Ball game’s over—the US is out of the AI chip market in China

A recent New York Times headline says it all: “Trump Approved a Nvidia Chip for Sale in China. Beijing Doesn’t Want It.” The administration endured substantial domestic criticism for approving sales of Nvidia’s H200 chips to 10 Chinese companies, but because of Beijing’s continuing objections, as of mid-May not a single H200 chip had been sold to Chinese companies. Following his recent meeting with Chinese President Xi Jinping—where, according to U.S. Trade Representative Jamieson Greer, chip export controls were not a major topic of discussion—U.S. President Donald Trump said China had not bought any H200 chips “because they chose not to. They want to try and develop their own.”

The fundamental reason for this posture is that China’s regulatory authorities no longer think U.S. chip companies are reliable business partners. As a result, there will be no going back to earlier levels of interdependence, where U.S. companies dominated the Chinese chip market. Even if Chinese companies wanted to train and operate their artificial intelligence (AI) models on U.S. chips, China’s authorities would not let them.

The global AI and chip marketplaces, however, are still up for grabs, and U.S. companies still welcome and hold an advantage there. The U.S. and China might not have really reached a detente or stalemate on the chips issue, however, and there are policy mistakes the U.S. could make in a misguided effort to keep China in a technological box. Such miscalculations would hurt the ability of U.S. companies to maintain their lead with international buyers. Chief among these potential policy errors are expansions of chip controls to third countries in the hopes of keeping Chinese companies from evading U.S. controls, measures to exclude Chinese AI models and applications from the U.S. and other markets, and restrictions on access to Chinese memory chips by U.S. and allied companies. Each of these ideas is under active consideration among U.S. policymakers and commentators, and each would accomplish little except isolating U.S. companies from the global AI and chip marketplace.

The story begins in 2019 when the first Trump administration faced an earlier technological challenge from Huawei, China’s leading telecommunications and network equipment supplier. As recounted by Chris Miller in “Chip War,” the U.S. feared that Huawei’s products were of such high quality and priced so attractively that they would shortly form the backbone of the next-generation 5G telecom networks throughout the world. If this new era of connectivity were built on a foundation of Huawei technology, wrote Edward Fishman in “Chokepoints,” “China would gain an enormous advantage in its deepening geopolitical standoff with the United States.” America’s dominance of the world’s tech infrastructure would be undermined, and China’s geopolitical clout would grow. The U.S. met this challenge by cutting off Huawei’s access to U.S. technology, including semiconductors.

Huawei responded to this embargo by developing its own semiconductor technology. In September 2023, it announced a new advanced smartphone, the Mate 60 Pro, with a domestically produced 7-nanometer processor made by China’s domestic chip builder Semiconductor Manufacturing International Corporation (SMIC). Huawei recovered from the ban and demonstrated that it could provide consumer products powered with advanced chips that were fully competitive with U.S. products.

In September 2022, the Biden administration significantly extended its chip export controls in response to the growing capabilities of AI language models. It announced a new series of cascading restrictions on exports to any company located in China—cutting off access to the high-end chips needed to develop frontier AI models, blocking access to the tools and equipment they would need to design and make their own high-end chips, and ending access to the components they would need to build their own chip manufacturing equipment. In 2023, these restrictions were extended to plug technical gaps, prevent legal circumvention of their intent, and combat illicit smuggling.

In January 2025, the outgoing Biden administration proposed a diffusion rule to further control access by Chinese companies to U.S. AI technologies. The rule divided the world into allies without export restrictions, countries of concern where chip exports were essentially banned, and all other countries, which faced substantial regulatory burdens and quantity restrictions to qualify for an export license.

Independent industry analysts like Timothy B. Lee warned that over time these middle countries would turn to China to get their increasingly capable chips rather than jump through these burdensome hoops, noting that “the rules appear optimized for a world where transformative AI emerges within 3–5 years.”

The incoming Trump administration seemed to agree. In January 2025, it announced that it would repeal the diffusion rule, and in May the Commerce Department did so officially, saying the rule stifled innovation, weakened diplomatic relationships, and created a competitive disadvantage for U.S. firms.

Also in January 2025, the Chinese startup DeepSeek startled the AI world with its highly capable AI models that were almost on a par with frontier models developed in U.S. labs despite being developed at a fraction of the cost. These DeepSeek models were optimized to work with Nvidia’s H20 chip, which had been designed specifically to comply with U.S. export controls. Hundreds of thousands of H20 chips had already been sold to Chinese customers, earning Nvidia $12–$15 billion in 2024. Chinese companies Alibaba, ByteDance, and Tencent rushed to buy the chip to equip their data centers to run the DeepSeek models.

The U.S. responded to this development by changing the export control rules. In April 2025, the Commerce Department declared that the once-permitted H20 chips that powered DeepSeek’s breakthrough AI model were now noncompliant with export controls. Nvidia took a $5.5 billion write-off to account for the resulting loss of sales to Chinese companies.

In July, however, the Commerce Department reversed itself and said it would grant Nvidia export licenses for the H20. In an August 2025 press conference, Trump defended this decision, saying the H20 chip was “obsolete.”

But Nvidia never sold any H20 chips. Chinese authorities told AI companies in China not to buy them, citing security concerns, and Nvidia stopped manufacturing them.

The Trump administration’s attempted relaxation of chip export controls continued, however. In December 2025, Trump announced that the United States would allow Nvidia’s H200 processors to be exported to China. These powerful chips were more capable than H20s but not as good as Nvidia’s top-of-the-line Blackwell chips and had not been permitted under the previous administration’s export controls. The idea was to get Chinese companies hooked on less powerful Nvidia chips, thereby assuring that they would not bother to develop their own capable chips and would always remain behind the cutting edge set by U.S. companies.

Despite this approval from U.S. regulators, Chinese authorities have not allowed domestic AI companies to purchase any H200 chips. One reason for this ban might be that Chinese officials are prioritizing developing the highest possible domestic capabilities farther back in the AI stack, in the chips that enable the training and day-to-day operation of AI models rather than in the models themselves.

Chinese officials often prioritize developing basic infrastructure capabilities over providing the best short-term service to companies closer to the end user. For instance, in the early 2000s, Chinese authorities refused to give a license for U.S. payment card companies like Visa and MasterCard to provide domestic payment services, even though the international brands had demonstrably better card networks than the fledgling domestic network operated by China UnionPay. The authorities knew perfectly well that this exclusion of foreign providers meant that merchants in China would have to live with less reliable and convenient electronic payment services. But it was worth that short-term sacrifice to develop a domestic competitor that could go head-to-head with the global brands in China and eventually around the world. When China lost a World Trade Organization case on the exclusion issue in 2012, it agreed to open up its domestic market in a reasonable period of time. It opened up slowly, however, granting Mastercard an operating license only in 2020. It took until 2024 until MasterCard processed its first domestic transaction in China. Visa is still waiting for a license, an issue that arose in the meeting between Xi and Trump this May. During this delay, China UnionPay had developed an excellent payment service that was fully the equal of the international networks.

Chinese authorities are playing the same long game with chips and AI models. Somewhat less capable models today are recognized as the price for world-class chips in the future. Of course, the Chinese foundation model startups DeepSeek, Moonshot AI, MiniMax, and Zhipu AI and the established tech giants Alibaba, Tencent, Xiaomi, and ByteDance would be delighted to access more compute from U.S. chips to train and operate their AI models. As a result of export control compute constraints, Chinese AI models are months behind the U.S. AI frontier, and Chinese AI companies don’t like being second best. But Chinese authorities seem to think that the U.S. lead in AI models is only temporary, like the lead that international payment networks had over China’s domestic champion. China’s chip designers and manufacturers will soon catch up with today’s global leaders, Chinese authorities seem to think, and that future abundance of compute will then put Chinese AI models on a par with the best U.S. models. 

The second reason for keeping advanced U.S. chips out of China is that Chinese authorities do not want to play Charlie Brown to the U.S.’s Lucy. What Chinese authorities must have learned from the H20 experience is that the U.S. would continually move the goal posts on chip controls. If China’s domestic companies learned how to use H200 chips to make competitive AI models, they reasoned, then the U.S. authorities would block access to those chips just as they did with the H20s. Chinese authorities could not see relying on H200 chips as a long-term platform for national AI development. The Western adage about fooling me once has its equivalent in the Chinese saying that falling into a pit once adds an inch of wisdom.

A third reason for China to block access to U.S. chips is that China’s AI and chip companies do not need U.S. chips. They can work around the U.S. chip embargo by innovating at different levels of the AI stack. Chinese AI model companies can design highly capable AI models that work with available domestic chips, as DeepSeek has done with its new close-to-the-frontier V4 model, optimized to run on Huawei’s Ascend chip. Chinese chip designers can innovate in chip architecture to craft highly capable chips that can be mass produced by Chinese chip builders like SMIC. Huawei might be on track to do this with its new “LogicFolding” architecture. The idea is to boost data transmission speed in transistors by stacking chips one on top of the other rather than focus on shrinking transistors to fit more onto the chip. SMIC could build these powerful chips without reliance on the export-controlled extreme ultraviolet lithography equipment made by Dutch company ASML. If it is successful, Huawai says, its chips would match the performance of TSMC’s 1.4 nanometer chip by 2031.

China’s larger economic development goals seem to include fostering “producer services” that will accelerate productivity in Chinese manufacturing. Developing a competitive chip design ecosystem—a Chinese Nvidia—would be a key element of this new emphasis on producer services. Allowing U.S. AI chips into the domestic market would undermine this drive for an indigenous AI chip and model ecosystem.

The fourth reason for the blockade on imports of U.S. AI chips is that Chinese authorities seem to have a different conception of AI success. They appear to share the perspective of law professors Yonathan A. Arbel and Matthew Tokson who reject the logic of an AI race to the finish line in model capabilities. Instead, they seem to think that rapid diffusion of capable AI models throughout the economy is the path to success.

This path to success in AI relies on models, but it is also a matter of energy, pricing, and government policy. China has a plentiful supply of energy to power the chips that train and operate AI models, generating more than twice the electricity as the U.S. In contrast, the U.S. is still struggling to adapt an aging grid to the new demands of AI data centers, creating a strategic bottleneck that may choke AI adoption in the U.S.  

China has affordable AI pricing. Its major models are all open source, which allows developers to build new models on top of the open models at no cost. For businesses that want to use the models, the cost of access is far less than comparable U.S. models, on average, one-sixth the price per token of U.S. offerings from OpenAI, Anthropic, and Google, according to some estimates. In the U.S., the practice of unrestricted “tokenmaxxing” in enterprises has exposed the high cost of using U.S. AI models for business tasks and caused companies to pull back from AI usage to avoid sticker shock.

China’s AI Plus policy directive prioritizes applying AI models to solve immediate economic and social needs on multiple fronts rather than rushing toward an ever-receding goal of artificial general intelligence. It encourages companies and promises them support if they use AI as a tool for scientific discovery and innovation, for industrial transformation, and as a consumption booster. For these purposes, AI models do not have to be top tier; they can be good enough to provide a meaningful boost to productivity in specific lines of business.

The fact that Chinese AI models are several months behind the cutting edge of U.S. models as measured by industry benchmarks is of less importance than these other factors. Being at the top of these benchmarks has proven to be decreasingly relevant in measuring AI utility to users. Business value appears to be more driven by scaffolding and harnessing that enable AI agents to perform valuable enterprise tasks in legal services, banking, investment, sales, and coding. Many industry observers, including influential adviser Mary Meeker, have concluded that for almost all relevant business purposes, the underlying AI models themselves have become commodities distinguished by price and ease of use rather than raw capabilities. Going forward, improvements from one AI model generation to the next, as noted by analyst Paul Kedrosky, are marginal rather than transformative. From a practical perspective, being a few months behind the cutting edge in AI models might not matter very much.

The major Chinese tech companies might want continued access to Nvidia chips to avoid dependence on Huawei. Huawei competes with Alibaba, ByteDance, and Tencent in cloud computing, AI models and applications, and a broad range of consumer services. Ideally, these other tech companies would not want to rely on a competitor for the AI chips they need to power their services. Some of them are taking steps to develop their own proprietary AI chips to mitigate this dependence. But this fear of competitive abuse from Huawei is not enough to drive them into critical dependency on U.S. chips, given the risks that this access could be withdrawn at any moment.

The main thesis of this commentary that the U.S. and China AI ecosystems are separating and will remain bifurcated for the foreseeable future seems to be threatened by evidence that China’s AI companies are using remote access to data centers located outside of China to evade chip export controls. According to several reports, non-Chinese data centers in third countries like Singapore and Malaysia currently obtain U.S. chips without a license and lease access to the chips to Chinese companies like ByteDance and Alibaba. Under this line of thinking, Chinese AI companies are still significantly dependent on U.S. chips but do not need to import them, since they can get all the U.S. compute they need through offshore data centers.

In addition, there are reports that institutions linked to China’s military still seek to procure Nvidia chips, suggesting to some analysts that China’s military and intelligence services, whether they like it or not, have no realistic alternative to Nvidia chips. These reports have fueled extensive discussion of the “cloud loophole” in export controls and whether and how to close it. The recent announcement that Commerce will require a license to export U.S. chips for “entities headquartered in China” even if the facilities themselves are located outside of China appears to be an attempt to close this loophole.  

This policy debate presupposes that the Chinese military and AI companies are so dependent on Nvidia chips through remote access that cutting them off would significantly degrade their intelligence and commercial capabilities. As analyst Paul Triolo shows, however, this assessment ignores developments in China’s AI stack over the last few years that enable the Chinese military and AI companies to use domestic chips for their intelligence and commercial needs. The scale of the reported evasion also seems marginal, not large enough to demonstrate substantial Chinese reliance on U.S. AI chips. The U.S. is still shut out of the Chinese market for AI chips.

Triolo sums up the situation clearly, writing that for China to approve substantial H200 purchases “could slow adoption of domestic alternatives, weaken incentives for Chinese cloud providers and developers to optimize software stacks around local hardware, and reintroduce a dependency on a U.S.-controlled technology platform that Washington has repeatedly demonstrated it is willing to restrict.”

Last year Brookings China analyst Kyle Chan warned that U.S. chip controls could soon reach the point of no return for U.S. chipmakers in the China market. That point has now been reached. U.S. chip companies have exactly zero market share of the AI chip market in China and have no prospect of returning to their once-dominant position there. The ball game is over, and the U.S. has lost.

U.S. regulators sought a decoupling from China through its chip export controls, and initially China resisted. The U.S. seems to be having second thoughts about the separation, but China has moved on. If being shut out of the high end of the world’s largest chip market seems a sad ending, perhaps the lyrics from Taylor Swift’s “All You Had to Do Was Stay” might provide some cold comfort: “Let me remind you. This was what you wanted.”