When AI is interposed in U.S.-China relations, the discussion turns to rivalry, surveillance, and a race we are unsure whether we can win. Yet the way AI is developing in the U.S. and China is very different and not necessarily at odds with one another. The conversation also comes at a time of heightened fear of interdependence around rare earths and critical minerals, and supply chains that hinge on the risky terrain of Taiwan. To understand the nature of U.S.-China competition in the tech sphere, the Monitor spoke to Paul Triolo.
Paul Triolo is a Partner at DGA-Albright Stonebridge Group, a founding member of DGA Group, and is based in Washington, D.C. As a member of the firm’s China practice and Technology Policy Lead, he advises clients in technology, financial services and other sectors as they navigate complex political and regulatory matters in China and around the world. A recognized expert in global technology policy, Mr. Triolo has worked closely with some of the world’s leading companies on AI, helping them track regulatory issues globally, develop long-term strategies on thought leadership and engage with regulators.
Mr. Triolo spent more than 25 years in senior positions in the U.S. government, analyzing China’s rise as a technology power and advising senior policymakers on technology-related issues. Mr. Triolo is frequently quoted on technology policy issues in media outlets including The New York Times, The Wall Street Journal, The Economist and the South China Morning Post among others. He speaks regularly at conferences and has authored many journal articles and book chapters on global technology policy and China-related issues.
Emma Brignall: Much of the political discourse around China’s tech development is centered on rivalry. How did we get here? Do you feel that the tech sector faces zero-sum challenges with China or more opportunities for exchange and cooperation?
Paul Triolo: The U.S.-China AI competition, or race, is something that’s been building for some time. I left government in 2016. My first experience was around 5G and modern telecommunication systems, where there was this sense of the U.S. and China racing to deploy 5G. At that point, people were talking about some similar issues, like access to data, but it was a very different world, where the argument really was, for example, that Western countries like the U.S. and countries in Europe shouldn’t be overly dependent on Chinese equipment because of the potential for espionage, or the potential for China to turn off this equipment as a long-term threat. I think both of those concerns were way overblown at the time, given the technical means of reducing and mitigating risk. But then, as we got into the 2020–2021 time frame and AI became an emergent technology, I wrote a paper in 2017 with Kai-Fu Lee, who was a big investor in China and AI, about how the U.S. and China were competing in AI. This was well before ChatGPT.
At that time, we divided it into things like internet recommendation algorithms, which were widely used then, particularly by Chinese companies like Alibaba and others. Then we had logistics-related AI, optimizing logistics and supply chains, and we had perception AI, which became a big deal in the U.S.-China context over things like facial recognition. That led to some early export controls on Chinese companies because of its use in policing and surveillance. And then there was autonomous AI. At that time, we broke AI down into those categories, and we tried to assess which companies in which countries were ahead of the game.
Fast forward a little bit to 2022 with the release of ChatGPT. At that point, AI becomes all about so-called generative AI. What we’ve seen since 2022, that we didn’t really see in the earlier AI race, is how fast the technology is developing across so many domains. Much more quickly, we’ve seen the national security implications of Gen AI in areas like cybersecurity, biosecurity, and things like loss of control. The race on the AI side is much more complicated. It’s more complicated because Gen AI, at the same time, is the classic sort of general-purpose technology. People say things like it’s dual-use, which is a little misleading in the sense that the vast majority of deployments of Gen AI are going to be in the civilian sector. Any sufficiently advanced technology will have military and other applications, but if you look at the amount of money invested, for example, by U.S. private-sector companies and Chinese private-sector companies, overwhelmingly, the vast majority of that investment is driving commercial applications.
AI is, at once, this supremely general technology that some people have equated to electricity, like AI will be everywhere. I use AI on my phone and on my laptop every day. At the same time, it has this aspect of nuclear weapons, which is that, when used in certain ways, it can have very dangerous national security implications. It’s very unusual and unlike any other so-called dual-use technology, or any other technology that has this nature, where it’s going to undergird economic development on the one hand and also enhance military capabilities on the other.
The other difference is that it does have this existential and longer-term threat, which is loss of control over these models and the idea that they may not be aligned with human values. We could have a human extinction event. Nuclear weapons have that sort of capacity, but no other general-purpose technology that we generally use, like electricity, is going to kill us. Yet there’s a sense that AI could. That makes this a much more complicated issue in U.S.-China relations, where both countries, and companies in both countries, are rushing to deploy AI across commercial applications and technology stacks.
I think the way to think about this is that there are multiple levels of competition. The really big competition that’s driving a lot of the tension, and a lot of things like export controls and investment controls, is the idea that whoever gets to a very advanced level of AI somehow wins and can then use that capability to its own strategic advantage in the geopolitical realm, or to force regime change. For example, Anthropic CEO Dario Amodei has issued the third or fourth of his essays. Dario is a seminal, critical thinker on all these issues. Of all the CEOs of the major U.S. labs, he’s the deepest thinker on the implications of AI for safety and national security. On the other hand, Anthropic and Dario are very much supportive of U.S. export controls and other controls on China, and he frames it as part of this almost existential struggle between democratic AI and authoritarian AI.
My personal feeling is that that’s the wrong framing because, if we accept that framing, it will make it very difficult for the U.S. and China to collaborate on putting some guardrails around advanced AI going forward. We’re on the eve of the first real Track 1 dialogue—government-to-government dialogue—on AI, coming out of the summit. I believe that was probably the most important deliverable of the summit that happened in May.
But this framing of the technology as a race and struggle to get to advanced AI first, between democracies and authoritarian governments, is really not the right framing because it will preclude the ability of the U.S. and China to cooperate. U.S. and Chinese companies are the leaders in this space, so without the U.S. and China collaborating in some way, and working together to put guardrails around this advanced technology, there’s little chance of reaching a global agreement.
EB: I want to ask more about AI and specifically Chinese-developed AI. In DeepSeek, we saw an alternative to ChatGPT, while Manus, Qwen, and Quark are newer Chinese AI agents and web browsers. With Tencent recently announcing the possibility of a built-in AI agent in its app WeChat, how does China’s progress in AI compare to the US and what are the implications of its development?
PT: You mentioned some companies—I think it’s important at the outset to avoid talking about the U.S. and China in terms of competition because it really is companies and ecosystems that are in competition. At a high level, the way AI has developed in both countries has been a bit different. In China, the industry has been driven primarily by applications and using AI in ways that solve a problem or improve a pain point. When I’ve been in China, I’ve talked to almost all the leading AI companies over the last few months. I’ve also talked to many companies that are deploying AI in their business models—from small companies and fashion companies all the way to bigger players. It’s a very application-focused approach to deployment.
In the U.S., on the other hand, there has been this sense that the biggest AI labs, like Anthropic and OpenAI, and Meta to some degree, xAI, and Google, are all rushing to get to some really advanced level of AI so that one of their models wins. The big difference has been that, for the most part, U.S. companies use closed-source models or closed-weight models. In other words, they have proprietary interfaces, and those companies are deriving a lot of revenue from API access, selling that access to companies, for example. But the model weights are not generally open source. In China, particularly after DeepSeek, all the major Chinese AI labs became open-weight model developers, with the exception of ByteDance. The other players—Alibaba, Tencent, Baidu, DeepSeek, Moonshot, MiniMax, StepFun, and Zhipu, which are all highly capable companies in China—are all, to one degree or another, using and releasing open-weight models.
That means that the cost structures in both countries are very different. In the U.S., companies like OpenAI and Anthropic are driving significant revenue from those models, particularly through enterprise deployments. If you look at their revenue growth over the last year, it’s been ridiculously steep, and that’s why they’re preparing for IPOs later in the year—to leverage some of that revenue and raise further capital to build out data centers and compute capacity. The other thing that has been a bit different, although I think it’s changing, is that in the U.S., financial markets and capital markets have really stepped up. They’re very deep and very broad, so companies that want to build out compute capacity can do that. We’re in this crazy transition from roughly an 80–20 training-to-inference development model, where a lot of the focus has been on training models, to the deployment of these models now that companies are finding them more useful—toward something like 80–20 inference. That requires a lot of capacity to service all the requests that users are making. All of this is driving huge demand for compute.
In the U.S., companies can tap capital markets, both equity and debt, to build out capacity. If you look at the numbers, they’re huge. Now, the problem is that in the U.S., the constraint is energy. Chinese companies have a lot of input into that energy part of the equation. As we look at Jensen Huang’s five-layer cake of AI, which includes energy, models, networking, internetworking, and, at the top, applications, Chinese companies are involved in all of those areas. In some of them, U.S. companies are dependent on Chinese suppliers or the Chinese government. So even though there’s competition between models, that competition is not really happening within the borders of the U.S. and China. In China, Chinese models predominate because U.S. models are not allowed. The Chinese government hasn’t approved OpenAI or Anthropic. Chinese companies find ways around that to use Claude Code and other tools, so there’s lots of slippage. And in the U.S., a lot of companies are building applications on top of Chinese open-source models.
There is, in a sense, some competition, although it’s not outright face-to-face competition. It’s more of a sub rosa competition between the companies. A lot of the competition will happen outside of China, in places like Africa and Latin America, where those countries don’t have their own model developers. They also probably have limited infrastructure. A cheaper Chinese model, like DeepSeek, running on lower-end hardware is going to be much more appealing than paying a huge amount for access to the Anthropic API to run models. Many companies won’t need the most advanced models; they’ll need something that’s pretty good, and Chinese models are pretty good. If you look at the Center for AI Standards and Innovation, which is part of the U.S. AI Safety Institute, it issued an assessment a couple of weeks ago comparing U.S. models on one axis and Chinese models on another. If you look at the divergence, yes, Chinese models are behind. The question of whether Chinese models will close the gap depends a lot on things like access to compute. But Chinese models aren’t that far behind. Consider, for example, that U.S. export controls on advanced compute have been in effect for almost four years, since 2022. The fact that Chinese AI models, in terms of benchmarks, are in many cases almost as good as Western models is actually quite impressive. It’s a complicated environment because of this transition toward inference and because U.S. export controls are placing restrictions on Chinese companies, which the Chinese government is trying to figure out how to overcome.
EB: Do you see Chinese-developed AI having success in the American market? Are there associated risks for security or surveillance?
PT: First of all, let’s shoot down the surveillance thing. The Chinese models are open source, so when they’re used in the U.S., they’re hosted on major hyperscaler platforms like Microsoft or Google, as well as places like Hugging Face and GitHub. I’ll give you an example: I recently downloaded one of DeepSeek’s models and ran it on my desktop PC. I’m downloading that model and the weights from GitHub, and I’m running it locally on my computer. There’s nothing going back to China, so DeepSeek has no visibility into what I’m doing on my PC. That’s the way Chinese models are being used by U.S. companies.
It’s a tricky thing to get a sense of how broadly Chinese open-source models are being used in the U.S. My sense is that it’s pretty broad, but there are different categories of use. There are researchers who want access to a model that’s fairly cheap to run and are probably going to gravitate toward Alibaba’s Qwen, DeepSeek, and other models like Moonshot. Kimi is very popular, as are other Chinese open-source models, because they’re readily available, they’re very good, and they’re cheaper to run. If you’re a research lab, you’re not going to be paying Anthropic or OpenAI big money to run their models through an API. You’re going to download a really good Chinese model and then build on top of that model. Nothing there is going back to China. You’re probably going to be running it either locally or in the cloud on a hyperscaler platform like Google, Microsoft, or AWS. Major U.S. companies are doing this as well. For example, Airbnb announced that it was using Qwen for one of its customer service applications. Perplexity is also using Chinese models.
It turns out that a lot of companies may be reluctant to acknowledge that they’re using Chinese open-source models because of the issues you raised and concerns about things like the models having built-in censorship of certain terms. What I find, however, is that with models like DeepSeek, a lot of the keyword checking and so-called censorship happens at runtime. If you’re not running those models on a Chinese server, which has those additional checks built in, and instead you’re running them on a U.S. server, you don’t get the censorship. Of course, the models are trained on data that may be somewhat more censored than U.S. data. But if you look at the overall datasets they are being trained on, the amount of actually censored data is pretty minimal. Depending on whether you’re doing coding or other tasks, it’s not going to matter that the Chinese model was trained on some small subset of data that was censored. And a lot of the Chinese models are using RAG, which allows them to go out and do real-time searches on the internet and use that data, including information that isn’t censored. So that issue has been overplayed a bit.
The bigger issue is how the U.S. government thinks about Chinese open-source models over time. Because one aspect of the competition, both outside China and more generally, is for mindshare around these models. If Chinese models become dominant in the U.S., it’s a bit of a strange situation where U.S. labs may be leading at the frontier, but their models are too expensive for many companies to use, so those companies end up using Chinese models instead. There’s a lot of effort on the part of U.S. companies, like NVIDIA and others, to develop open-source models. Google is doing this too. Over the next year, it will be critical to see whether Chinese models can continue to improve and whether they will continue to be used by U.S. companies, especially if there are alternative models that are nearly as good.
So far, Chinese models at the open-source level seem to be ahead of U.S. models. Xiaomi came out with a really good model the other day, and Tencent brought in a guy from OpenAI earlier this year to head its model development efforts. Their latest model, Hunyuan 3, is pretty good and has been getting a lot of attention. When you look at this issue in China, each company has a very different business model, and those business models differ in terms of how they integrate AI into their own operations. Just like in the U.S., you have two different kinds of companies. You have the hyperscalers in China, like Alibaba, Baidu, and Tencent. In the U.S., you have Google, Microsoft, and Amazon. Those companies have big cloud infrastructures, but they’re also developing models for their own internal use, for their own products, and for broader development. Those companies have some advantages because they’re deploying their models on their own infrastructure, which enables them to control costs. Then, at the strict model-development level, in China, you have companies like Moonshot and DeepSeek. Those companies are not hyperscalers; they are focused primarily on model development, and each is targeting different niche markets for its products. They’re more analogous to Anthropic and OpenAI, which are doing similar things.
For example, I’m in possession of the ByteDance agentic phone. In December, ByteDance released this really cool phone that was built with agent capabilities from the ground up. You can talk to the phone and say, “Make me a reservation in Chengdu for a hotel and a flight or train.” The agent will then act on your behalf and do all of that for you. This created a big problem in China in December because the agent wanted access to WeChat, Alipay, and all the other apps on your Chinese smartphone. Those companies initially blocked the ByteDance agent from accessing them. Just recently, Tencent started allowing some agentic access to its user interface on the phone. But ByteDance is very much focused on that kind of application: How do you make the user experience good? TikTok, Douyin, and all the applications that ByteDance is famous for are AI-enabled, and that’s why it’s developing AI. ByteDance isn’t really trying to compete on having the most advanced AI models. Every company in this space typically has some niche area in terms of the data they have access to for training models. For example, in China, Baidu has 20 years of search data, Alibaba has all the logistics and e-commerce data, and ByteDance has all the video data. All of these companies have some advantage on the data side, and they all have different ideas about how they want to apply AI.
WeChat and Tencent have been very careful about deploying AI because they have, what, 1.4 billion WeChat users. They don’t want to disrupt that user experience because, when I travel to China next month, I’m going to have WeChat and Alipay on my phone, and I need to be able to rely on them. They’re connected to all my rideshare apps, bike rentals, and pretty much everything else in China. WeChat and Tencent want to make sure they don’t disrupt that experience through AI applications. On the other hand, everybody wants more AI capabilities. The Chinese consumer, unlike the U.S. consumer, is very interested in getting cooler and spiffier kinds of AI applications.
The other interesting dilemma is that something like 80% or 90% of consumers in China want AI and want more AI. In the U.S., we’re moving in the other direction, with people being afraid of AI, opposing AI, and wanting AI to be regulated. It’s almost like an 80–20 split in either country, and that’s a really interesting dynamic to watch. Chinese consumers, in general, view technology as a positive. They’ve seen it improve their lives over the last 20 years in ways that were unimaginable for most people in China. They’re not resistant to AI, and there isn’t this big “doomer” debate that we have in the U.S. about whether AI will eventually kill us all. That debate exists in some places in China, but it’s not dominating the conversation. People like Dario Amodei and Sam Altman in the U.S. have been publicly talking about the dangers of AI in terms of job loss and other risks. As a result, the public perception of AI in the U.S. is very negative, whereas Chinese AI company leaders tend to project a more positive message. They don’t scare Chinese consumers. Although, in China, there has recently been more attention paid to the issue of job loss. But because of the demographics in China and its population decline, that issue has a different spin there than it does in the U.S.
EB: NVIDIA has been pushing to sell more chips to China. Jensen Huang appeared in Beijing at the recent Trump-Xi summit, and the Department of Commerce approved sales of its H200 chips to 10 Chinese companies back in January. However, China has yet to make purchases, reportedly over concerns about bolstering its domestic industry. What are the implications of increased trade of technology between the U.S. and China and why has the process stalled on both sides?
PT: We have to go back to the president’s Truth Social post in December, which really was the initial break that resulted from Jensen Huang’s lobbying in Washington with the president and with White House adviser David Sacks. They were making the argument that the U.S. should allow some less advanced NVIDIA GPUs to go to China so that China would not become completely dependent on domestic sources of hardware. That argument is very subtle and nuanced because, basically, David Sacks and Jensen Huang were saying, “We need to keep Chinese developers working with U.S. hardware so that there’s not a complete transfer of those developers.” I think Jensen says that half the world’s AI developers are in China. Jensen doesn’t want to lose all those developers to a Chinese AI stack. The argument has been that the U.S. should ease up a little bit on the controls and allow H200-class GPUs, which are now a couple of years behind the cutting edge, to be sold to China.
Now, the problem in China is that the Chinese government has looked at this differently. Its view is that, eventually, it wants most Chinese companies to use domestic hardware. Even though the U.S. has changed the game a little by allowing H200s, we don’t know what’s next because the technology is changing fast and the Trump administration has not articulated a clear long-term policy. Is this just a one-off decision where Chinese companies are going to get access to H200s and then be cut off forever from all more advanced NVIDIA chips? Or is there going to be a threshold or a time period after which the most advanced U.S. technologies will be available for export to China? There’s no certainty there. That uncertainty makes the Chinese government reluctant to keep its companies addicted to U.S. technology. It’s a complicated issue because most of those AI developers in China were trained on NVIDIA hardware and NVIDIA software, such as CUDA, which is a development environment that is very complicated to master. Once you’ve mastered it as a developer, you generally don’t want to move to another development environment.
The problem is that, as the U.S. restricts Chinese companies, some Chinese companies are now making the calculation that they can’t count on having access to U.S. hardware, whether through smuggling or remote access. The other complicating factor is that Chinese companies can still use remote GPUs located outside China. For example, they can access GPUs in Japan or Southeast Asia that are housed in data centers run either by their own companies or by third parties. That’s a loophole that some in Congress are trying to close. But it’s really a product of the fact that U.S. export controls were not designed to address this kind of situation, where companies can remotely access hardware that itself may be subject to export controls. It’s an artifact of the fact that the U.S. is using export controls originally designed for weapons of mass destruction, nuclear technologies, and chemical weapons, and trying to apply them to the very different and much more complicated problem of AI development. What the U.S. was really trying to control were the workloads. As a result, the hardware itself becomes much harder to control because of smuggling and transactions where the legality is often murky. Markets, Chinese companies, and intermediaries have found ways to meet demand in China. The U.S. controls were never really designed to prevent some of these workarounds.
The semiconductor situation is complicated because we’re in a transition period where, probably within another two to three years, most Chinese companies will be using a Chinese-centered AI stack, with Huawei processors and GPUs from other Chinese companies like Biren, Moore Threads, and Enflame. Many of those companies were founded by engineers who previously worked at NVIDIA and AMD, so their hardware tends to be more compatible with NVIDIA hardware and NVIDIA software. We’re in this weird transition period between U.S. and Chinese hardware stacks. Each Chinese company has a different calculus in terms of what it uses for training models, what it uses for inference, and how it divides up its compute capacity across those workloads. They’re all struggling a bit with that transition because, unlike companies in the U.S., they don’t know what their access to advanced compute will look like in two years. They may not have access to NVIDIA hardware, and the domestic ecosystem may not have produced sufficient quantities of advanced processors for them to use. As a result, they’re in a bit of a dilemma when it comes to planning their technology roadmaps going forward.
EB: Taiwan has remained at the forefront of the semiconductor industry, prompting concerns about the security of the world’s advanced technology development if China were to attempt a military takeover. How do you think Taiwan should leverage this technology-based playing card? How should the United States navigate semiconductor trade with China and Taiwan?
PT: It’s important to understand where we are generally in U.S.-Taiwan-China relations and then understand the impact of the semiconductor issue. Traditionally, the relationship has been based on a set of documents agreed to more than 40 years ago—and in some cases 50 years ago—between the U.S., China, and Taiwan. These include the so-called Three Communiqués and, on the U.S. side, the Taiwan Relations Act and the Six Assurances, one of which was that the U.S. would not negotiate around arms sales to Taiwan. When these documents were developed, everyone was eager to have U.S.-China relations begin on good terms. The problem of Taiwan and resolving the Chinese Civil War between the KMT and the Chinese Communist Party was essentially kicked down the road. And it has continued to be kicked down the road since 1982, or even earlier than that.
But since 1982, we’ve seen developments that nobody could have envisioned at the time. You wake up in 1982, and Taiwan is an important but relatively small island. It wasn’t even a democracy then; it was a dictatorship under Chiang Kai-Shek. It was a flashpoint in U.S.-China relations, but now an agreement has been reached and broader relations begin to develop. Taiwan gradually becomes more of a back-burner issue in U.S.-China relations. Then you wake up in 2022, and Taiwan is the epicenter of the global IT economy. TSMC, the semiconductor foundry that Morris Chang founded in the late 1980s, has become the key player in manufacturing and packaging advanced AI hardware. Who could have conceived of that in 1982? Nobody could have, right? Then you fast-forward to 2026, and as I’m looking at my screen here, tens of trillions of dollars in capital expenditures and market value across U.S. stock exchanges and exchanges around the world are now dependent on Taiwan remaining a viable entity and on TSMC functioning 24/7, cranking out advanced semiconductors. It’s an inconceivable situation compared to where we were in 1982.
The problem is that those relationships and those documents still form the foundation of how people talk about the issue. People point to them and say, “Nothing has changed. U.S. policy hasn’t changed. Chinese policy hasn’t changed. Taiwan’s policy hasn’t changed.” But, of course, the surrounding environment has changed drastically and radically. My contention has been that we need a new way of thinking about this problem because it was acceptable in 1982 to keep kicking it down the road. But now, when we do exercises for companies involving scenarios where Taiwan is blockaded, for example, the consequences are much more significant. A full military exchange would be catastrophic for the global economy; there’s really no way to de-risk that. There are also less dire scenarios, such as China blockading the Taiwan Strait and the airspace over Taiwan for a couple of weeks. Even that would have an immediate impact on the global economy. Some studies have estimated something like a $15 trillion hit to the global economy, depending on how certain factors are calculated. Just a couple of days ago, I heard someone cite a figure of $8 trillion. I think that vastly underestimates the impact because, as I often point out, the market capitalization of NVIDIA alone, which is something like $6 trillion, is totally dependent on Taiwan. If Taiwan disappeared tomorrow, NVIDIA stock would presumably go to zero, and the pension funds of U.S. retirees would be heavily impacted.
The AI dimension has complicated things even further because companies like Anthropic—and its CEO, Dario Amodei, whom I mentioned earlier—have published a series of essays in which he is essentially arguing that the U.S. needs to race to achieve advanced AI first, in part to enable regime change in authoritarian countries, by which he really means China. If you’re sitting in Beijing, and the leader of the leading company in the world on AI development is saying the reason we’re developing advanced AI is so we can force regime change in China—and the basis of his whole company are GPUs that are manufactured in a province of China or a country, depending on where you’re sitting, that China considers part of China—as you’re sitting in Beijing, you’re thinking, Well, that’s an interesting twist. I’ve argued that this raises the risk that China may reconsider its calculus on Taiwan. It’s complicated because this isn’t the only issue Chinese leaders would weigh when deciding how to approach Taiwan. There are longstanding concerns related to Taiwan’s political status, questions of independence, and the optics of relations between Taiwan and the mainland. So semiconductors, TSMC, and AI are now part of the equation, but the issue is far more complicated than any one of those factors alone. At a minimum, however, those issues increase the risks surrounding Taiwan.
The other piece is the export controls. U.S. export controls have dragged Taiwan and TSMC into the mix. U.S. controls now extend extraterritorially, and the result is that many leading Chinese companies can no longer use TSMC as a manufacturing base in Taiwan. That’s yet another factor that increases risk. If you’re a leader in Beijing and your companies are being cut off from using a manufacturing base that you consider part of China, you’re going to view that as undesirable.
We need a totally new approach to the problem. The main proposal from the existing foreign policy establishment in Washington is to arm Taiwan and pursue a policy of deterrence. The problem is that Taiwan is not Ukraine, and China is not Russia. The idea that you can arm Taiwan to the point where you would deter China is, in my view, dangerously naïve because it’s a very different situation. The historical roots are different, and Taiwan is really not defensible. If you go to Taiwan and see where the fabs are on the west coast, and you take the high-speed rail from Hsinchu down to Tainan, that’s it. That’s where the world’s epicenter of AI is. It’s a very narrow strip. It’s not like Ukraine, where there is at least a buffer and where Ukraine can mount a credible defense. There’s no way to do that in Taiwan.
On the other hand, yes, there needs to be a policy of deterrence. But I would argue that it needs to be coupled with a policy of assurance. There’s no military solution to this problem. The military solution, if you look at the numbers, essentially amounts to the destruction of the global economy. That shouldn’t be an option. The problem is that the mindset in Washington is stuck in this deterrence framework instead of thinking outside the box. I’ve proposed, for example, that we need a fourth communiqué. We need to think about potentially making an agreement with China and Taiwan that would make Taiwan a global commons. That would be complicated because the U.S. would then have to lift those export controls, and there would have to be some agreement among all parties around AI so that the idea of the U.S. using Taiwan as a hardware base to undermine the Chinese government is taken off the table. There would also have to be a lot of confidence-building measures on both sides.
I’m afraid that, given the direction we’re on now, we’re potentially headed toward conflict over Taiwan, with all the tremendous economic, political, military, and social risks that would entail. I think the best outcome would be for the technology component of this—AI and the broader technological revolution—and Taiwan’s critical position within it to help produce a peaceful resolution of the Taiwan issue. Ideally, both sides would agree that Taiwan is too valuable to fight over, let alone destroy. The idea of missiles destroying those fabs in Hsinchu is absurd. So is the idea that the U.S. should bomb the fabs if there is a peaceful resolution of the issue, which some people have discussed. That’s equally absurd in my view and dangerously naïve. You’re going to bomb the fabs in Taiwan that your leading technology companies depend on—for what gain? To prevent China from gaining access to advanced semiconductors? There isn’t enough outside-the-box thinking on this issue to get us out of the current cycle of China conducting military exercises, Taiwan responding, Congress wanting to supply more arms to Taiwan, and China hawks in Washington arguing for more deterrence. That cycle is really dangerous. Unfortunately, there are very few people arguing for a different approach. What we need is a complete rethinking of the status quo among all three parties, and of how critical issues like semiconductors and AI affect the equation. Too often, these issues are treated as an afterthought. The policy debate typically ignores what we’re talking about here, which is the semiconductor dimension of the issue. And I think we ignore that dimension in these discussions at our own peril.
EB: I’m curious how this might extend to U.S.-China interdependence. With the tariff battle, we saw this sudden resistance to the effects of globalization and interdependence, especially with rare earths and critical minerals, and realizing how much of a stake American markets have in China. With the aftermath of tariffs and under this administration, I’m wondering how you think U.S.-China relations are shifting and what the effects on interdependence will be for administrations to come?
PT: The rare earths issue has really exposed, on the U.S. side, a massive blind spot. If you’re going to start weaponizing supply chains—which began under the Biden administration and, to a lesser degree, under the first Trump administration—the question becomes: Where does that end? What is the ultimate goal? How do you prevent it from getting out of hand? Clearly, the Chinese government was aware of its companies’ dominance in rare earths and had briefly used that leverage against Japan in 2010. However, it did not really decide to weaponize that capability until the U.S. put in place the October 2022 export controls. Those controls were very sweeping because they targeted a critical part of the semiconductor industry—not just GPUs, but also the equipment needed to manufacture semiconductors. From Beijing’s point of view, this was not simply about national security. It was also about China’s right to economic development. In fact, Xi Jinping raised this as a red line with President Biden in April 2024 and again later in 2024 at the APEC Summit. He argued that China’s right to economic development was a red-line issue on par with Taiwan. That was China’s response to the U.S. export controls.
Since then, we’ve seen additional controls on critical minerals, as well as a range of other legal measures, including the Anti-Foreign Sanctions Law and the Anti-Blocking Law. Then, in April 2025, China implemented the broad licensing regime for heavy rare earths, which had a significant impact. In October 2025, there was even more regulation of rare earths, including extraterritorial measures covering rare earths, critical minerals, and related equipment. Earlier this year, we also saw the rollout of State Council Directives 834, 835, and 837, which provide a higher-level framework for China’s broader response to what it sees as the weaponization of supply chains. Those directives effectively provide the policy scaffolding for China’s pushback against these measures.
We’re at a point where both sides have demonstrated the ability to damage each other’s economies. When China cut off rare earth exports last April, magnets ran out at a Ford facility in Detroit within a couple of months, and the company couldn’t produce Ford Explorers. Over the last year, we’ve had this back-and-forth over rare earths because China has been slow to issue licenses. It is issuing a lot of licenses, but not fast enough or broadly enough. It’s controlling things like yttrium, which is critical for aircraft engines and semiconductor manufacturing tools. It’s also controlling cobalt-based magnets, which are critical for many applications, including automotive uses as well as military and defense systems. We’re in this strange situation where both sides have weaponized supply chains. Where does this end? How do we get off this bandwagon?
It’s also starting to impact AI supply chains. Reuters recently did a piece on indium phosphide, which nobody had thought about. We were already aware of it because we’ve been working in this space, but indium phosphide is one of those obscure technologies that few people understand, yet it’s a critical part of the photonics supply chain. The U.S. is intent on building its own AI stack, but that stack is largely built across Taiwan, South Korea, Japan, and China, with critical inputs sourced from China. Those inputs range from specialized materials like indium phosphide for photonics to more mundane components such as transformers, power semiconductors, and turbine blades, all of which have substantial Chinese content in their supply chains. What we’re finding out is that the idea that the U.S. can completely decouple from China and build an entirely independent technology stack is really a mirage. When you throw in Taiwan, it becomes even more of a mirage.
Yes, onshoring some manufacturing in the U.S. is good. It creates jobs and reduces dependence on Taiwan to some degree. But it still doesn’t solve the deeper problem of interdependence, of which rare earths are really the poster child. It turns out that semiconductor export controls are not truly a chokepoint because they’re only partially effective—perhaps 50% or 60% effective—whereas China’s control of 90% of the rare earths supply chain for critical minerals is really a chokepoint. The U.S. unleashed the chokepoint weaponization idea, but only China has real chokepoints.
Some politicians in Washington have not yet come to grips with that because there’s still a degree of denial. One of our clients spoke with some senior U.S. government officials, and they were not even aware of some of these critical dependencies, even at this arguably late stage. I found that amazing. That just highlights the fact that there’s a lot of denial in Washington, and a lot of hope that the U.S. can reduce its dependence on rare earths in one or two years, which is totally unrealistic. The two countries have to figure out a way to coexist. Neither side can fully decouple and de-risk its entire semiconductor or rare-earths supply chain. Doing so would cost an enormous amount of money, take a decade or more, and be highly wasteful. In the meantime, companies are in a very difficult position because they often have no alternative. Either there is no alternative source available outside China, or China could cut off the sources they currently rely on.
As an example, a Japanese company that gets tungsten powder from China has been unable to obtain tungsten powder for months because of the Chinese controls, which were largely implemented in response to U.S. controls. Japan has also been caught up in this because of comments the prime minister made about Taiwan. That company supplies leading fabs such as TSMC, Samsung, and SK Hynix. It has informed those fabs that it will run out later this month of products that are critical to manufacturing advanced semiconductors. There needs to be a rethinking of this policy of relying on chokepoint technologies without any clear endgame in sight. I and others are advocating for ways out of this situation, but it’s difficult because there is so much tension. There is very little trust between the U.S. and China, or between China and Japan. It turns out that once you start down this path, it’s extremely difficult to get off the treadmill. We need to think about ways to get off that treadmill before there is even more collateral damage.