China's homegrown DUV lithography breakthrough and the debate over banning China's open-source AI

Two big stories this week. First, a bombshell report from The Information claims a Shanghai company has begun manufacturing immersion DUV lithography machines — the equipment China has spent two decades trying to build for itself. Greg and co-host Adam Goodwin unpack what "immersion DUV" is, why five machines this year is both respectable and decades behind ASML, and the buried detail that may matter most: the Japanese components quietly helping China claim self-sufficiency. Greg argues the real prize isn't new machines but China's growing ability to service its huge installed fleet — and why every day BIS delays burns leverage the U.S. won't get back.

Then: the Trump administration goes on record about Chinese open models — the Kratsios and Bessent statements accusing Moonshot AI of distilling Anthropic's Fable, the GB300 smuggling angle, Beijing's "AI hegemonism" rebuttal, and where the MATCH Act stands. It closes on the open-source debate itself: Dario Amodei's response to the NVIDIA–OpenAI–Google letter, and why the real question isn't open versus closed but what safety testing should apply past a certain capability threshold.

Chapters:
(00:00) Cold open & the week's two big stories
(01:02) What immersion DUV lithography actually is — and why it matters
(05:46) Why China wants to build these machines itself: chips as the new oil
(09:11) A decades-old goal: Made in China 2025 and the roots of indigenization
(13:19) Five machines this year — is that a big number? ASML for context
(18:15) China's disinformation campaign vs. real scoops, and the Japanese-components problem
(24:09) The spare-parts endgame: servicing the installed fleet as the real prize
(30:19) BIS Under Secretary Kessler's testimony and "incomplete" allied cooperation
(36:57) The MATCH Act, the foreign direct product rule, and the NDAA path
(39:51) Kratsios and Bessent on Moonshot AI, Kimi K3, distillation & GB300 smuggling
(45:32) China's Ministry of Commerce hits back: "AI hegemonism"
(53:31) How industry is reacting — the letters, and who signed which
(59:31) Wrap-up

See Greg's full research archive and the Decision Tree newsletter at decisiontreeresearch.com.

Transcript

SPEAKER_00 0:09

Welcome back to the AI Power Podcast. We've got a great episode. We're gonna focus on two big stories. The first, a story on a claimed breakthrough on Chinese semiconductor manufacturing equipment. And the second is about government and industry responses to the Chinese open AI models. Let's start with a pretty amazing report from the information on July 27th, which says a Chinese state-backed company has begun manufacturing one of the key pieces of equipment used in chip manufacturing for the first time. A significant development in Beijing's drive to reduce its reliance on foreign technology to make chips. The Shanghai-based company is making immersion deep ultraviolet lithography machines used by chipmakers such as Taiwan Semiconductor Manufacturing Company, TSMC, in the production of chips, according to two people familiar with the situation. Greg, I know this is a big deal, but let's start with the basics. What is immersion DUV lithography equipment?

SPEAKER_01 1:07

There's a lot of different types of machines that are required to make modern semiconductors. What you are doing is you are taking a big silicon wafer, effectively a big uh circle of silicon, and you are turning that into computer chips. A ton of different machines, a whole manufacturing facility full of machines are required to make that happen. And lithography is arguably the most important one. That is the machine that is printing the design of the semiconductor chip onto that wafer. The original lithography is like the printing press. And so here we're printing with light, and we are using lasers to print the circuit design onto these chips. There's different generations of this technology of lithography going back many, many, many decades. The hot one right now, the one that folks have probably seen in the news a lot, is EUV, extreme ultraviolet lithography. What we are talking about here is one generation older than that. Immersion DuV or immersion argon fluoride, because it's about the chemical and what wavelength of light it produces. So immersion means it's immersed in water. There's also a dry, meaning using air version of argon fluoride. And then we go back to krypton fluoride and eyeline. Now, these technologies are older than you might think. Like the latest and greatest EUV machines have a legitimate claim to be the most complicated, the most advanced machines that humans make today. I would actually argue that, in fact, they are. The latest and greatest EUV machines are the most advanced machines that humans make today. Well, the first EUV machine was introduced in 2013. So we've been here for a while now. And the first DUV machine, which was krypton fluoride-based, was introduced in the late 1980s. Then they switched to dry argon fluoride in the 1990s, and finally argon fluoride immersion DUV in the mid-2000s. So you might think the difference between dry and immersion DUV is just add water, but actually, no, it's way, way, way more complicated than that. You have to change everything about the internal makeup of the system. So what we're effectively saying here is that China has matched a technological achievement that the Dutch and the Japanese, the other two big providers of lithography systems worldwide, they reached in the mid-2000s. So you might say, oh, well, okay, China achieves something that the Dutch achieved 20 years ago. What's the big deal? Well, immersion DUV is the generation of technology that can get you closest to modern chips. And by using techniques such as multi-patterning, you can actually stretch DUV to make stuff that is pretty advanced. So the latest and greatest generations of chips, we're talking about two nanometer or one nanometer. And those names, despite sounding like they're measurements, actually don't precisely correspond to the measurement of the transistors on the chips anymore. They just really refer to the performance nodes that loosely translate to meaningful step increase in performance. So the with Immersion DUV, SMIC, a Chinese chip manufacturer, has demonstrated that they can make chips corresponding to the seven nanometer node. That is what iPhones were using in 2019 for their most advanced chips. It's what NVIDIA chips were using, I want to say as late as 2020. They were still making seven nanometer chips. So right now, Chinese chip manufacturers like Smic, when they're using immersion DUV systems, they're buying them from ASML. And what this is saying is now they can produce those systems locally, but like at what quantity, at what reliability, those are still the big unanswered questions from this story.

SPEAKER_00 5:25

Understood. Understood. We understand that chip manufacturing is a global enterprise, right? It is a global industry. But why would China want to produce DUV machines locally? And what is the specific strategic importance to them for that manufacturing?

SPEAKER_01 5:41

Yeah, so there's uh two basic reasons why they would be interested in this. The one is just the overarching goal of China dominating manufacturing. The reason why they would want to make these machines is the same reason they would want to make industrial robots. It's the same reason they would want to make cars, it's the same reason they would want to make airplanes. It's just a big market where their companies can grow in technological sophistication, capture market share, et cetera, et cetera. That's the generic reason. There's also a very specific reason that relates to semiconductors, artificial intelligence, and the US-China geopolitical rivalry in AI. So to talk about this, I think it's worth grounding it in like what are semiconductors as a strategic asset. Sometimes you hear chips are the new oil, and uh that's not precisely true, but it is a useful analogy in a lot of ways because if you can't get access to chips, it really hurts economically. We actually have a case study of this, which is the 2020 COVID pandemic. Um, that caused a supply chain issue in the global semiconductor manufacturing ecosystem. Especially US automakers couldn't get enough of the chips that they needed to make cars. And the Commerce Department estimated that just that shortage in semiconductors shaved a full percentage point off of GDP. That's like cutting US GDP growth in half just because you can't get enough chips. So if you want to be, you know, we often hear like the phrase energy independence. Like a country would want to be strategically not reliant on foreign oil. Well, China has for a long time had the strategic goal of not being reliant on foreign chips. And in fact, every year, China imports more by dollar value semiconductors than they do oil. And so if you want to be independent in semiconductors, you can be independent by buying a lot of foreign machines. But then that begs the question: what if you get cut off from those foreign machines? And since 2019, China has been unable to buy Dutch EUV machines, extreme ultraviolet machines. And there have been discussions going back to the National Security Commission on Artificial Intelligence report from, I think it was 2021, that was calling for also using export controls to ban Chinese imports of immersion DUV machines, argon fluoride immersion machines. So China, seeing that they're already cut off from EUV machines, seeing that people are proposing cutting them off from DUV machines, they view having a domestic capability in manufacturing these machines as a critical strategic goal.

SPEAKER_00 8:42

So, Greg, understanding that uh China has been working on this and they have been cut off from certain parts of uh lithography technology since 2019, how long have they actually been working on this? What does their long-range goal look like here?

SPEAKER_01 8:59

Yeah, so I think if you read like the news, you will find people who are saying, oh my gosh, this is all a reaction to US export control policy. And there's parts of that that I think are consistent with the evidence pattern, and there's parts of that that I think are inconsistent with the evidence pattern. So to begin, China has been saying they want domestic DUV immersion machines for a very, very, very long time, right? In Made in China 2025, that was a policy that was implemented and came out in 2015 in China. So, what meaningful export controls apply in 2015 on semiconductor manufacturing equipment? Like none, right? We're willing to sell China anything they're willing to buy in 2015. And what did their official industrial policy goals call for there? It said 50% localization of 90 to 32 nanometer process equipment before 2020, 50% localization of 90 nanometer lithography machines before 2020, 30% localization of 20 to 14 nanometer process equipment before 2025, domestic production of immersion deep ultraviolet lithography machines before 2025, and domestic production of extreme ultraviolet lithography equipment before 2030. And the policy roadmap said this is a full quote meeting domestic demand, improving the self-sufficiency rate of integrated circuit products, meeting national security needs, and occupying the strategic product market have always been the greatest demand and driving force for the development of the integrated circuit industry. And it's not just made in China 2025, right, which is uh a policy dating back to 2015. Even earlier policies, um, such as the medium long-range plan for the development of science and technology explicitly called for self-sufficiency in semiconductor technologies in 2006. And China started investing tens of billions of dollars in this localization drive with the big fund, which got started in 2014. So, on the one hand, yes, part of the reason why China wants this technology so urgently is because they are afraid of being cut off from foreign access. But even before cutting them off from foreign access was under discussion, they wanted to have this independent capability. And as I say often, if you look at other technologies like electric vehicles, like solar panels, nobody was ever talking about cutting China off from foreign technology. And they indigenized those industries fully too and kicked out all the equipment manufacturers of solar panels. So, yeah, there's there's parts of this that I think you can point to and say, hey, US export controls did meaningfully shape the strategic calculus of Chinese companies. I think the strongest case is that that started back in 2018-2019, right? When we first cut off Huawei from manufacturing chips at TSMC, or cut off ZTE from manufacturing chips at TSMC by application of US export controls. And then again in 2019, when we cut them off from EUV machines and similarly advanced machines of different types for semiconductor manufacturing equipment. So by the time the Biden administration rolled around, by the time the second Trump administration rolled around, the Chinese policy was already fully baked.

unknown 12:42

Yeah.

SPEAKER_00 12:42

Yeah. And I think taking that quote, if you simply remove integrated circuit products from the quote, it could apply to a number of different things that the Chinese government is pursuing. Totally, period.

unknown 12:52

Right.

SPEAKER_01 12:53

Jet engines, rockets, cars, construction equipment, robots. Yes.

SPEAKER_00 13:00

So let's assume for a moment that this report is entirely true, right? The sources interviewed by the information claim that the unnamed Chinese company will make five such machines this year and 20 next year. So put that in context, right? Because these machines are fairly exquisite, very difficult to manufacture. Uh is that a big number? Is that a small number? What does that look like for a producer of this equipment?

SPEAKER_01 13:23

So for a first-year production number, that's a perfectly respectable number, right? If you go back to the history of ASML, the Dutch lithography company, um, they're fab testing argon fluoride immersion DUV machines. They were beginning testing in real fabs in November of 2004. And they probably made 13 or so machines in 2005. Last year, because these machines are still in demand in China and elsewhere around the world, although that this is the upgraded variant of DUV machines. But last year, ASML sold 131 of these machines. So making five is not like an embarrassing number for China's first year, um, but it's also nowhere near where ASML is today in manufacturing these numbers of machines. There's also the question of just like, how good are the Chinese machines? When you're making circuitry at the near atomic level, reliability matters, speed matters, um, downtime matters, these are all the metrics that are separate from how much the machine costs that really impact what the return on investment on those machines are. And ASML is awesome at all of those things. And so the question is like those five machines that uh supposedly this company is going to be selling to Chinese customers like SMIC, like Hahu Hong, are they selling those five machines because the customers actually want to use them? Or are they selling those five machines because the Chinese government said, we will make it free? You can just have them, and we want you to help this company grow in technical sophistication. So, like, are they selling five machines that are gonna be there, or are they selling five machines that are remotely as good as even ASML's cheapest and lowest performing product? We don't really know. I mean, one number that I didn't say, which I should, is uh if you add up all the machines that ASML has ever sold of immersion DUV, it's probably something like 1,700 to 1800 immersion scanners. A lot of those, not all of them, but a lot of them are in China because ASML has been selling a ton of these machines to China. Now, it's it's worth pointing out the export controls that we've had in place uh since 2022 and updated several times since did restrict sales of the most advanced ASML immersion DUV machines, but the older generations of such machines um are still allowed to be sold to China with some restrictions. So that's like kind of the question is like, you know, how much additional marginal capacity growth is China getting from these five machines, from these 20 machines? And if they continue the ramp up that we saw from ASML over time, they could probably be getting to making something like 50 machines a year in a few years. So they're going to do this. But it's it again, it's worth putting that in context. It's it's an immersion DUV machine, but how good of an immersion DUV machine is it? And this is technology that ASML uh was selling on the open market back in 2006. So, like they're still decades behind the global state of the art.

SPEAKER_00 17:03

100%. So if true, with the caveat, if true, uh we have to imagine that China is very excited about this, but it's also uh information that was not uh coordinated with uh President Xi's previous speech, and it seems that a lot of uh kind of uh AI breakthroughs, whether it's the Kimmy K3 rollout or other things, were coordinated with that speech to make it seem as though China is really advancing uh kind of all at once. So I do you think this is true, Greg? And is China very excited about this?

SPEAKER_01 17:39

So I think it's it's worth pointing out that there is like an active disinformation campaign coming out of China that is designed to make the US export controls seem futile, and they're trying to highlight various breakthroughs with the hopes that US policymakers will be like, oh my gosh, you know, if this is true, then the export controls are stupid and we never should have done them and we should terminate them. Um, examples of this sort of thing is like how much the Chinese government hyped uh use of particle accelerators as a light source for EUV machines, which, like, if China's wildest dreams come true, that's gonna solve their problem in like 20 years. Um, another one was like some account that was promoted by state-backed um propaganda accounts on X that was claiming that China had an EUV machine, and then it showed a picture that had like the label EUV on it. Um, but that machine is not an ASML lithography machine, it's not even a lithography machine, it's like a totally different thing. So, like they are willing to lie. There is an active goal of lying about where they are, trying to make the export controls seem like a bad idea. That said, the information has had great reporting on stuff that China didn't want to get out, including on chip smuggling, including on equipment smuggling. Like, if folks out there don't subscribe to the information and you're interested in the AI and semiconductor beat, like subscribe to the information. They get a lot of good scoops on this sort of thing. So that leaves you asking the question. If it's not a disinformation campaign, which in this case I don't think it is, then it brings the question that you asked, Adam, which is like, why are we getting this from an information interview with two unnamed sources who won't even name the Chinese company that's making these machines because they think it's too sensitive. Why are we getting it from that as opposed to like a government-backed press release? And I think the answer is in the information article. Um, so here's the quote from the article that I thought was really, really juicy. Quote, the Chinese DUV uses mostly domestic components, though some key parts come from Japan, the two people said. Delays from local suppliers have held back production this year, they added. And I think that tells you why China didn't want this news getting out. They didn't want to say it loudly and proudly, because there's an obvious policy next step that the United States government needs to do. Who in Japan is helping China do this and how do we stop them immediately, right? And this is the point. Is China making an immersion DUV machine, or are they like taking critical Japanese components and assembling them into a quote-unquote Chinese machine? And that really matters when we're talking about how quote unquote self-sufficient is China in this story. If they're still critically dependent upon foreign components, then what you have here is, at least according to the reporting of the information, Japanese companies who are helping China and hurting Europe and hurting other Japanese companies that are critical input suppliers to ASML. And if you are the Japanese government, uh, I hope that it is obvious to you that this is utterly unacceptable and that you have like one tiny company that's helping itself by hurting your overall strategic position vis-a-vis China. So I really hope that the United States intelligence community, frankly, is going after this story and trying to figure out what is the supply chain that is enabling this level of capability and how do we shut it down. Now, again, we're talking about 2006 era technology here. So it's not like we're gonna stop China from having this forever. But old technologies can still be strategically important. If you have a 1950s nuke, I'm still scared, right? If you have a if you have a 1990s jet engine, you can still uh compete in the commercial market and you can compete. Compete in the military with those sorts of technologies. And at least in the terms of semiconductors, it really does matter strategically what share of these machines are in China, what they can do locally. And every additional year that we can make it hard for China, I think is useful in terms of American and Allied economic security, minimizing China's coercive potential vis-a-vis, increasing our coercive leverage vis-a-vis China. And it's also strategically important in artificial intelligence because obviously China's long-term goal here, and they would prefer it be a midterm goal, is domestic mass production of EUV machines so that they can be completely independent and have more AI chips than us for the next phase of the AI competition. And while immersion DUV machines are not useful in making the absolute bleeding edge state-of-the-art of AI chips, they are still useful in making AI chips of a certain level of quality here. And so I don't think it's in America's interests to make it easy for China to get access to the machines that can help them build that. The other part of this that I think is quite important is spare parts and components. So in the past, people have told me that Chinese semiconductor manufacturing equipment, not just in lithography, but like in a lot of stuff, is just a straight millimeter for millimeter copy of foreign equipment, right? Whether you're talking about etching, deposition, lithography, inspection and metrology, all the sort of big components of semiconductor manufacturing equipment, Chinese machines are often copies of foreign machines. And that's useful for multiple reasons. It's useful not just because it shrinks your research and development window, right? Instead of like having to figure out how to do something, you just have to figure out how they did it. But it's also useful because it means that your domestic supplier can plug into the existing fleet of machines that already exist. So like one of the things that China is really scared of right now is that they are going to be cut off from spare parts and repairing services of all the equipment that's already in China. And some of this stuff, you know, like the light source in an EUV machine, I think degrades after only one or two years. So, like some of the parts in semiconductor manufacturing equipment is just like valves, right? And China can make the valves locally. But some of it is super fancy, ultra precision technical equipment, where if they can't get spare parts from abroad, when that stuff breaks down, it shuts down the whole machine. And so China has been buying an insane number of immersion DUV machines from ASML for the past few years as they're trying to capture market share in legacy semiconductors in order to give themselves choke point control of legacy semiconductors as another economic weapon similar to that of rare Earths. Well, if we were to suddenly cut off servicing of those machines and cut off exports of spare parts of those machines, that means all of those machines could go down, right? But if the Chinese machines are millimeter for millimeter copies of the ASML machines, then it means at the same time that they're growing in sophistication of local production of the machines, they're also growing in local sophistication of their ability to repair the existing installed fleet, which I expect is a priority strategic goal for them. And again, these priority strategic goals, like nothing we say to China is gonna change their mind that they need to eliminate their dependence on ASML as fast as humanly possible, right? The the brick is already on the gas pedal, the brake lines have already been cut. Um, we cannot persuade them not to do this, but we can make it complicated and expensive. And that's what I'm thinking about here. And and that is the subtext of this point is not just how many machines can they make per year, but what might be their spares and repairs manufacturing capacity in the next few years?

SPEAKER_00 26:27

Understood. Understood. So getting back to the actual horse race part of this, um, let's assume for a second that the basic outline of this information is true. Uh, how much of a breakthrough is this for China, right? What does it suggest about where they might be in five years per this technology and the rest of the world?

SPEAKER_01 26:49

So I think they are fundamentally achieving a milestone that ASML achieved 20 years ago. But ASML also had their first EUV machine uh prototype, I think in 2010. Um, so that was like an important milestone. That's uh 15 years ago. So if they're hitting immersion DUV now, if EUV is still a crash national priority, a very simple, oversimplified, frankly, analysis would suggest that they might be near the EUV milestone in a handful of years, right? Um, certainly at the lab prototype level. And once China has a domestic EUV machine, um, that is going to be more strategically significant than this milestone. And I think it's worth pointing out, right, like this is what we were continuing to sell, and it didn't dissuade China from choosing to indigenize it. And so the question that I think I go back and I look at the Biden administration policy on immersion DUV machines, frankly, as a failure, because now you have this massive install base of DUV machines that China is still going to try and indigenize. And now China is a huge share of the global manufacturing of chips that you need immersion DUV from. So if you cut them off from access to the spares and repairs market, you know, you're you're you're meaningfully affecting global chip production in a way that could be painful economically. And so now China is able to unfortunately hold the world hostage as a supplier of legacy chips in a way that they probably would not have been if we had not embarked upon this road. So again, I think the most important thing for China near term is not necessarily their ability to make these machines locally, it's their ability to keep the machines that they have locally working. And that's probably what they're thinking about. That's probably the most strategically relevant development so far.

SPEAKER_00 28:56

Understood. So uh you you mentioned administration policy kind of and how it affects uh kind of these developments. So I want to talk about what, if anything, U.S. policymakers are working on that is relevant to this issue. Uh, we've spoken a bit in prior podcasts about the Commerce Department's Bureau of Industry and Security. Uh, can you reflect on what we learned from the BIS undersecretary July 14th in his testimony on this topic?

SPEAKER_01 29:25

Yeah, so this this did come up. Lithography, machines, semiconductor manufacturing equipment did come up in the testimony. I think probably the most pointed moment was a question from Representative Kim. And here's what he said, and here's Kessler's response. So, quote, last June you testified before our subcommittee on South and Central Asia that export control alignment with Japan and the Netherlands remains, quote unquote, incomplete. But that's been one year since you said that. So Dutch and Japanese companies continue to sell chipmaking tools or provide service and repair services to more advanced chip factories in China that US companies are prohibited from selling to or working with. So, can you provide any updates on BIS's progress on aligning our export control on chipmaking equipment? And have you set any specific deadlines? And here's Kessler's response: Quote, Well, thank you for the question again. You know, I I what I can tell you, and I need to respect the confidentiality of diplomatic discussions that I have, I can tell you that there is cooperation with foreign governments on export controls, but I I unfortunately, unfortunately, I have to report that the cooperation remains incomplete. There's still more work to do, and that's a conversation. End quote. Not exactly the most inspiring of confidence response that I've ever heard. And I think that the right question is the one that Representative Kim asked, which is like, what have you been doing for the past year? Okay, this is a really, really big problem. And the crisis of not addressing this problem won't be felt today, but you are lighting the fuse on a bomb that's gonna go off a few years from now. And every day that you don't deal with it matters. And like this is not stuff that you can procrastinate on and there's no consequence. Every day that goes by, more of these machines, more of these spare parts are going to China. Every day that goes by, ASML is training more local technicians on how to repair these machines themselves. That's the that's one of the key points here is that a lot of the ASML repair workforce are local Chinese nationals who are probably not gonna leave the country if ASML says you're not allowed to repair DUV machines anymore. So they're probably just gonna quit and go work for Chinese companies. So cutting off access to the spare parts now is the strategic leverage we have now. And if we don't do anything now, then in not that long, we're not going to have the leverage that we want anymore. We won't be able to do anything, it'll be too late. And so I think it's shocking to me that BIS has been so unable to not just reach agreement with allies on updating these controls, but updating its own controls separate from what the Allies are doing. It's worth pointing out here that the White House said they wanted to do exactly this. In the White House AI Action Plan, which came out, I want to say, in July 2025, they specifically called for strengthening export controls on semiconductor manufacturing equipment. Congressman Mast, in that same congressional hearing, he referred to that in his opening statement. We are nowhere near implementing on that goal. Like, let me just read what the AI action plan said. Quote Plug loopholes in existing semiconductor manufacturing export controls. Semiconductors are among the most complex inventions ever conceived by man. America and its close allies hold near monopolies on many critical components and processes in the semiconductor manufacturing pipeline. We must continue to lead the world with pathbreaking research and new inventions in semiconductor manufacturing, but the United States must also prevent our adversaries from using our innovations to their own ends in ways that undermine our national security. This requires new measures to address gaps in semiconductor manufacturing export controls, coupled with enhanced enforcement. It's been a year. There's been no new measures, right? Like I don't know how many machines have been sold to China in that time, but definitely dozens. And also, apparently, if the information reporting is right, there's Japanese components that are directly supporting China's efforts to eliminate their foreign dependence on American companies, Japanese companies, and Dutch companies. So, like, what are we doing? Like, get on this ball.

SPEAKER_00 33:52

Yeah, uh, Greg, I think using a term like incomplete cooperation from allies is one of those phrases that really tests the poker face of the staffers sitting behind a principal when they're in a hearing. Uh, and is it's certainly not something that indicates a lot of forward movement from whatever uh diplomatic and coordinated uh strategy is being exercised.

SPEAKER_01 34:14

I know, and what's so mind-blowing is that, like, this is the Trump administration, right? Like, have they ever seemed unwilling to make tough requests of allies? No. And it's like, okay, let's focus on the things that really matter strategically for American national security, for American economic security. If you want to put pressure on allies, don't put pressure on them to like buy more random farm goods from some part of uh America's. Yeah, don't put pressure on them to buy like American random American farm goods. Like, really focus on the sectors where we can make a choice that will have strategically beneficial impacts for the next decade, for the next two decades. Like, if you're willing to twist American allies' arms, here's a great place to twist their arm because honestly, it's not just in our best interest, it's in the allies' best interest too.

SPEAKER_00 35:14

Correct. So um, let's talk about Congress just for a minute here. We kind of mentioned them off-handedly as far as the hearing is concerned. Um, but the MATCH Act is under consideration right now. How does this news affect the debate per the Match Act?

SPEAKER_01 35:31

Yeah, so this is directly relevant uh to the MATCH Act. The Senate version of the bill, and this is a bipartisan bill at this stage, um, covered semiconductor manufacturing equipment, and that included, according to their definition, quote, at a minimum, all deep ultraviolet immersion photolithography machines, end quote. And they're talking about not just export controls that would apply to America, but explicit use of the foreign direct product rule that would apply it to allies such as um Japan and the Netherlands. So the Match Act, if it's passed, we're not negotiating with the Dutch and Japanese anymore. We're applying the foreign direct product rule and saying pretty intensely, right, we're going to punish your companies through our financial system, through our criminal justice system, and by cutting off access to key things that you need to import from us technologically if you don't go along with this plan. And that is a bill that you should update us on the current status in Congress of.

SPEAKER_00 36:32

Yeah, yeah. So, Greg, the Match Act in the House side was considered as part of the NDAA. Uh, the NDAA process in the House has concluded. So that bill has been sent to conference. Uh, there are some challenges associated with the NDAA in the Senate version. So I know we are still trying to get to a conclusion of that process. There are amendments still being considered there. Uh frankly, we need to check on a as of right now basis to see if the MATCH Act was included in a manager's package there or was actively being considered. Uh, but it is something that is up for conversation and consideration as the NDAA moves to a conference process later this year. Uh certainly there are a couple of things in the way right now uh for that process, and we don't want to get too much further into the weeds on uh all of the But it's a big it's a big deal if it's in the NDAA. Uh if it's in the NDAA, uh it will pass, right?

SPEAKER_01 37:29

Right, because NDAA is must-pass, so being included on that.

SPEAKER_00 37:33

Yeah, the NDAA is one of a very small handful of bills that are considered must-pass bills any given year. It's a moving vehicle. So if you can get something, especially something into a bill that has moved out of a chamber and has passed out of a chamber into the conference process, there is an extremely high likelihood that that thing will pass unless it has to get shoehorned out of the conference process to slim it down in order to just expedite pass it.

SPEAKER_01 37:57

I got it.

SPEAKER_00 38:00

Last week we talked about the Chinese company Moonshot AI's Kimmy K3 model being very high performer that was making waves in the stock market. But at that point we really hadn't had official statements from US officials on the release of Kimi K3. Uh, now we have two big statements from Trump administration leaders. Let's start with Michael Kratzios' tweet on July 22nd, which said, quote, We have information that Moonshot AI distilled Anthropic's fable for the development of its K3 model. To do this, they developed a sophisticated internal platform to conduct large-scale distillation against US models, allowing them to quickly switch between multiple methods of access to avoid detection. Moonshot AI has also acquired GB300 equipped servers and has accessed G B300s in Thailand, likely to train its AI models. The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open source frameworks, and open weight models. Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem. However, large-scale covert industrial distillation aimed at stealing proprietary US technology and undermining American research is unacceptable. Then on the same day, Treasury Secretary Scott Bessent tweeted that we support open source AI and the innovation it unlocks. But open source is not open season on American IP. When PRC firms conduct covert industrial-scale distillation attacks that cross the line into IP theft, sanctions and entity list designations will be on the table. Greg, what do you make of these comments?

SPEAKER_01 39:50

Well, it's a big deal on both of these statements. So starting with what Kratios said, number one, this is again acknowledgement by the US government that chip smuggling is taking place into China, and specifically chip smuggling of the GB300s, the best chip that NVIDIA sells right now, the Blackwell generation. Trump himself said that he did not want blackwells getting into China. Here is Kratzios saying not only did the chips get into China, they were used by Moonshot to train this latest and greatest version of the model. The second thing here is that the government is drawing a line between open source AI and what they're saying is uh open source AI rooted in intellectual property theft. And I think this is a big deal, right? You can say there is a difference between coming up with a competing product and giving it away for free, right? Like let's just say that a company has a cool new software product and it's very similar to Microsoft Office and they give it away for free, but they developed it on their own, that's not a problem. The US government's not gonna fight you on that. But if you create a pirated version of Microsoft Word, Microsoft Excel, and you give it away for free on the internet, and it's just Microsoft Excel in a wig and a mustache, yeah, the US government is gonna say we have legal problems with what you're up to here. And so they are saying we support open source, we don't support the path that Kimi K3 has taken and that China writ large has taken in terms of getting very high quality open source. And I think that their argument is very, very credible. I think the performance gap between American open source AI models and Chinese open source AI models, I don't know if it's all of it, but I think a huge share of that gap is just coming from the fact that American companies are respecting intellectual property of the frontier model developer leaders, open AI and anthropic, and Chinese ones are not. And that is a huge problem for the United States government. And again, what Bessent is saying when PRC firms cross the line into IP theft, sanctions and entity list designations will be on the table. So that basically says we're not relying on the AI companies to come up with some kind of magical technological protection that makes distillation impossible. Or at least we're not solely going to rely on that. We are considering active punishment for the Chinese firms involved in executing this activity. And sanctioning Moonshot AI, putting them on the entity list, that's a big deal. That basically means that, for example, Google, which has recently made Kimi K3, I think, available through its cloud platform, if that's not up and running right now, it will be very soon. Um, they're not going to be able to do that anymore, very likely. And so, what share of global computation is American cloud hyperscalers? A huge, huge share. And if they're not allowed to make these models available, then they won't be all that available. That's a really big deal strategically. Uh, it's a big choice that the government has to make, and a lot of different people have opinions on it.

SPEAKER_00 43:26

So China has certainly not taken those statements laying down, right? So during a July 27th press conference, uh, the Ministry of Commerce spokesman said, quote, the US disregards such facts as the release dates of the AI models of the Chinese companies that are concerned are very close to those of US frontier models, and that some Chinese models are already lead in capability. Instead, it pins labels on Chinese firms, that they rely on distillation, that they steal US intellectual property. And further threaten suppression and sanctions. These practices lack any factual basis and have no foundation in law. In practice, they apply a double standard. They are a textbook instance of artificial intelligence hegemonism. It is our understanding that many US AI companies have distilled Chinese models in their own research, development, and training. It is worth noting that the numerous companies within U.S. industry have publicly opposed the U.S. threat to sanction Chinese AI firms. Nearly 200 U.S. startups have urged the US government not to cut off access to Chinese open source models, holding that doing so would weaken the competitiveness of U.S. companies. With regard to any conduct that substantively harms Chinese interests, China will take all necessary measures to resolutely safeguard its own legitimate and lawful rights and interests. Greg, what do you make of these statements?

SPEAKER_01 44:53

There's a lot in there. So the first is they're rebutting Kratos' claim that Kimi K3 is specifically a distillation of Fable V. And they're making an argument based on the time lag between the release of Fable V and the release of Kimi 3K3, basically saying there wasn't enough time to pull off such a distillation attack. To be perfectly honest, I don't know if that's specifically a clear argument. I know that there was a time in the history of AI when that probably would not have been enough time to pull off a distillation attack. But as Kratios is saying, China has set up this industrial scale distillation effort where they have tens of thousands or hundreds of thousands of accounts that are constantly asking questions, receiving answers, asking questions, receiving answers. And it's worth pointing out here that there's the base model training run, and then there's also fine tuning training runs. And I think the idea that they could not have executed a fine-tuning training run doesn't strike me as credible at all. And whether or not they could have done a base model training run with that updated data, I don't know. Like possibly maybe not. The other thing I would say is that Kratius is saying, like, we have information that. And one thing that I just don't know what to make of as I think about stuff the government says anymore, is that based on what we saw with OpenAI autonomously hacking Hugging Face, based on what we heard from Senator Warner about Mythos autonomously hacking the NSA, there's just a part of me that wonders like, is it the greatest time ever for cyber intelligence in the US intelligence community right now? Like the idea that we are deep in the internals of every Chinese AI company thanks to AI augmented hacking, um, I obviously have no personal knowledge of any of that kind of information. It would obviously all be classified. But it strikes me as at least plausible, right? That the quality of our intelligence on Chinese AI companies, on Chinese hyperscalers has meaningfully gotten way, way better in just the past handful of months. That's at least plausible to me. And so when Crazio says he has information, he doesn't tell us what that information is. Maybe it's like a palm reader told him that, or maybe it's like they really got deep into the internal servers of Moonshot. They looked under the hood and they know exactly what they found. I realize it's difficult for people out there because they say, you know, how do we know what to make of any of this information? But I at least want to make it clear, right, that it's possible that US intelligence is having a very good few months. Um and when we say stuff like this, it's going to be grounded incredible in a way that it wasn't uh before. That's one thing. The next thing is China always says the US is lying. Like when we said they were building islands and militarizing those islands, they said, no, we're not. And then we're like, here is the satellite photos of you building airfields on those islands, and here are the military airplanes on those islands, and they are yours. And they said, No, we're not. You're lying. I mean, just like China is saying that the United States is lying adds no information uh to the ecosystem, unfortunately. The last thing that is very important here is where they say that they will respond, they will take necessary measures to resolutely safeguard their own legitimate lawful rights and interests. I think what they're basically saying here is if you come after our AI companies with entity listings and sanctions, we will retaliate. And they don't say what that retaliation is, but I think, frankly, we've seen that the Trump administration is scared of Chinese retaliation. I think that's uh a pretty clear interpretation of the Trump administration's behavior since the rare earth metals dispute. And so, in the same way that Bessent is effectively saying, hey, we will protect our AI companies with cutting off access to our financial system, like sanctions, for example, China is saying we will protect our AI companies with other tools of our geopolitical power, which could include cutting off rare earths, it could include something like going after American companies in the Chinese antitrust system, which is something that they absolutely weaponize in the most vicious ways, and it really shapes the behavior of American companies and American CEOs.

SPEAKER_00 49:38

Yeah, right. Obviously, there's a full understanding within the Chinese government of not just the impacts of sanctions that China puts on the United States, but also uh the US election cycle. So normally the Chinese government keeps those sanctions and those retaliations somewhat closely matched to the subject matter we're talking about. Rare Earths are there. Uh they also well understand that we're getting towards November and sanctions on soybean farmers in middle America hits really hard and very, very loudly. So they they understand how to get everyone's attention when they really want to. It's just a matter of where they are in that process.

SPEAKER_01 50:16

Yeah, I mean, I remember reading at one point that China had a map of which were swing state elections, and they had identified key goods that were relevant to the economies of each of those swing states in the election. They absolutely think about this and game this out strategically.

unknown 50:32

Yeah.

SPEAKER_00 50:33

Uh so there was a mention within the remarks from the spokesman of the Ministry of Commerce specifically about U.S. industry and U.S. industry trade associations, but let's delve a little bit more into that. There certainly have been a number of different uh trade organizations that have been set up over the last couple of weeks in response to this policy conversation. How is industry reacting to this and to the statements coming both from the Trump administration and from the Chinese ministry?

SPEAKER_01 51:00

Yeah, so the letter that the Chinese spokesman was referring to was a group of 200 startups self-identifying as the little tech ecosystem. Um, probably the heaviest hitter in that group was Y Combinator, the incubator and venture capital firm. What's interesting is that that was the letter that says, don't ban Chinese open source models, very explicitly, very directly. And it didn't include really that many heavy hitters. Now, some people have echoed that statement without signing on to that letter. Mark Zuckerberg recently said publicly in a Wall Street Journal op-ed, don't ban Chinese models, said the same thing in an interview with the Financial Times. Um, Jensen Huang said in an interview with Axios on July 21st that, quote, these Chinese models are excellent, open source models that are excellent should be used. He also hinted at his financial incentive for this when he said that this would be good for NVIDIA because, quote, free AI should be great for hardware. In other words, that the cheaper the software is, the more people will want to buy his hardware, which I think is factually probably true. So there are powerful people who are saying don't ban Chinese AI models. But the letter that China referenced was not the most powerful letter that's out there right now. For sure, the most powerful letter that's out there right now is the one that was coordinated, I think, by NVIDIA originally. Literally Jensen Huang's first tweet ever was announcing this letter, and that included the signature of NVIDIA, of OpenAI, of Google, and that one said don't ban open source, but it did not say don't ban Chinese open source. And the reality is like there's not that many people out there who are arguing ban open source AI. I mean, like maybe the Eliezer Yudowski, like AI is gonna kill us all, and soon crowd would be in favor of banning open source AI. But like that's not what Anthropic is arguing for. And Anthropic, which did not sign on to the letter, felt strongly enough about it that they had to post a statement from Dario Amodei. And here's what he wrote that I think is worth quoting at length. This brings me to the open letter. I agree with much of it. Open weights expand access to the AI economy, they strengthen competition, at least for some use cases, and they give customers greater control. Concerns about distillation should be addressed through targeted legal and commercial frameworks, the same measure I described above, but I don't agree with the letter's assertions that open weight models necessarily make it easier to develop safeguards, or that broad access to capabilities necessarily helps defenders more than attackers. It seems at least as likely to me that the opposite will be true. For example, I worry that biology will have a strong attacker-defender asymmetry where sufficiently capable models may be able to quickly weaponize pandemic-level viruses with widely available materials. Whereas defense against these agents is a multi-year operational attack in the best case, as we saw with Operation Warp Speed. Then later he says, quote, to summarize my and Anthropic's position, we have not and are not advocating for a ban on open weight models as a category. We should instead focus on keeping powerful chips out of authoritarian hands, stopping industrial-scale distillation, and requiring safety testing of all sufficiently capable models open and closed. And I think it's that last point, right? Require safety testing of all sufficiently capable models that I think is the really important point here. Because saying that you're open source should not be a hall pass that allows you to engage in all kinds of otherwise illegal and unethical behavior, right? If a company said that they wanted to create an AI model for generating child pornography, you'd say, that's bad and illegal. And if they said, but it'll be open source, would that change your mind at all? Right? And similarly, if they said they want to create a model that's going to make anyone able of executing a nation-state level cyber attack or bioweapons attack, that would be bad and unethical. But if they said, but it'll be open source, would that change your mind at all? I think no. And so the debate is not about banning open source models, but it is about saying when they start having certain levels of capability, do the same regulations that apply to closed source models also apply to open source models? And I think if you want positive outcomes, if you want to prevent child pornography, if you want to prevent bioweapon scale disasters, I don't see how you can make the argument that giving a complete exemption to open source is going to allow you to achieve those goals. And it's definitely true that open source has played an important role in cyber defense and will continue to play an important role in cyber defense. But here I just have to say, I think Dario's correct that past certain capability thresholds, I think the regulations have to apply to everybody.

SPEAKER_00 56:13

Correct. It's not a matter of open or closed, it's a matter of what do you do with mythos under any of those circumstances. Well, Greg, that concludes today's episode. Uh thanks to all in the audience for listening. And Greg, thank you for walking us through another very busy week in AI policy.

SPEAKER_01 56:31

Thanks, Adam. That was great.