that's one easy way to make that distinction. DUV immersion uh systems are not as powerful as UV systems. They can't make you know the very most advanced nodes that are at the frontier at TSMC, but they are still extraordinarily powerful pieces of technology. They are certainly capable of making AI chips, and critically, they are the systems that are responsible for China's AI chip production.

7:52 That is sort of the frontier of where China is at, and that's an artifact of our exports into China, and we can get into all of that. So, you know, I, you know, when we say that they are the foundation of the US AI advantage, there's lots of ways to think about that. But maybe a starting point is that that is where, at a technical level, the line is where those are the machines that China can get access to, and those are dictating how many chips that China can make, and therefore the relative gap between what the US can produce and what China can produce.

8:22 Got it. And then SAFE, I know that you've been following this sector for a long time. You've been involved in export control policy for a long time. So can you kind of walk us through the stages of how lithography machines have intersected with American export controls to China and take us way back, right?

8:42 Like think about what was this like 20 years ago, what was this like 10 years ago, and then what were the major milestones to get us where we are today in terms of what export controls on lithography machines look like to China.

8:56 Sure. So the concept of export controls on semiconductor manufacturing equipment actually goes decades back. You can find US government documents from decades back where it was even used as a tool in the Cold War where the Soviet Union was building semiconductor manufacturing capacity, and the US government at that time, with allies, was imposing controls that would limit Soviet production of advanced chips.

9:17 Oh, yeah. I've read some of the old declassified CAA papers about them trying to stop Soviet smuggling of Western semiconductor manufactured equipment. Surprisingly readable and kind of a remarkable echo of our era today. But keep going.

9:33 Yeah, absolutely. And then even on China, it's been kind of a long-standing piece of export control policy. Um, you can go back to a report that I think was published around 2022 by the Government Accountability Office. It's an office that's affiliated with Congress, where they talked about the history and the then policy of semiconductor equipment controls to China.

9:52 At that time, there was a generational policy where the US government and allies in multilateral frameworks would kind of deny access to the most advanced or maybe the two most advanced generations of semiconductor nodes. That is like so basically keep China one or two nodes behind the frontier of semiconductor manufacturing.

10:10 And that was a policy that, as far as I can tell, was maintained for uh some period of time, where as the leading edge advanced, then you would apply controls just to those most advanced generations and move the threshold up over time. And so that's been uh true for a long time. Lithography has been a core aspect of semiconductor manufacturing since basically its inception.

10:30 And so lithography controls have been a core part of this. If you go back decades, it was a more distributed market where you had a number of US companies that were also making the tools. Uh, but over time, as we've seen in many parts of the semiconductor industry, you've seen concentration due to fixed costs and more RD to develop the next frontier.

10:48 So you've had many companies drop out, and now we basically see something like a near monopoly from ASML. Certainly a monopoly at EUV, but then even for DUV immersion, uh ASML, the one company in the Netherlands, has something like 99% global market share, with Nikon, the other only commercial producer.

11:04 And of course, China is attempting to indigenize, but in our assessment has not done so quite yet. And so that's been a core aspect of export controls. You fast forward to the Trump administration, and there was a debate in the context of all of this policy about uh whether EUV, which was subject to export controls based on all of the past kind of work on keeping license requirements on the most advanced lithography tools.

11:27 There was a license requirement in place. There was a question of whether the Dutch government was going to grant a license. Ultimately, I think this was around 2019, uh around December that year, where the Dutch government, in negotiations with the Trump administration, ultimately decided not to approve a license for EUV tools to China.

11:43 And that would have been the first EUV tool that went to China. And now that policy is held since then. And as far as we know, no EUV tools have ever shipped to China. And that has limited their ability to make some of the most advanced chips at five-nanometer technology and better. And then in the Biden administration around 2023, there was a series of controls for the US starting in 2022, but then allies, Netherlands and Japan, instituted additional controls on DUV immersion lithography.

12:10 And this, if you look at data that was released from the Select Committee on China last year in a report, shows that the very most advanced DUV immersion tools, and I think this is maybe the four most advanced generations of those tools currently, are actually not now shipping into China at all. So there's controls, and licenses per that data appear to not be approved.

12:30 That said, the same data showed that for less advanced tools, which are also subject to license requirements, many of them are allowed to be shipped into China, despite enabling seven nanometer capability that is essentially used by Huawei to make AI chips in China. So yeah, there's been a lot of work over the last few years in export controlling lithography.

12:49 This is not a new issue. The key issue is that the job is not done and there are still gaps, and China's able to scale production because of those gaps.

12:56 So you were in the Biden White House and participating in some of these decisions. And I'm I'm certain that you can't say everything that you know about how these decisions went down. But let me ask a couple questions. The first is the counterfactual question. You know, we've had these EUV license denials, the sort of de facto ban on exports of EU V lithography machines to China for gosh, seven years now, I think is the right time window.

13:27 Where would China be in semiconductor manufacturing if we had never done that?

13:34 Yeah, and this is a point I think, Greg, that you've made, but just to repeat it, you know, you look back at 2019 or so, this was when Huawei began to actually enter the global top 10 of chip designers in the world, and which was unprecedented. They were quickly growing as one of the leading chip design companies in the world.

13:51 They had released an AI accelerator chip called the Huawei Ascend 910, which, measured by raw performance, was actually more powerful than the then leading edge NVIDIA chip, the NVIDIA A100. Obviously, there's been questions about the maturity of Huawei software stack. On net, I would say probably at that time, the NVIDIA chip was still more useful, or AI because of the maturity of the software stack.

14:12 But basically, Huawei was gaining. If we had not done any of these export controls, say, control EUV would have limited China's indigenous ability to manufacture the most leading edge chips, plus also putting Huawei on the entity list and subjecting to the foreign direct product rule so that it cannot continue to make these chips at TSMC.

14:30 I would say probably today, Huawei would be a peer competitor to NVIDIA at least. Right. And so we are in a world where export controls have worked to limit Huawei's capability very substantially. We're still in a world where I think Huawei chip production is something like maybe two or three percent of NVIDIA's production, if that.

14:49 Yeah. And I think it's worth adding one thing to the story as you told it, which was SMIC, which is generally regarded as the most advanced logic chip manufacturer in China, they had a remarkable defection of talent from TSMC. A lot of the best folks in TSMC's RD and engineering units defected and went to go work at SMIC.

15:11 And since then, SMIC has made astonishing progress in chip manufacturing technology. And if they had had access to EUV machines, not only would they have that awesome former TSMC engineer workforce, they would also have the smartest customer service people at ASML teaching them how to do anything and everything with their new EUV machines.

15:36 So I think Huawei, as you mentioned, already had world-class chip design. SMIC already had world-class chip manufacturing talent. And ASML and EUV machines would have provided the missing link for China to implement a policy that was on its books, right? The 321 plan, which says mandatory usage of domestic AI chips phased in over time.

16:04 And so, you know, NVIDIA loves to talk about how export controls have destroyed their Chinese market share. I think Chinese policy has destroyed their market share. And to the extent that it's been protected, it's been protected by export controls, which is why I think this sort of counterfactual analysis is so useful in understanding the stakes of this story.

16:26 So, Nick, when we talk about lithography in the media, really the breakdown is not exactly a fine-grained breakdown. It's usually just EUV, immersion DUV, dry DUV, and then maybe, maybe somebody will mention argon fluoride every once in a while. But the reality is there's actually a finer sort of within generations degree of technological progress for all of these machines.

16:52 And the way that the Biden administration designed the export controls on these machines is that they're targeting some types of machines with some restrictions and then other types of immersion DUV with tougher restrictions. I think that they thought originally that that was going to be enough to hold back China's AI progress and AI chipmaking capability in a massive way.

17:18 Why are you now calling for a more extensive set of restrictions? What's like help us understand the double click on what SAFE said about the different categories of immersion DUV?

17:29 Yep. Happy to talk about that. I think maybe for our purposes, we can within DUV immersion, we're sort of zooming into the DUV immersion block of the capability ladder. You can segment that out into what you might call three categories. The most advanced DUV immersion, those have the technical specs that are at sort of the top of the range.

17:50 For the real nerds out there, that's like ASML's NXT 2100 and its peers. Um and then there's what you might call the intermediate level of DUV immersions, and that's like ASML's 1980, and we'll talk a lot about, no doubt, ASML's 1980. And then there's even earlier generations, and that's like ASML's 1950 and even earlier.

18:12 And um, as you were alluding to, Greg, these are controlled differently by the US. And then we can talk a little bit about what the Netherlands does. The way the US controls work is that the most advanced tools are controlled throughout China. They cannot enter China. They're not allowed to enter the country.

18:30 The intermediate tools, um, which is ASML's again, 1980, they are allowed into China, but they are restricted from entering a specific set um of entities, uh which include a lot of China's advanced fabs, like the fabs at SMIC and CXMT. So that's what the US does. For what the Netherlands does, it's actually a little bit harder to know.

18:52 In the Netherlands case, it actually controls both sets of tools countrywide. It's not exactly clear how it then handles the license applications for any particular export.

19:02 Aaron Powell Let me just jump in here and add a little bit of plain English. The categories of export controls, this is kind of oversimplifying, but basically you can think about three buckets of restrictions: countrywide, end user, and end use. So countrywide says everybody from that country requires a license, and there might be a presumption of denial.

19:23 So the sort of default answer when you apply for a license is no. That's sort of what we mean when we say it's banned. Then you have end user. So if somebody like SMIC's most high-end facilities are on the entity list, then an additional set of requirements might apply to them, and that might be presumption of denial.

19:42 And then you have end use, which is what's going on in any factory in China, whether or not the owner of that factory is on the entity list. And if they're making fancy AI chips, that's a restricted end use. And so the license requirements would reply to them and potentially be reviewed with a presumption of denial.

20:01 So, as you're saying, the Netherlands doesn't quite do it this way. They just only have countrywide restrictions. And now continue your thought.

20:10 That's right. You're a great phraser of ideas, Greg. The Netherlands has just countrywide controls, but um, as I think we alluded to, there is ample public data, and we've sort of exhaustively combed through this in the report that demonstrates that a lot of licenses are granted within China. And that is specifically for that intermediate bucket of tools, and most particularly for the NXT 1980.

20:35 We've sort of found that there are vast numbers of these tools that are flowing into China. And also some reasonably compelling evidence that some of them are even going to some of the advanced fabs in China. There's public data that suggests that fairly strongly. It's again not exactly clear how the Netherlands is granting which licenses, but we know that a vast number are going in, and maybe some even specifically to advanced fabs.

21:00 And if I could add to that, in theory, this kind of regime could work if you have kind of very good oversight about which facilities the tools are going to, ensuring they're not diverted in some way to make chips that you don't want China to make. In practice, policing diversion is very hard. There's many versions of it.

21:16 There's obviously the case where you could just physically lift up a tool and put it in a new location, but then you also just have cases where the tool stays in the same place and now you start running an advanced wafer through it. And it's very hard to tell necessarily what exact type of chip is being made by that tool, or you may have a fab that at some point was a legacy fab and is upgraded to an advanced fab, and suddenly it's making a chip that unawares you didn't want them to make.

21:41 So one of the challenges is a kind of regime where some tools are allowed to some locations but not others. That can work. But given the challenge of diversion, a cleaner policy would be to not allow any of these tools to China in the first place. And this is especially doable in this case because one of the concerns you get about A China-wide ban for some semiconductor tools is that, well, okay, if you ban it to all facilities in China, you might impact legacy production, which maybe the export controls are not aimed at affecting.

22:10 Obviously, we can debate that point, but that has been kind of a part of the framework for US export controls, at least. Now, the good thing is for even these intermediate level DUV immersion tools, it turns out that they are not necessary for making legacy chips. We have a lot of analysis in the report that shows the yields that you get using the dry DUV tools, which are entirely uncontrolled and would remain uncontrolled under our proposals, can support, say, 28 nanometer production at SMIC.

22:38 And so a countrywide ban will actually have that surgical effect of limiting China's ability to make AI chips and other advanced chips.

22:45 Interesting. And you know, if you had to comment, we just did the counterfactual of what would happen if we had never had the EUV export controls. Now I want you to go the other direction. What if we had had full spectrum immersion DUV export controls starting in, say, October 2022?

23:10 Maybe to give a couple of pieces of background, this gets into why is immersion DUV so important. One fundamental fact is that it is essentially impossible to make AI chips without immersion DUV systems. We've got an analysis in our report that sort of gets into exactly why. Basically, it turns out that the nodes, the production nodes that are possible with earlier systems, dry systems, just do not allow you to make chips that are remotely competitive, either in computational performance or in power efficiency.

23:42 I think we should mention here that immersion DUV was the state of the art when it was introduced. And I want to say like 2003, I mean, this is an old technology. If you want to make AI chips with something older than this, you're losing a lot. Keep going.

24:02 For sure. And you know, we have an analysis. If you were to take one of these frontier scale clusters like OpenAI Stargate, we work through the numbers and to try to do it with the chips that you can make without a DUV immersion, it becomes laughable pretty quickly. It is simply not feasible. Now, in our report, we we have a sort of an analysis of China's installed base over time and their immersion installed base.

24:26 So how many DUV immersion tools of what generations and how that changed over time? And what you'll see is that by October 2022, uh, which is when I think the first set of chip controls um had come out, um, it was something around, I think, 50, uh maybe a little bit over 50, between 50 and 100 immersion tools.

24:44 Um, that has since exploded to somewhere around 340. Wow. So 7x increase in the number of these machines. I think it is it is not an exaggeration to say that if in October 2022 we had cut off all of the immersion systems, China's AI chip production would truly be crippled into the 2030s. It would be virtually impossible for them to uh make you know anything approaching uh large-scale chip production.

25:11 So, in that sense, when you look back, not to blame uh safer or anyone else in particular for this, but when we we look back, it is a strategic loss.

25:18 Yeah. Now let's, I'm kind of jumping ahead here, but I want to ask about reports that China is now making its own immersion DUV lithography, which I know critics of your argument are gonna raise. But before we do that, can you just explain something that you said before, which is that the legacy chips are still viable with the older technologies?

25:42 The immersion DUV is viable for certain categories of AI chips. And so I think, you know, maybe it's worth here you just sort of explaining why were the controls as we had them not crippling to China's ability to make AI chips, which I think was part of the theory of victory of the Biden administration.

26:04 In fact, I think part of the theory of victory was that like SMIC was gonna have to shut down uh their seven-nanometer line once they couldn't get enough spare parts. And so I'm just curious, what is, for example, multiple patterning? Why is it allowing China to continue making progress? And then when you say that the older machines could still allow China to build legacy chips in whatever quantities they want, is that with like multiple patterning on older chips, or what do you mean here?

26:31 So before Nick jumps in with the full technical explanation here, which he's best positioned to provide, I do want to say it's worth saying, you know, these controls are indeed having quite a significant effect. There's controls on edge tools, deposition tools, and the litho tools, they are not entirely uncontrolled.

26:46 As Nick talked about, they are subject to countrywide license requirements. There are ways to please how they're being used. And right now, production is quite low. And so one of the key things here is if you just give China the raw capacity in the long term, ultimately, as they potentially decide to mobilize and take more heavy efforts to divert the tools, they ultimately will have a very sizable capacity to be making AI chips at a comparable level as what the United States would be doing at that time.

27:14 But I do want to emphasize that the controls have been effective. We're just saying they can have been effective, but we can do more and basically finish the job to ensure that the US will have a very strong lead in the years ahead. But I'll turn to Nick to give the more specific explanation here.

27:28 Yeah, maybe just one more comment on that topic. There are obviously controls on lots of different complementary SME. There are controls on litho, but there are also controls on other segments on etch, deposition, metrology, a lot of other uh segments. It takes a village to make a chip, and the village is made out of semiconductor manufacturing equipment.

27:47 I will say that in our view, um, in a lot of ways, you can make the argument that the controls on some of these other segments have actually been stronger than the controls on lithography, that in some sense we have done a better job in these other segments. And then as a result, because we fully agree that the controls are having a massive impact, China is, after all, limited to just a couple of percentage.

28:10 And we think that in many ways, the controls that are biting them right now are the controls on etch and deposition. But one topic, and we'll no doubt hit this when we get talk about indigenization, a strategic weakness in our current position is nonetheless that China is indigenizing its SME supply chain.

28:29 And so over time, it's going to eat through all of these segments, and there will be a Chinese version of an etch and deposition tool. And the strategic loss is that we have very strong reason to believe that lithography will be the hardest segment for them to indigenize. It'll be the last tool that they indigenize.

28:46 And so as they get better and better at the complementary inputs, whenever they do that, they will unlock a larger fraction of their latent lithography capacity. And so we may see just surprising progress as they make these easier tools.

29:01 Oh, that's interesting. So you're kind of saying that we haven't even suffered all that we could potentially suffer from the immersion DUV that we've already sold them because they're bottlenecked on the other stuff that is more tightly restricted. And once they unlock that, they'll be using their immersion DUV to its full potential.

29:21 That is exactly what we are saying. Unfortunately, like if they were able to utilize their existing stockpile of immersion DUV to their full potential, we would be in a dramatically different situation. And one challenge is that they will probably eventually be able to utilize it to their full potential.

29:38 And we have got sort of some extensive modeling that shows what that looks like. We should definitely get into that high-level strategic picture because there are some relevant facts that fall out of that. But very much it is the case that the place uh where we had the strongest leverage, lithography, is where we have in some ways done one of the worst jobs of controlling.

29:57 And there's a loss of leverage there that we may look back on and regret in the coming years.

30:02 Wow. Very, very remarkable. Okay. So let's touch on what I briefly mentioned because you're saying that lithography is going to be the toughest thing for them to make progress on. We have been export controlling EUV in almost a total sense. And then in immersion DUV, there has been reporting that China is going to make five immersion lithography machines that are going to be shipped out to labs, and then 20 such machines that are going to be shipped out to uh semiconductor manufacturing facilities next year.

30:36 So it sounds like they are indigenizing what's missing from that story that I just told.

30:42 So, our high-level conclusion, having looked into this at some length, is that China is unlikely to make the versions of immersion DUV systems that are necessary for AI chipmaking at a large enough scale to materially expand their AI chipmaking capacity until into the 2030s. Now, maybe I can sort of talk about components of indigenization and then how that relates to the news that you were talking about last year and then also this summer.

31:13 I think roughly, in order for them to make immersion systems that really substantially expand their chipmaking, they sort of have to do three things. First, they have to master the basic fundamentals of the technology and of the science to the point where they can basically make a prototype that can even begin to do the node.

31:31 That's sort of milestone number one. Then you have to actually convert that prototype system into something that is good enough to be used in the lab. You might call that, you know, a commercialization, to the point where people can actually rely on it. Sort of as anybody in manufacturing knows, a prototype is quite a ways away.

31:48 And then there is a third step that I will have to say is almost entirely neglected in most analyses of indigenization, which is you have to scale up production of those commercial systems to a point where they materially expand AI chipmaking. And again, because in some ways, because we have given them so many immersion systems, we've already given them 340 immersion systems, to move the needle, uh they have to scale it up quite a bit.

32:14 They would have to be making high tens of these systems every year for a number of years to really move the needle. Now, to the point that you raised about the reporting from last year and this year. That's of Shanghai Yuliangsheng last year, and then it sort of got folded into Aishang Na this year. That is a DUV immersion system that is capable of 28 nanometer production.

32:37 And now, for the viewers out there that aren't Uber nerds, 28 nanometer is firmly in what's called legacy production. That's one of those nodes that is not sufficient for making AI chips that I mentioned, where there is a profound disadvantage if you try to make AI chips at that node. Meanwhile, the tool that we are currently allowing China to obtain, again, ASML's 1980, that's the tool, that's the name to remember, is sufficient for seven nanometer production.

33:05 That's, I think, about five nodes more advanced than 28 nanometer.

33:09 And I think SMIC would even claim that they're getting there at six nanometer. I think tech insights assessed one of their most recent chips as a six nanometer equivalent.

33:16 Yeah, arguably. And we actually, one of the things we did in this report is we sort of built a full yield model of um how these tools perform at different nodes, and we can talk about that if there's interest. But one very important point is that if you look at how long ASML, which is the world leader, they are the masters of lithography, how long they took to go from their 28 nanometer system, which is NXT 1950, to NXT 1980, they took seven years to do that.

33:43 So if you assume that China takes that long, that would already put us at 2032. Now we don't have to assume it takes that long, but that's just a benchmark. So there's a question how long it'll take them to go from the system they have to a prototype capable of seven nanometer. And I think it is reasonable to expect at least a couple of years and plausibly as long as ASML took.

34:02 There's reason to think they might go faster or slower. Then there's commercialization, as I mentioned. That is at least a year, a year or maybe two. We looked at how long this took in the past with ASML systems, and it doesn't happen overnight. It doesn't take a decade, but it takes a year or two. And then there's scale up.

34:19 And you know, we've looked at the lithography industry, and the fastest rates of scale up in the past are something like 60% year over year production. So when you stack all of those factors for China's companies to be making, you know, high tens of systems every year or over 100 systems a year, which is what it would take to really move the needle, that almost certainly is into the 2030s.

34:43 It is still quite a ways off. And maybe one more point I will make, and then I'll stop. We run the numbers on a lot of different scenarios. And I will say I've run the numbers on a scenario that is where they do unbelievably fast indigenization, where next year they make a seven nanometer capable prototype, and the year after, they commercialize it.

35:03 And then every year after they double production, 100% year-over-year production increase, faster than anything that's ever been seen in the lithography industry. And I will say that it does not change our fundamental conclusions, which are that with their current stockpile and what they will make indigenously, they will not match USAI chip production, almost certainly, but that continued imports of ASML systems can actually change that if they were to import at the pace that they have been thus far.

35:32 And I want to just add a couple more important points. And again, here, Greg, a point that you've made before about China's indigenization trajectory, how much are additional export controls actually going to change that trajectory? Or are we just handing them a bridge until they have their own indigenous tools anyway?

35:48 Why should we help them in the interim? That's one point. Another point worth noting is that China's actually never commercialized at scale any lithography tool across decades. Any, any, any that's dry DUV or advanced. Correct. Yes, yes, dry DUV and more advanced. And I will say if you look at the commercial market shares today, you know, with there's DUV immersion where they still don't have any real commercial market share, although they're starting to ship out maybe some prototypes.

36:15 You look at the dry DUV segments, and this includes argon fluoride dry, which is the more advanced dry DUV, and then Krypton fluoride, which is a less advanced dry DUV. And these are tools that go back into the 90s. China still has, as far as we can tell, and according to Tech Insights data, 0% commercial market share that we're being able to identify for those tools.

36:35 And these are generations behind the tools that we're concerned about here. And you would expect that they're going to climb the tech tree, you know, step by step, the less advanced tools you're able to indigenize and commercialize, then you finally get to DUV immersion. And that's another data point as well.

36:50 And then the last thing I'll say is that I remember tracking this issue all the way back in 2020, where you had Shanghai Microelectronics, which was their original champion for DUV immersion tools. Around 2020, they were announcing, aha, we have a 28-nanometer prototype. And I remember seeing these press articles about this, and then nothing really came of it.

37:08 And then I remember in 2022 or something like that, there was another piece of press. And then in 2025, there was another piece. They just keep coming. And so we've actually not seen a great deal of evidence that differs from these prototype announcements that we've been seeing for many years. The real key challenges are going to be the ones that Nick described on commercialization and scaling.

37:26 Yeah. And I think there's like a whole ecosystem of China boosters and export control opponents in social media, where every single time, you know, China has one of those announcements of like, hey, we've got a prototype. Everybody says, wow, doesn't this prove that export controls don't work, will never work, we'll only backfire?

37:43 And it's like, yeah, guys, you got to read past the headline. You got to actually think about the strategic implications, which is what you've done in this paper, and you've done it, I think, very, very quantitatively and scenario driven. And so you've already talked about a lot of what's in this paper, but let's like just go high-level, step by step.

38:05 What were the biggest scenario models you ran? And what kind of outcomes do they make for AI chips, for high bandwidth memory? And what should policymakers know? And why is it persuasive that we should do what you're recommending, which is a total countrywide ban on immersion DUV, if I'm not mistaken?

38:27 That's right. Maybe we can start with the high-level findings and then drill down into uh the very gnarly details. I alluded to this just before, but to sort of summarize, there's maybe two, really just two fundamental pieces of information that I think are policy relevant. One is that with their current stockpile, with China's current stockpile of DUV emergent systems, which is a very large stockpile, and arguably, as I mentioned, a large strategic loss that we've allowed them to accumulate that stockpile.

39:00 Nonetheless, if they were to take every single one of those systems, including the ones that are in legacy fabs, and divert them to their advanced fabs, and they were to take 100% of those systems at advanced fabs and dedicate them to AI chip production, and if they were to indigenize every single other piece of SME, all of the etch, all of the deposition, everything else is indigenized, and they were to achieve world-class yields, the same yields that TSMC and SK Heinex achieve.

39:29 If all of those things happen, which is an extreme scenario, truly, even then, they will not match USAI chip production through 2035, I think, by any reasonable analysis. They are just physically limited by their litho stockpile. This is like a fundamental strategic fact that every policymaker should have tattooed across their forehead.

39:51 China's litho stockpile is insufficient to match US AI chip production. Another way of saying this is that the United States has one, at least if you're looking at aggregate chipmaking capacity. However, there is a part two to this. The United States has one in that sense if we close the barn door. Because the part two to this is that if you just look at China's pace of imports over 2024 and 2025, like how many of these DUV systems did they import?

40:22 And you assume that they do something similar to that, maybe a little bit higher, although even not necessarily, each year in the coming years, into the 2030s. And then you assume that they do, you know, a moderate amount of mobilization. Not every tool, but they achieve the same allocation of their tools to AI that is achieved outside of the United States, and assume that they indigenize the other SME bottlenecks.

40:45 And that that last assumption, of course, there is some uncertainty there. They have not, as we talked about, indigenized all the rest of SME. But as I mentioned, there is extremely good reason to think that they will indigenize all the rest of SME before Litho. What's the extremely good reason? So there's sort of two ways of looking at it.

41:03 One is that they are making progress. If you look, CSET has a great report on their progress in etch and deposition, deposition which has got some very tough tools. You know, those were the initial set of tools controlled in 2022, and that are really biting them. AMATS, their Endura Volta, and similar systems, they're gaining market share.

41:19 They gain maybe four points of market share. Don't quote me on that, but go read the report instead. Meanwhile, of course, in Litho, they're gaining no market share and they have no market share. There is also just many other fundamental reasons, just when you look at the technology, the RD expenditure, the cost of these systems, which is between three and twenty times more than other kinds of SME, how long it took to develop them, how many companies flamed out in trying to develop them, to have extremely high confidence that litho is by far more difficult to indigenize than these other systems.

41:54 So if we assume that China does indigenize the rest of SME, which again, we should not underestimate China, I fully do expect them to indigenize the rest of the supply chain. But if they do indigenize the rest, there are scenarios in which their chipmaking, assuming they achieve the same allocation as outside of China, where it approaches parity, and even some where it could surpass the US if the US has slightly reduced and they have slightly higher.

42:20 At minimum, the gap narrows enormously. All of these are normalized to H-100 equivalents. This accounts for the fact that the US can make better chips.

42:39 So I think another very, very important conclusion based on what you just said is EUV not enough. All the tough controls we have on etch, deposition, et cetera, not enough. If China continues flooring the gas pedal for the next decade, there is a scenario in which their compute output per year in H-100 equivalents surpasses the United States, driven entirely by immersion DUV.

43:12 A remarkable conclusion.

43:14 That is right. Now, you know, to be clear that there is a wide uncertainty here. Yeah. But uh what is undeniable is that the possibility certainly remains open. And the counterpart is that the possibility is not open if DUV immersion is controlled. You can achieve near certainty for a number of years.

43:33 And that is a pretty dramatically different strategic situation. There is really a vital choice here for the United States governments as to how seriously we take competition with China and AI.

43:44 And one piece of color on that is just zooming out to the AI competition as a whole, you look at a lot of the factors of production in AI, data, algorithms, investment, talent, these are all things that China has. The two places where we differ are US. Has a compute advantage and the tools to extend it further.

44:03 And China has an energy production advantage. And so these Huawei chips that they're making at seven nanometers, yes, they're a little bit worse. They cost a little bit more to run. They're more energetically expensive. But one thing we already see is that China is willing to build up electric capacity to run these chips at somewhat worse economics, but from a strategic standpoint, they will have the computing power to develop advanced AI systems and run them.

44:27 And so they're kind of basically using that advantage in order to offset some of this disadvantage of not being able to make chips with EUV that are more efficient per chip or per computational unit of performance.

44:39 So it's like the calculation is flops per gigawatt, and we can do a lot more flops per gigawatt for every chip, but they can do a lot more gigawatts. Correct.

44:49 Yeah. That's right. Just one more point to add on to that. One of mine and many others' pet peeves, I think, in strategic competition, you know, chip issues, uh, which is much better now, I will say, than it was, is that there's often a reluctance to think about quantity and scale and how many chips.

45:06 You know, people often want to compare a chip versus a chip. Uh, but ultimately, a flop is a flop is a flop. From a software perspective, if you can obtain the flops, even if you have to spend a little bit more energy to do so, it really doesn't matter. At the end of the day, you're just shifting bits around.

45:21 So it is, you know, EUV does enable more economical, more energy efficient chips. But if you are willing to just floor it on the total chip production, if you've got 100,000 seven nanometer chips, you better believe that you're gonna have more computational power available to you than 103 nanometer chips.

45:40unknown Yeah.

45:40 Coming from a former Google software engineer, a flop is a flop.

45:45 So I think maybe it's helpful to take your point about people focus so much on the one and not so much on the scale. Elon Musk has this line that he uses to describe the challenges faced by Tesla and SpaceX as they were scaling up production. And he basically says designing a rocket is easy. It's like completely unimpressive to design a good rocket.

46:14 Making a single prototype of the rocket is hard. But compared to making the factory that makes a lot of rockets that don't blow up over time, designing the rocket was literally nothing. Making the prototype was barely anything. And like 90% of the story was making the factory that makes, you know, 100 rockets a year or what have you.

46:43 And so I think that is the failure of the policy discourse on AI chips, is we say, oh my gosh, look at this Huawei chip design. Look at this Huawei server design, right? And what we really need to keep our eye on is what is the factory producing what quantity at what yield. And then when they put it in a server or in a data center, do they all work together well as a team?

47:14 The different levels of scaling in the AI story, there's big failure modes at each one of those. And as we're telling the story to policymakers, you've got to keep your eye on that ball. So you've talked about the numbers of machines. There are 340 of these machines already in China, I believe you said.

47:33 How many are we shipping per year? And per year, how does one of these machines operating at full bore, how many AI chips can it make?

47:43 Aaron Powell Yeah. So I will say the the shipments, you know, they they vary year to year. Certainly in 2024 through 25, which are the full years where we have uh most of the data, China imported a really large number. We estimated around 90 in each of those years.

47:57 Aaron Powell And what share of like ASML's sales of these kinds of machines is that?

48:02 Yeah, I think there is this fundamental question like, is 90 a lot or is it a little? Um and I can say definitively, in immersion DUV standards, 90 is a lot. That's over half, I think close to two-thirds of ASML's entire immersion market for the world, for the entire world. It's over $10 billion. And more critically, from a strategic perspective, when we're thinking about AI and AI chips, I think a number that is worth keeping in mind is that we mentioned that these 1980 systems, they enable seven nanometer logic, right?

48:32 That's the logic that's used for Huawei's Ascend chips. And one of the fundamentals in this paper that sort of has allowed us to draw the conclusions that we have drawn is we've modeled out not just their yields, but also their throughputs, which is how many chips can these systems make if they are installed in a FAB?

48:49 And again, in the same way that we care more about the total computational power than how much any one single chip can do, throughput matters a lot because at the end of the day, you care about how many chips do they have on hand. And what we have found is that the 1980, ASML's 1980, can do on the order of hundreds of thousands of always chips.

49:09 It depends on the yields, but even assuming fairly poor yields, even assuming the yields that we think that China's getting at SMIC and CXMT, 20%, 30% logic, it is still capable if it is paired with complementary SME, the etch and deposition of producing on the order of 100,000 SN 910 C's a year. That is a vast quantity.

49:31 Oftentimes, certainly in the last year in the licensing debates over the H200, there's debates over, you know, shipments of 50,000 or 100,000. A single one of these systems pumps out 100,000 every year. So that is the strategic significance of every single one of these machines. And so when you think about, well, 90 of them went into China in 2025.

49:52 I think that can start to bring the strategic impact into view.

49:56 And if I could just add to that, it would just take one or two tools of yearly production at high yield to get an aggregate amount of flops that is in XAI's first Colossus supercomputer that was built in Tennessee. So Colossus is big. So we don't want to give that many tools.

50:13 Got it. Now let's shift gears to implementation, right? Let's just say it's Monday morning at the Commerce Department. BIS has just read your guys' paper. It's also at the White House, they've read your paper, they find the argument compelling. What's the right way to structure a crackdown on immersion DUV exports?

50:37 So I think there are there are a number of different ways that this could happen. The first way, a way that has, you know, been tried in the past, and you know, in many ways, the the preferred solution, I think everybody's preferred solution would be that the countries that make these systems control them multilaterally using their own national authorities.

50:57 So that means the Netherlands controls it and that controls ASML systems. Japan controls it and that controls the Nikon systems. This is ultimately the preferred solution. I think everybody would would like to keep it to national authorities. However, one thing that we go into detail on in this report is that the US has a number of tools at its disposal if that solution is not viable, if the Allies are not able to be convinced.

51:22 Tools that it has used in the past and that are readily available in this case, to briefly go through them. One, I think, very important point is that in fact, the US and Japan, if they could reach agreements, they within their own national authorities have sufficient authority to cut off all sales. The reason is because the US and Japan are the soul makers of what's called the light source, which is this extremely important subsystem within a DUV immersion system that if you don't have a light source, you don't have a DUV system.

51:50 So if Japan and the US were to jointly agree we are going to control light sources for the end use of incorporation into a DUV that is destined to China, it would have the effect of controlling all DUV worldwide. Now, putting that solution on the shelf for a moment, the US also has uh well within its authorities the ability to unilaterally and extraterritorially control those systems.

52:14 And there's sort of two relevant authorities there. One is the foreign direct product rule, which allows it to control systems that are the direct product of U.S. technology, and all DUV systems worldwide meet that criterion. There's also the De Minimus Authority, which allows it to control systems that incorporate more than a certain amount of US content.

52:33 This gets into the weeds of export control regulations, but you know, that's uh specific US parts that are physically inside the machine. And it's pretty well known that ASML systems at minimum meet that bar. So um there are a wide array of tools that are already have statutory authority that the US can control all of these systems uh nationwide to China.

52:54 Aaron Powell And I would just add that there are sufficient executive branch authorities and diplomatic capabilities to get this done without any new legislation. But of course, this is all in the backtrap of a really important bill that is currently being considered in Congress, and it's in the Senate NDAA, which is the must-pass year-end defense authorization bill.

53:13 And there's now conferencing going on between the Senate and the House on whether to include this in the final package. And the MATCH Act has a slate of kind of mandated emiconductor equipment export controls, and it has a period where it mandates the U.S. government to negotiate with allies to achieve controls on a number of pieces of equipment.

53:32 And then if a certain deadline passes, to use some of the unilateral authorities that Nick has described to control those tools just with executive branch authorities. And one of the most important pieces of this bill is that it would mandate a China-wide restriction on DUV immersion tools. And I would say probably the most important part of that bill to be preserved.

53:52 And one other benefit of going the congressional route is a big question of how to manage export controls right now is the Chinese response and retaliation. Now, honestly, there's obviously always going to be a risk that China retaliates and policymakers are just going to have to weigh the benefits and the risks.

54:08 In my view, this is one of the most strategic actions you can take. And I think there is a lot of asymmetric leverage for the US to do this without incurring any comparable pain in response. But I think one critical aspect of how to do export controls is that, and I think that for the US government, it's not done as well as it could, is don't put export controls on the negotiating table because that heavily incentivizes China to retaliate in the hopes that it could cause the United States to withdraw export controls.

54:35 This happened with the affiliates rule that BIS issued last year. And then other export controls on EDA software and the like also got caught up in trade negotiations. And I think this does just create an incredible incentive for China to retaliate. One benefit of doing it with the MATCH Act is that Congress will just pass a bill that is in statute, cannot be rolled back.

54:54 Congress, as hard as they work, is unlikely to pass another bill that is going to roll back these controls. So it at least reduces that incentive. I'm not going to say that this doesn't mean China is going to respond, but nevertheless, I think it at least kind of reduces some of the chances of the most dramatic retaliation outcomes that we could see.

55:11 Very cool. Well, gentlemen, this was an amazing conversation, and I'm really grateful to everybody who listened to it. But what they should also do is go read the paper. So what's the right way for them to find you, find your work? Techstatecraft.org/slash DUV. Nice. You got the URL before anybody else.

55:29 All right. Well, Nick, safe. Really enjoyed the conversation. Uh, I've enjoyed following your work for a long time now. And thanks for coming on the AI Power Podcast. Thanks, Greg. Thanks for having us. Thanks for listening to the AI Power Podcast. If you liked what you heard, there's an easy way you can help.

55:47 Please rate us five stars on your favorite podcast platform. And tell your friends every bit helps.