Have America’s chips controls backfired?
The Spectator magazine is home to wonderful writing, insightful analysis and unrivaled books and arts reviews. Subscribe today for just £12 and receive a 12-week subscription in print and online, along with a free £20 John Lewis or Waitrose voucher. Go to spectator.co.uk forward slash voucher.
Hello and welcome to Chinese Whispers with me, Cindy Yu. Every episode I’ll be talking to journalists, experts and long-time China watchers about the latest in Chinese politics, society and more. There’ll be a smattering of history to catch you up on the background knowledge and some context as well. How do the Chinese see these issues?
Beginning in the first Trump presidency and expanded under Joe Biden, the US has taken a strategy of technologically containing China through restricting its access to cutting-edge semiconductors. As Chinese Whispers has looked at before, these chips form the backbone of rapid advances in AI, telecoms, smartphones, weaponry and more. Washington’s aim was and is clear to widen the technological gap between the two powers. But has the strategy worked? Lately, this has become a hot topic of debate as Chinese tech companies such as Huawei and DeepSeq have nevertheless made technological strides. Some even argue that the export controls have spurred on Chinese innovation and self-reliance.
In this episode of Chinese Whispers, we’ll be debating this issue with two very informed, very smart guests. Ryan Fadasik is US Director of the Future Society, an independent non-profit organization focused on AI governance and former advisor for US-China bilateral affairs at the US State Department. Steve Hsu is Professor of Theoretical Physics at Michigan State University and a startup founder. He also hosts the podcast, Manifold. Steve and Ryan, thank you so much for joining Chinese Whispers.
Ryan, you were actually working in the State Department to help formulate some of these export controls. To start with, can you tell us what the US government saw as the goal of these controls and how exactly do they work? Sure. Well, I don’t want to speak on behalf of the US government and can only talk a bit about some of the ways I was thinking about controls. My origin story here actually began as an analyst a few years prior to that at a place called the Center for Security in Emerging Technology, where we were trying to think about how to control or govern inputs to AI development so that it might not be used by malicious actors against the United States.
Quickly, we found that the US export control system was never intended to deny resources to the world’s second largest defense industrial base. It was stood up in the 1980s as an instrument to put selective pressure on rogue states and companies involved in nuclear weapons proliferation or human rights abuses, with the promise of an off-ramp if they would comply with international norms and regulations. But during the Biden administration, and even earlier under the first Trump term, it became clear that there were certain strategic and emerging industries in the United States that warranted a closer look. Both administrations placed controls on a number of industries, particularly AI and the advanced semiconductors used to train and operate AI systems.
We’ve seen a number of rules beginning in October 2022 and updated each year since then, which were basically taken, I would say, for three reasons. The first reason is to prevent military modernization and the development of systems that would eventually be pointed toward the United States and its allies. The second was to prevent egregious cases of human rights abuse, surveillance, repression, and particularly surveillance systems installed in Xinjiang. The third reason was increasingly a response to unfair trade measures taken by the People’s Republic of China. I would say to protect U.S. economic competitiveness writ large. It’s this third reason that’s really been viewed as somewhat controversial.
For years, the United States and China have been trading non-tariff barriers to trade and beginning this week with tariffs as well. I think this is kind of the story of the evolution of these controls. The background for this is that the U.S. and its allies, it’s fair to say, dominate the most advanced level of the semiconductors that is needed for all of these areas that you mentioned in China. So it’s the thought that if you deny China access to these things, you can at least slow them down or hobble them.
What is the goal here? That’s right. Well, I would say the goal is kind of the three prongs I’ve just outlined, preventing military modernization, cracking down on human rights abuse and promoting American economic competitiveness. But to your point, yeah, I mean, chips are viewed as a core element in the supply chain and development of AI systems. It happens to be the case that the leading chip designer, NVIDIA, is a U.S. firm. The leading chip producers are firms in Taiwan and South Korea. The leading producers of the equipment used to produce those chips are manufactured primarily in the Netherlands as well as in Japan.
So with this kind of collective geopolitical arrangement, it became clear that it would be possible to shape and modulate the structure of the global chip supply chain. However, there’s been recent debate about whether or not these export controls really are working, not just in AI, but in Chinese technology in general, especially in light of what’s happened with DeepSeq over the last couple of months. Steve, I think you’re a skeptic of these export controls. So just tell us why.
I don’t have to be a skeptic. I can just be descriptive of what happened in the last few years. I think it’s a matter of record. If you actually go and look, you can see what actually happened. So we don’t have to speculate. Ryan made an interesting point about the regime under which these sanctions could be administered. Really, maybe we’re thinking of targets more like Iran, for example, or North Korea than China. The key thing is that if the entity that you’re imposing the sanctions on isn’t able to build the technology itself, then you can really dictate terms.
But what you risk when you sanction a country that is as or more technologically capable than the United States is that you risk them basically indigenizing the whole supply chain. Eventually, you end up crushing your own companies. You didn’t intend to do it, but you end up cutting off your own companies from huge markets that end up being dominated eventually by Chinese companies. The most succinct way I would describe it is that the U.S. government solved the coordination problem or the alignment problem between Chinese companies.
If you go back well before the sanctions, there was this China 2025 plan and something called the Big Fund, which is a big industrial policy government fund to build up Chinese capabilities in semiconductors. That was already something they tried, and it was largely a failure. You could ask, how do you know it was a failure? So you could just look at, to what extent did Chinese companies at different points in the semiconductor supply chain take market share away from Western firms? Your Western is always going to refer to Japan, Europe, the United States, Taiwan, South Korea. So Western just means ex-China more or less.
To what extent did these small companies in China take pieces of market share in the supply chain away from Western competitors? In the era of the first big fund, the answer is de minimis. They did not succeed. Whereas now they have succeeded quite a bit. The main question you have to understand is why did they not succeed earlier and why are they succeeding now? To understand that, you actually have to understand how technological innovation works. Unless you’ve worked at a startup that is introducing deep tech into an established industry, you may not be aware of these dynamics.
A small company can, if it’s well-capitalized by venture, hire really top engineers from top companies—people who know what they’re doing, who want to be multimillionaires, not just salary workers, people that are particularly agentic and ambitious. That’s what typically startups are made of. It is seldom the case that the startups are lagging the big companies in technology. Because you have such a good team, founders are usually very unusual individuals. Typically what’s happening is the startup has better technology than the incumbent, but it can’t get the customers to buy. So who is the customer here?
Let’s specifically talk about SMIC, which is the leading fab in China and is gradually getting to the point where it will compete directly against TSMC. In fact, I would say in some sense, SMIC, for people in the industry, is much more formidable right now than where Samsung is. If I’m the guy at SMIC and I have to achieve a certain yield and I have to compete against TSMC, the low-risk thing for me to do is to use the same suppliers that TSMC uses. So I’ll buy stuff from Tokyo Electron, KLA, Land Research, Applied Materials.
If a guy down the street from me in Beijing or Shanghai has better technology or equivalently good technology, I’m very unlikely to use that because this little startup could run out of money. It’s highly unstable. If a few of the founders leave or get hit by a bus, the whole thing can collapse. Why would I risk my multibillion-dollar process on that new technology? That’s a situation that every Silicon Valley startup faces.
Now, let’s imagine the U.S. government says, though, you know what? There’s a good chance we’re just going to cut you off completely from ASML and Tokyo Electron. There’s a possibility, there’s some tail risk now, you’re going to be cut off completely from these Western suppliers. The very first thing I’m going to do at SMIC is say, hey, my buddy down the street in Shanghai, we both went to Baida together. You know what? I’m going to start using your stuff.
There was a coalescence post these sanctions of effort and cooperation. Instead of competition, cutthroat, dog-eat-dog competition, you switch into a mode where, with some help from the central government, everybody is collaborating, everybody is cooperating. That accelerated the catch-up of the semiconductor ecosystem to the point now where I would say most people looking at the industry can confidently state that China is going to start dominating the mature and legacy part of the chip business. This would be roughly 28 nanometers and up. But that is most of the revenues. If you look at sales by how advanced the process is that makes the chip, those legacy and mature nodes are the dominant part of revenues even today in that industry. We’ve basically given it to them.
That’s my description of what happened in the last few years. I think it was a disaster. As a shareholder of applied materials and some of these companies, I just feel that the U.S. government’s actions, although perhaps well-intentioned, have produced complex real-world consequences. This is why we constantly have debates in America about whether industrial policy is good or bad. People in the industry who have worked for twenty years in a particular field will fight about whether government should or shouldn’t intervene. It’s often a function of how complicated the industry dynamics are. I just don’t believe Washington is the best place for crafting these kinds of strategies.
Right, Steve. I’ve just come back from China recently, where I was in Shenzhen for a few days, talking to some people at Huawei. One of the senior executives there basically told me a similar thing: Thank God for the American sanctions. Now we have all these Chinese suppliers. They’ve basically pushed us into doing that. He was a bit smug about it, I have to say. So, Ryan, what do you think about that argument? Is it just cope on the side of these Chinese tech companies, or is there some truth in that?
To be honest with you, I do think a lot of it is cope. I’ll give you some reasons as to why. First of all, I want to step back and just take a look at the framework I laid out about why we have export controls at all. The goal is to inhibit or slow military modernization, human rights abuses, and unfair economic practices. And within that framework, I do think that U.S. controls up to this point have been moderately successful in slowing Chinese companies’ progress in developing AI systems or at least imposing progressively larger marginal costs for their development.
DeepSeq’s own founder, Liang Wenfeng, reportedly told the Chinese premier earlier this month that U.S. restrictions on the export of advanced chips to China were still a huge bottleneck for the company. A few weeks ago on this podcast, Ava Do pointed out the costs borne by Huawei after it was added, for example, to the U.S. entity list. Ultimately, DeepSeq succeeded in producing a sophisticated technical breakthrough. They did it by relying on a large stockpile of NVIDIA’s A100 chips, which they acquired before U.S. export controls were ever imposed late in 2022.
The fact is that the United States has pursued this strategy of controlling AI chips specifically to stop this kind of thing from happening in the future. If they had imposed these controls earlier, it’s possible that it would have taken DeepSeq much longer to develop capabilities that have seized the world’s attention in January. However, there’s also been a lot of confusion around the significance of the company’s technical breakthrough, and whether that might diminish the importance of computing power as a core driver of AI development. I’m happy to talk a bit more about that.
The bottom line is that leading AI labs in China, like Baidu or Alibaba’s Gwen, are facing similar constraints as DeepSeq’s today, struggling to purchase the cutting-edge chips that are available on the consumer market. It’s important that people understand that export controls were never meant to be a silver bullet to stop or inhibit the progress of Chinese companies altogether or to prevent the emergence of an AI industry in China. But many of these companies have reported that their progress has faced significant delays. They have communicated this directly to Chinese leaders, reporting months to years of lead time, by some estimates, that U.S. companies have been able to achieve due to these controls.
One common argument made against the efficacy of these controls, and one that aligns with Steve’s claims, is that they’ve somehow sped up or prompted the Chinese government to expedite the indigenization of chip supply chains. However, I would argue that this process was already well underway, obvious since the 20th Party Congress, but even years before. Greg Allen at the Center for Strategic and International Studies has documented this in great detail, outlining a number of policy documents, one of which Steve mentioned—the 2015 roadmap of major technical domains required for implementing China’s Made in China 2025 plan—which called for substantial indigenization of mid-level process node equipment.
In 2019, one of China’s leading advanced chip memory manufacturers, YMTC, launched a full-scale de-Americanization campaign. It targeted the entire workforce structure while relying on over 800 staff. With this full court press to remove American technology from Chinese supply chains, it was only a matter of time before industry leaders at a company like SMIC began cooperating and consolidating their supply chains, ultimately resulting in larger and larger benefits for their industries.
While I acknowledge that this has indeed occurred, as someone who carefully considered whether and how to impose export controls on this industry prior to their implementation, I can confidently say that the probability of Chinese indigenization accelerating was already internally recognized. The focus was always on imposing some costs on the development of AI systems that could be misused for military modernization, human rights abuses, or unfair trade practices.
Alright, Ryan made multiple points, and I think the critical issue here is that people in government have a tendency to overestimate their level of control or impact on a dynamic market system. The capability of Beijing to compel their companies to indigenize is not as robust as many in Washington believe.
In fact, it was really the threat of being cut off that materialized for these companies, prompting them to adjust their behavior. This is something that many who haven’t been involved in technology development may not truly appreciate. It’s too easy for bureaucrats in Washington or Beijing to abstract away the complexities and assume that if they issue directives, compliance will naturally follow.
Shifting back to AI, I would correct some of the statements regarding the current situation with NVIDIA chips and DeepSeq. First, China currently has a surplus of NVIDIA GPUs. This situation exemplifies the disconnect between Washington’s perspective on what regulatory measures accomplish and the actual developments on the ground. Last year, NVIDIA’s sales indicated that half of those sales went to the broader Sino sphere, including significant purchases by Taiwan, Singapore, and China itself.
So, even if the U.S. aggregates all sales across its tech giants like Meta, Google, and Microsoft, a notable portion of those GPUs undoubtedly found their way into China. There have undoubtedly been instances of smuggling from nearby regions into China, contributing to this surplus.
While it was an issue for DeepSeq for a time, it’s important to note that many Chinese government data centers and major tech firms, much larger than DeepSeq, apparently are not facing significant shortages of NVIDIA GPUs at present. If you evaluate the street price of GPUs in China, they are currently quite low—lower than they are in the U.S. This price drop indicates a glut resulting from the sanctions not functioning as intended by the U.S. government. So again, I’m just emphasizing the complexity of the real world versus the intentions of policymakers. But if I could jump in, I mean, I would push back on a couple of points just now.
First of all, it might be the case that there’s a temporary glut created by stockpiling of AI developers in China of chips that they knew would likely be restricted by future US controls. And that’s exactly what happened. So the temporary street price of those GPUs has diminished. But just to clarify, I mean, the thinking very much at the time of the policy debate in Washington was such that we understand that today’s chips are available and will be available in China. The question is about setting some kind of line in the sand after which future developments in chip technology will become harder and harder for leading labs to access. And that’s exactly what happened.
If you look at NVIDIA’s H100 chip, you’re saying it’s widely available and relatively inexpensive. The newer B200 could have achieved those advances in training or inference a lot easier and a lot faster than what would have otherwise been available without the controls. And so I think that, like, the purpose of enacting any kind of line or drawing some kind of control around some piece of technology is the idea that, okay, after we reach this point, you know, future developments in the technology will be better controlled.
I’m optimistic that we’re just starting two years later to see the initial result of some of the original restrictions imposed in late 2022. Right. So the future impact of B200, whether they’ll be able to get them through smuggling, whether they won’t, how much that will affect further AI progress, we don’t know the answer to that. So that could go either way.
I just think people who are thinking about this and are a little bit too focused on the AI part of it are forgetting that you may have basically given the chip industry, or at least the majority of revenues in the chip industry, which is mature and legacy nodes, to the other side for this speculative gain that post-2025, future model training in China might be negatively impacted by the sanctions, right? I mean, there are trade-offs here, and I just want to be very clear about what has been traded away.
If China gets a dominant position in mature and legacy chips and decides to restrict US access to those chips, that affects the whole economy. That doesn’t just affect a few labs trying to build AGI. It affects the whole US economy. So you can have a huge debate about just whether that trade was a good trade.
Two points. One is, I think your overall framing is generally right, that this is the trade-off. It’s the question of how we’re affecting the overall chip industry today, borrowing against future expected improvements in AI systems. On that point, I would say that with the hindsight of the meteoric improvements in LLMs seen over the past two years, the gamble was absolutely the right one to make.
The second point I would just add very quickly is that I do think it’s worth pushing back and thinking a bit more about the global chip supply chain. And a bit of hubris to think that US controls on leading-edge chips, not the vast majority of legacy nodes, would have so massively shifted the environment in China. We’re already talking about a political and economic environment where the Chinese state was demanding domestic production, dual circulation, ripping out American components and hardware.
And so I just really struggle, as one of the bureaucrats Steve is talking about in DC, but also as a technology analyst and a political analyst, to understand and really believe that the change in US export control at the frontier side of semiconductors in 2022 motivated a shift in the supply chain that was already happening. You could look at studies published by semi-analysis and others looking at the composition of legacy node supply chains over the past 20 years. And it’s very clear that the China model that has happened in every other industry was and continues to happen in legacy chips.
I don’t believe for a second that that’s been motivated by the US decision to impose export controls on advanced chips. Yeah, I think we should agree to disagree on that. If you ask the actual people involved, I think they’ll tell you that the US sanctions did trigger an alignment of incentives and collaboration among Chinese companies, which didn’t exist before. And I think if you haven’t been in business, you just don’t understand how hard companies compete and how individuals respond to incentives.
So if I’m running SMIC, I’m going to use the best stuff I can. I don’t want to lose my career. I don’t want to have a failed production run. You know, even if I literally went to college with the other guy, I’m not going to use the stuff from his startup unless I really have to. And it was risk analysis based on de-sanctions that caused the shift. And I think we should just agree to disagree on that. I mean, ask people in industry—not analysts who analyze technology, but people who actually run companies, and you’ll get very divergent analyses of what happened.
Steve, when it comes to, let’s say, those Chinese alternatives then, I feel like part of the debate is underpinned by how quickly China can catch up. If indeed it can catch up, you know, America currently has the lead. So another topic we haven’t gotten into is how fast they’re going to catch up in EUV. We’re talking about the frequency of light used for lithography. So that’s actually the way that they actually create the features on the chip.
And you’d like to make those features as small as possible. So you need high frequency or a short wavelength light. It’s incredibly challenging to produce extreme ultraviolet light to make the smallest features on chips. And that is, you know, maybe the biggest remaining bottleneck is that currently the EUV lithography products are only made by one company, ASML, in the Netherlands. And so the open question, which I think nobody really knows the answer, and even when I was in China, like people wouldn’t really say very much about where China is on EUV, but there’s a huge range of speculation about where they are.
But anyway, so we don’t know the answer, like how long will it take them to catch up in EUV? And of course, whether this whole thing in hindsight, when we look historically back on this strategy by the U.S., whether that turns out to be a success or failure, depends a lot on what exactly happens in the future with EUV.
No, I don’t have much to add on EUV, and I think that Steve’s point on that is actually basically right. But I did want to turn back to our earlier conversation and just one final point. And it’s worth pointing out that in 2019, we heard the exact same argument from the CEO of Huawei, that, you know, look, we’ve stockpiled all these chips. American sanctions are not going to have any effect, and they’re really not even worth imposing to begin with. It’s only going to backfire and harm the American economy and the innovation of American semiconductor firms.
Well, guess what? In May 2019, the Trump administration added that company to the end of the list. And for the next six straight quarters, the company saw declines in revenue amounting to billions of dollars and lost 13% of its profit. And so the export controls worked. They worked. They harmed the company. They slowed its progress. They led to later releases of phones. Huawei’s access to the international market has never been the same.
And I personally do think that we’re going to see similar effects take place in the chip market. You know, there are reasons to believe that as a result of its reliance on domestic Chinese components, SMIC is still going to be a few years behind leading chip manufacturers like Taiwan Semiconductor Manufacturing Company in terms of process node. The 5 nanometer line of chips, the most advanced that SMIC is working on, has been plagued by low yields and high production costs.
It’s likely going to be able to use alternative methods that are replacements for the EUV that Steve is talking about because they don’t have access to it right now. And I’m skeptical that its line of chips, the Kirin 9020, is going to be as competitive on the global market or as efficient, at least, at the frontier end of what we need to train AI systems used in military modernization and human rights abuse as those that are produced by other companies.
And so that is the motive, that is the line of thinking that has underpinned some of the decisions to impose these controls. And I do think, just to reiterate a point I had made earlier, it’s really too early to tell on a lot of this stuff. We’re producing supply chains today that will become relevant to training and operating AI systems that don’t yet exist and potentially really sophisticated capabilities that could come to pass in the next two years.
And so at that point, I think it’s worth looking back and auditing the effectiveness of these controls. Yeah, so now we’re discussing timescales, which I think is the most important thing about this. So those issues like military modernization or the ability to, say, suppress dissent in Xinjiang or something like this, right now there’s very little LLM impact on any of that.
So the stuff that’s a really glamorous AI development, Gen AI stuff that people talk about, has so far not had any impact on military stuff or ability to suppress dissent, human rights, et cetera. It will take a while for it to get there. So that was the goal, right? That’s what Ryan wanted to accomplish, to prevent the new developments in Gen AI from influencing those other areas to slow.
Or at least slow down, yeah. Or slow, okay. But in any case, it hasn’t happened yet. So let’s suppose it takes a little bit of time for that to happen. And by the time that happens, they’ve got EUV. In that case, you didn’t really accomplish your goal. And I would say you sped up their catching up in an extremely strategic set of technologies.
So Steve, you’re saying that they kind of went a bit too early, that this tool is… It went in too early. It went in too early, and they scared the other side. And the other side realized, oh, my God, these guys could cut us off and destroy the whole company. So I guess we just better go all in now. And so that’s what happened. I think they went too early.
And I think they went too early also, again, like I’m an AI founder. So one of my companies builds applications of Gen AI. So I know exactly how fast it’s being adopted in the real world. And it’s still de minimis. Go and read the quarterly reports of all the biggest companies in the world. And in those quarterly reports, the CEOs will report how much Gen AI is contributing to their bottom line. It is still de minimis.
Open AI only made $4 billion last year of revenue. That $4 billion is all the leading-edge revenue that they got from running those models. That’s all they got. So the multiple of how much it impacted the world can’t be that much bigger, okay? And this is a fact. So it’s just taking longer for Gen AI to have this magical impact on missile targeting or whatever it is you think it’s going to do. It’s taking longer than you think.
And in that meantime, they may catch up with EUV, and the whole thing will have been a mistake. You know, I think that the fundamental problem facing the United States around 2019 to 2021 was, okay, we can see China’s technology supply chains and industries indigenizing. We can see them attempt to rip American tech out of their supply chains. And we can see the fear created around some of the initial actions against ZTE and Huawei, but still some of the underlying policy developments and pronouncements, document number eight, made in China 2025, and so many other promulgations from the top that made clear where China was at in terms of its technology supply chains and industries.
And so with this on the horizon and an understanding and expectation that SMIC and other major Chinese companies were racing toward new ways to fabricate leading-edge chips and legacy chips and to try to surpass the United States and sell to a global market, the question became, well, what should we do about it? And the answer was to slow them. The answer was to try to buy some time.
As I said at the beginning, the understanding around the U.S. export control system is that it is not set up or equipped to crack down on and successfully prevent the transfer of every computer chip flowing into an entire country, let alone the largest country in the world. And we understand that China is awash in these kinds of capabilities and chips.
Within that framework, though, others in the space, Dario Amadei, the Anthropics CEO, have pointed out a couple of things. Okay, first of all, new developments are happening in the field of AI every day. Companies are, I don’t know if it’s fair to say racing, but they are all rapidly unveiling new capabilities in transforming their transformer architectures, new ways of tweaking those architectures to run more efficiently on underlying hardware so that they are trained faster, better, cheaper.
That’s exactly what DeepSeq was able to accomplish last month. But it’s something that keeps happening, that almost every company has its time in the limelight. So it’s not something that I think we can reasonably try to stop or shape entirely. But the point of a control in this sense and focusing on computational power as a key driver of AI development is that these chips are physical components. They’re countable. You can watch them fall off the back of a truck conceivably.
And so even if there are holes in the system, if you wanted to constrain the availability of resources in what would essentially be a temporary race or try to achieve certain capabilities and benchmarks on a shorter or longer time frame, then chips are naturally the thing you would first try to look at to impose any kind of structure or control around. And that’s exactly what ended up happening.
I think, you know, just a final kind of speculative point is that the Biden administration’s project to augment inputs to and development of AI technology was very much a situation of building a plane while it’s being flown. OK, it was very unclear where the trajectory of this technology would lie. There were many skeptics around 2020 about whether scaling laws would hold, whether today’s AI systems would be as capable as they are in 2025. In fact, people doubted that they would be so capable in 2035 or 2050.
And so the fact that we have got here, the trajectory of the industry has been such that it really is a world-changing technology, I think is a huge credit to the Trump and Biden administrations for understanding the salience of computational power as a commodity that can conceivably be controlled and tracked in the 21st century.
Steve, just to take you back to one of the points that Ryan started with by saying that even Liang Wenfeng, the founder of DeepSeek, has said, you know, this is our biggest bottleneck now. I guess the validity of your argument or the strength of your argument, rather, lies in balancing two impacts. On the one hand, the impact of not having as many advanced chips as they would like, not having as many as they can afford.
And then on the other hand, you know, the impact of solving this coordination problem, as you call it. How do you balance those up? Because it feels to me that the second bit has to be of greater benefit to the Chinese than the first bit is a cost to the Chinese. But it’s not clear to me that that calculus is obvious.
Yeah, I want to clarify a few just factual aspects of what you just said. So DeepSeek, as everybody knows by now, started as a quant fund. And they were not actually regarded in China as one of the leading LLM companies, Gen AI companies. It’s only recently that the country woke up and DeepSeek is one of the leading companies. The limited compute resources that DeepSeek had was for that reason, because they were spending their own money.
They were spending their own money from revenues from quant trading on building this up. And it was actually initially a kind of side project. It’s not fair to describe it as a side project now, but it really was initially a side project. And on the other hand, the ones who were considered the national champions, like ByteDance, Tencent, Alibaba, what has gotten unnoticed in the U.S., most people are not aware of this, within a week or two of DeepSeek R1 coming out, state-of-the-art models came out from Alibaba.
So Quinn Max, ByteDance, shipped Doubao 1.5, another startup called Moonshot shipped a model called Kimi that competes with O1 on the reasoning benchmarks. So people just have a very, like, incomplete understanding of what is going on in China. And those companies were not compute limited, okay? So you have a special case of a group of very, very smart guys, very idealistic founders. They were resource limited. That caused them to do all kinds of really clever engineering, et cetera, et cetera, that make their models so efficient.
But at the same time, the gap between DeepSeek and incumbent companies, which are, you know, basically control monopolies, are similar to the analog of Google or Meta or Microsoft. They didn’t have a compute problem. When you say they’re not compute limited, they were limited in the sense that if they want NVIDIA’s most advanced chips now, they couldn’t get it legally. That is a limitation. They might not be close to that limitation. They couldn’t get it legally, but they don’t have a shortage of compute.
But this is exactly my point. Like, they have access to the chips illegally. There’s been a glut of them because for years, Chinese companies saw the writing on the wall and bought lots of them, so the price is temporarily deflated. But as the years go by, and as they become harder and harder to find, the price is going to improve or increase. I’m just looking backwards in time. They didn’t pre-buy, they literally got smuggled ones that were sent to Taiwan and to Singapore.
But Steve, isn’t that an argument for enforcing the export controls better rather than not having export controls at all? That’s a whole, yes. So we could have a separate argument about if you’re going to have export controls, maybe do it competently. And, you know, I don’t want to come off as an anti-government guy because, for example, when I tell people the story of ASML, which I think people don’t appreciate, ASML is a product of 20-plus years of U.S. government industrial policy.
It’s really U.S. government industrial policy that is responsible for the existence of ASML. Good story, though, because when the technology was first invented at the U.S. national labs and by U.S. companies, they could not commercialize it. And they later sold it off for almost nothing to Philips, and that eventually became ASML. And the Europeans threw in a bunch of money to support that development. So I’m not against government. I’m generally for industrial policy. I think industrial policy can work.
But generally, when people in bureaucratic offices try to shape the world, they have a difficult time understanding how complex the world is. And the smuggling is just an example of that. So they just did not succeed in limiting the supply of H-100s in China. I think that’s literally what happened. I mean, I’d love to push back a little bit and just kind of defend my own record here.
I mean, my kind of – the crown achievement of my career, you know, up to this point has been looking record by record, you know, at hundreds of thousands of purchasing records made by the Chinese military. And trying to investigate the question of how it is that PLA units and state-owned defense enterprises were getting their hands on NVIDIA-designed chips before and after the controls went into place. And the answer is that there are huge gaps in our export control system. It was not set up to deal with this problem. It’s something I’ve written about constantly, and I’m happy to make a couple of points on this podcast today.
The fact is that today, as of the time of recording, there are 2,077 Chinese entities on the entity list subject to U.S. export controls. Seventy-five percent of those were added in the past four years. Okay, but at the time I started looking at this problem in 2021, it became clear that Chinese military units were successfully ordering AI chips by going through the kind of smuggling channels that, Steve, you’re talking about. The way they were getting those hands on their chips is very difficult to control.
And the fact is that the current way we add companies to the entity list is very strenuous. It takes a lot of work to get added to the entity list. And so the result is that some Chinese suppliers have made an entire business out of sourcing foreign data and components, NVIDIA design chips, and reselling them to sanctioned targets in China. That is very difficult to prevent. It is an endless game of whack-a-mole.
But I would just say, you know, I want to make a couple of points. The game is worth playing. Okay, enforcement is getting better, especially as government agencies in the United States, like the Bureau of Industry and Security, get more resources, funding, and access to better technology. That’s why Howard Lutnick, the nominee for the U.S. Commerce Secretary, made exactly that point during his confirmation hearing.
Second, though, is that even if some of these chips are getting through, contraband goods carry a premium on the black market. You know, it becomes more and more expensive as the years go on for a constrained resource to become available to maligned actors. They do pay a price. And so imposing costs to augment that behavior is the fundamental task of government. You might think that it’s difficult or it’s counterproductive or it’s not working so well. And I would agree with you there, but I think that, you know, it’s worth trying to reform that system and think of ways to do it better rather than throwing up our hands and saying, let them buy whatever they want.
My point was not that because enforcement is hard, we shouldn’t try to do these things. I just pointed that out as the world being much more complex than, and I’m not talking about you personally, but the way it might look to Mr. Lutnick sitting at, you know, top of a big government bureaucracy. I mean, he might make a decision and not be aware of how complex this is really going to be.
And I guess it is worth making the point that, you know, NVIDIA shareholders wanted those GPUs to eventually find their way to China because otherwise their sales would have been literally half in 2024 if the sanctions had really worked. All those chips that fled through Singapore and Taiwan, it’s comparable to the U.S. sales. I just want to keep saying that it’s comparable to U.S. sales of NVIDIA GPUs that went into the sinus sphere. And so, you know, it’s much more complex than some people might think. I’m not saying Brian doesn’t understand the complexity, but a lot of people, I think, do not.
Steve, I take that point very well, and that, you know, it’s not a blunt tool. But just going back to that question of just weighing up, because I think what you’re saying is counterintuitive. You know, to listeners who are politically minded, they think, OK, government makes something harder. It’s surely better that it’s done than not. But what you’re saying is that it’s actually counterproductive, that the benefit it gives China is more than otherwise.
The real counterintuitive aspect to this is to understand how difficult it is to control a dynamic market system. So it’s very easy to do some first-order reasoning like we’ll do X and the consequence will be Y. But, you know, all those players have a vote, including the people who run NVIDIA. They have a vote, OK? All the data centers in Singapore who maybe can make more money by just reselling the GPU back into China or data centers in Taiwan that can do that. Everybody has a vote.
And so it’s just, again, I’m not a libertarian. I’m not a free-market absolutist. But they have a point when they say, hey, government guy, you guys don’t know what you’re doing. When you reach in, you’re just as likely to mess it up as get the result that you want. And a lot of people in Silicon Valley and in Shenzhen and in Shenzhou where TSMC is headquartered, they all believe that. You’re going to find that opinion in their brain.
I’m really mindful that we’re starting to run out of time. And so I want to kind of just move on to what’s been happening in the last couple of weeks, really. Speaking of, you know, when you reach in as a politician and you mess things up. Ryan, as we’re speaking, you know, in the last few days, Donald Trump has put in these tariffs on China. China has retaliated.
But more importantly, in the past, he has pledged tariffs on Taiwan, on Taiwanese chips as well, because he’s setting out America first. He’s not going West first. He’s America first. So do you worry that Trump’s kind of explosive approach to foreign policy might actually damage some of the unity that, for example, under the Biden administration, the West was trying to find when it comes to tech?
It’s a good question. And I want to just start and say, I’m not sure. I’m not working in the U.S. government. I left my old role at the State Department in November. But I wanted to, you know, as an outside observer, I’ve had some similar questions. And, you know, to be frank with you, I think arranging, you know, something as delicate as a control on an industry as large as AI or the semiconductor industry takes a lot of political capital. And that’s not easy to accomplish.
It’s also not an easy conversation to have with a U.S. partner like South Korea or Taiwan. But I do think that, you know, the president’s been clear that he wants to use tariffs to extract concessions to the benefit of the United States. I think, frankly, he’s been very successful at doing that over the past couple of weeks with Canada and with Mexico. And so it would make sense to look at a commodity that’s extraordinarily valuable for which demand is basically fixed.
Everybody needs these really advanced semiconductors, whether they’re training a frontier AI system or running a highly computationally dependent task in a data center. And so if you want to extract revenue, I guess look at chips. And what about aside from export controls? Because, you know, if that is the, you know, in the U.S. government’s narrative, that’s the way to kind of widen that gap between the U.S. and China. Surely there are other things that the U.S. should be doing in terms of strengthening its own AI industry, Steve.
I mean, one thing that comes to mind is personnel, for example, keeping AI researchers in the U.S. The number one thing the U.S. can do to compete with China is to make the U.S. stronger. That is it. It’s past the time where we can really do much to them. Okay. Look at shipbuilding. Look at EVs. Look at photovoltaics. Look at batteries. Look at, you know, there’s so many industries now where there’s literally nothing we can do to them.
Semiconductors is one of those that’s still kind of on the edge. And my opinion is we messed it up. But what we should be doing in the U.S. is trying to train as many brilliant scientists and entrepreneurs and innovators as possible, attracting as many of those brains, the really talented ones, not the ones who are currently coming on H-1B visas, but the really talented ones, we should be trying to suck into the U.S.
Including the Chinese? Yeah, actually, it’s just at a tipping point now where if you’re a very talented kid who could really make a significant contribution to Gen.A.I. progress, it’s kind of 50-50 whether you would, even if you got an offer at Open.A.I., whether you would come here. A few years ago, it would be definitely, yeah, I would go to NVIDIA or Open.A.I. if I got the offer. Now it’s like maybe, maybe not.
But still, the U.S. should do its best to brain drain those people out of the country. In a way, we’re kind of doing the opposite because, again, I’m not putting this on a high end, but the same kind of what I consider juvenile red scare thing that led to maybe what they did also has prevented many, many talented Chinese students from being given visas even to pursue PhD programs in the United States. Like we just shot ourselves in one foot, I would say, with the semiconductor stuff. We shot ourselves in the other foot by limiting the amount of brainpower that we could brain drain out of China the last few years.
And it’s literally just poor thinking on the part of our leadership in D.C. That’s my view. Ryan, keen to hear your view on this. It’s a really tricky political question. I think it is. But, you know, actually, this is an area where Steve and I are in violent agreement. You know, I think the most effective response to China’s talent recruitment plans and aggressive efforts to poach the rest of the world’s talent pool is one that relies on streamlining visas for foreign scientists, expanding opportunities for recent PhD graduates to stay in the United States, and basically creating experts or creating opportunities for experts who might otherwise be attracted to go work in China.
I also think it is important that the United States address the push factors that make China an appealing option to begin with. Although the United States isn’t, you know, in danger of losing its immigration advantage outright, it’s not exactly rocket science to figure out why some Chinese-American experts are choosing to leave, given the shameful academic job market, some of the hustle and work environment that they have to navigate, and the mercurial legal landscape that is required to stay in the United States after graduating from a top-tier program.
And so I think fixing those issues is still top of mind. It is probably an endemic advantage the United States holds over China in a long-term technology competition. And so hopefully it’s one that we’ll continue to recognize as the years go on. I’m glad that we found agreement. And just to kind of wrap up this conversation, I want to take a step back. Are we basically saying that the two superpowers in the world are engaged in this AI race?
I mean, Steve, you and I first met at an AI war game hosted by the British tech company Faculty AI, where we were on the PRC team strategizing what China should strategically do to win this race. But, you know, outside of the war game, it’s not exactly a reassuring thing to think about that the two major superpowers are racing for AI, considering it’s a technology that has even more existential risks than nuclear. Ryan?
I guess I would draw a distinction, you know, descriptively of what U.S. policy has sought to achieve over the past five, six years. Unabashedly, the objective, as articulated even by National Security Advisor Jake Sullivan, has been to make the most of this decisive decade, to maximize a temporary technological lead in the capabilities of frontier AI models, and in the capabilities afforded by the most advanced computer chips currently available on the market.
And I do think that the United States has adopted and pursued policies that have achieved some moderate success in achieving that objective. Whether it’s a good idea is an entirely separate question, and it’s one where I’m inclined to agree with you, Cindy. This is a world-changing technology with real risks. We’re talking about the capacity to unleash AI accidents that could put people in real danger.
It’s not simply a matter of technical malfunction or misuse in a military environment. It is no longer purely a question of misusing advanced software for military advantage, human rights abuse, or economic malfeasance. We’re talking about a potentially catastrophic episode that could arise as a result of progressively more capable digital life forms that are smarter than the human beings who are creating them.
And so it’s that understanding that I think people in Washington and I hope in Beijing are also waking up to that has motivated the outgoing National Security Advisor to warn about the risk of an AI catastrophe. I don’t think this is just a talking point, and I certainly don’t think that AI should go the way of climate change as being some kind of esoteric issue to bargain with or to borrow against the future on. I don’t think either country simply has the time available to it to make it that kind of luxury.
I do think it’s much more appropriate to think about what kind of international safeguards both countries can put in place or sign on to try to mitigate some of the potentially worst-case scenarios. Steve, do you feel optimistic about that? Well, you know, you referenced this war game that you and I participated in, and my counterpart in the war game, the U.S. leader, was Stanley McChrystal, a former general.
And that war game ended on an optimistic note because, so this is being made into a very major documentary film, so I hope I’m not giving too much of a spoiler. But the key issue was, could the two sides trust each other enough, in a similar manner as, say, with nuclear weapons, to allow inspections and prevent the worst kind of maximalist race in AI, as you might have had with nuclear weapons?
And, okay, I won’t say the ending, but yes, I’m optimistic that that could happen. So both sides could say, hey, you know what? We don’t want to make this into a race. This is going to benefit humanity. We’ll implement transparency in our labs. You guys implement transparency in your labs. We’ll have inspections, things like this. It’s possible that could happen. Like, if I had to bet money, would I bet that’s going to happen? Maybe not. Maybe not.
I do want to say that, like, again—how many NVIDIA chips would you put on that bet? Yeah, I don’t know. But, again, like, I don’t want to come off seeming like I’m a government basher because, look, I’m for industrial policy. I think the U.S. needs industrial policy to keep pace with China. But I would say most of the progress on Gen.AI in the U.S. has literally nothing to do with government policy.
Most of the time, it was just mainly being afraid that government was going to try to overregulate it the way they do in the EU. And it’s mainly due to the commitment of individuals who are running our big tech monopolies. Their core interest and understanding of the potential of AGI and their willingness to fund it at extraordinary levels, approaching, you know, 1% of U.S. GDP in hyperscaling and stuff like that.
So I think it’s really a completely private sector story so far in the U.S. And we just luckily avoided government crippling, you know, AGI progress so far. But the gap between the U.S. and China right now is very thin. I mean, it’s incredibly slim right now. And I wouldn’t actually, despite this issue with Blackwells and stuff like this, the next generation of NVIDIA chips, I wouldn’t say with high confidence the U.S. is going to keep a lead.
There’s too much brainpower in China, and they have more than enough compute. They lag the U.S. in compute, but the extent to which the NVIDIA advantage actually matters for this next stage of development of reasoning models is not really understood by most people. Most of reasoning progress looks like inference compute. It doesn’t actually look like gradient descent compute.
And so the NVIDIA advantage is much, much less. So the Chinese have these Huawei Ascend chips, which they make on a 7 nanometer process. And those seem to be just as good for inference compute as what NVIDIA has. So it’s a very fluid dynamic situation. It’s very hard to know what’s going to happen.
I’m going to stop this conversation now before we run the risk of turning into an AI podcast. Steve and Ryan, thank you so much for joining Chinese Whispers. Thanks so much for having us. It was great. Thank you for listening to this episode of Chinese Whispers. I hope you enjoyed it. If you’re listening to this podcast on the Best of the Spectator channel, remember that Chinese Whispers has its own channel as well.
If you just search Chinese Whispers wherever you get your podcast from, you will always get the latest episode first there. If you have any feedback, positive or negative, but preferably constructive, please do email me at podcast at spectator.co.uk. And I’d also love it if you left a review or told your family and friends about the podcast. It’s the way to help us grow. So thanks so much for listening.
Join us again next time.