In a strange alignment, U.S. national security hawks have sought to stop Nvidia from selling downgraded versions of its artificial intelligence chips to China, and Chinese hawks are escalating attempts to stop Chinese AI firms from buying them. After initially banning their sale in April, U.S. Commerce Secretary Howard Lutnick explained why the Trump administration reversed course and lifted the ban in July: “You want to sell the Chinese enough that their developers get addicted to the American technology stack.” Beijing is worried the U.S. strategy could work, leading Chinese regulators to warn firms not to buy the chip. Nvidia has since halted production of the H20 and is now reportedly hoping to sell an even more powerful chip tailor-made for China.
Of course, China would rather have its companies training AI with Chinese chips, just as the United States has wanted its firms to avoid Chinese rare earths. While both countries have a strong incentive to move to domestic suppliers, neither has managed to conjure enough high-quality, price-competitive local alternatives to convince the firms in their jurisdiction to buy domestically and eliminate these mutual dependencies. This is an important factor contributing to stability in U.S.-China relations.
Recently, some hawks have argued that Lutnick is wrong and the United States should instead impose a full blockade on both AI chips and chipmaking equipment to maximize the U.S.-China gap in computing power now and make AI competition “as lopsided and unfair as possible.” They imply that this strategy is both good for security and effectively costless to the United States because China is already full steam ahead for indigenization of AI chips, so selling it chips “will not meaningfully diminish these efforts.”
Whether Lutnick’s strategy is worth pursuing fundamentally depends on whether continuing to sell China tuned-down AI chips (the H20 at issue is far behind the cutting edge for training AI models, but not as far behind for running and improving their performance) slows its push for self-sufficiency. How much weight should be put on short-term AI capabilities that China gains from U.S. chip purchases today, when those sales may generate medium to long-term leverage? It is extremely challenging to draw the right line for U.S. interests—one that avoids failure by preventing too many powerful chips from being exported to China, but also does not go too far in pushing China’s AI firms to be all in on self-sufficiency.
What blockade advocates tend to leave unsaid is their assumption that Beijing will not fight back. But China has learned from the United States and built an export control apparatus that should not be underestimated. It has also shown a willingness to use it in retaliation for U.S. controls. After all, the arguments for chip bans from the U.S. perspective—that they can be used for the military, to prevent stockpiling, and that leverage should be maximized before the competitor can work around the chokepoint—are even more immediately true for bans on China’s chokepoints, most notably in critical minerals, which are important to many U.S. military supply chains and cannot be replaced in the near term.
If the United States cuts off one of the few things that China genuinely depends on it for, then it will have unilaterally reduced a Chinese dependence without even a chance of fatally wounding its competitor. That foregone leverage means reduced deterrence and retaliation against future aggressive Chinese actions—an invasion of Taiwan, for example.
What have the effects of export controls been so far? U.S. export controls blocked Chinese firms from buying the best AI chips in October 2022, just before ChatGPT started a race to build larger data centers. Chinese firms could only legally buy chips that were weaker for AI training.
As a result of both controls and major investment at home, the United States has gone from rough parity with China in 2022 to a large and widening lead in AI computing power, even as analysts estimate that Nvidia sold around 1 million H20 chips to China last year. The United States compute lead is secure, as millions of even more powerful chips come online in U.S. data centers.
Chinese chips have not been able to fill the gap. Despite more than a decade of serious investment to indigenize, Chinese chipmakers still cannot produce even close to as many high-end AI chips as its companies need to train and run advanced AI models. Export controls on key chipmaking equipment have made this harder, but analysts expect supply constraints will fade over the next few years. Homemade chips are poor quality in comparison to U.S. ones—and they are also not produced in the huge numbers needed. As a result, combined with a software ecosystem for AI training that is far inferior to Nvidia’s, China’s best AI models remain trained on Nvidia software and hardware.
At the same time, the gigantic compute advantage over China has translated to an underwhelming gap in AI capability. U.S. AI models are consistently state of the art, but despite the chip embargo constituting its “biggest challenge,” DeepSeek’s powerful V3 and R1 models were only months behind U.S. ones when they were released in late 2024 and early 2025, respectively. Since then, many Chinese firms—including Alibaba, Zhipu, and Moonshot—have released even better models. Many factors have contributed to this result, including stockpiling, smuggling, renting access to chips, software improvements in AI training that make weaker chips more useful, increased usefulness of inference, plentiful electric power in China, and abundant AI talent.
Thus, export controls have neither failed nor worked as well as hoped. Tightening them further may not have as much impact on Chinese capabilities as advocates hope. The controls made life harder for Chinese AI companies, but it is not obvious that capability gaps today will translate into a significant military advantage when military adoption of technology lags behind the consumer sector by many years. Technology is hard to predict—it is far from certain that huge computing gaps will remain the key differentiator for strategically relevant AI capacity.
There is no denying that China is already investing significant resources in every step of the chip sector to weaken U.S. controls. Export controls, however, have certainly influenced this effort by helping the government better align diverging incentives between its AI developers, chipmakers, and chipmaking equipment companies. A blockade would force China’s highly talented AI companies to be all in and shorten the controls’ shelf life.
Despite China’s power, AI companies have their own agency and interests. To some extent, Chinese AI labs are forced to work with firms like Huawei and Cambricon to signal alignment with government objectives, but they remain focused on beating their domestic rivals in a cutthroat domestic market. Switching to Huawei is disruptive and could lock them to a buggy software ecosystem with fewer, worse chips, risking falling irreversibly behind rivals who keep using Nvidia chips.
Progress in the semiconductor sector only occurs with intense collaboration upstream and downstream. Taiwan Semiconductor Manufacturing Co. (TSMC) can only advance the cutting edge by working with ASML and others to integrate advanced tools into its manufacturing and receiving better designs from players like Nvidia. The power of AI training chips comes not only from raw hardware, but also the software ecosystem. AMD for example sells few AI chips despite strong hardware because of software weaknesses. If Chinese firms continue to use U.S. chips and software, since most Chinese AI models are open source, any innovations that they stumble upon will also improve the U.S. software ecosystem, furthering the gap between it and Huawei’s software.
But with time, China will keep making better chips. Just as chip controls created an incentive to make domestic chips, chip equipment controls have boosted China’s chip equipment makers, giving Chinese players like the Semiconductor Manufacturing International Corporation and Huawei an incentive to work more with domestic equipment producers—even when doing so means more defects than just buying the best equipment from abroad.
As a result, China has made significant progress in many kinds of chipmaking equipment. Some have even been integrated into advanced chipmaking at market leader TSMC. Extreme ultraviolet lithography, ASML’s specialty, is the biggest bottleneck with no near-term prospect of a Chinese replacement. But in the long run, with enough years and no shortage of funds, China’s semiconductor equipment sector will eventually innovate past the thresholds that the United States has tried to set.
One long-standing doctrine in export controls is to allow the export of sensitive dual-use goods when there is “foreign availability” of a similar product that is not controlled. After all, if the adversary will get the product anyway, then the control was more than useless because it both deprived a U.S. firm of revenue and the United States of knowledge of a competitor’s capabilities. U.S. Rep. John Moolenaar’s “rolling technical threshold” policy proposal is a good example of policy that would automatically update controls as Chinese domestic chips improve to ensure that Chinese firms will be able to get better chips from the United States than they can from domestic producers.
From the perspective of Chinese firms, that U.S. policy makes it a better strategic choice to at least delay the full move to Huawei. The initial H20 ban had the opposite effect, leading Chinese firms to invest more in helping Chinese chipmakers by sending the message that they have no way forward with U.S. chips.
One other challenge for the embargo policy is that the United States’ ability to impose worldwide controls it has used to, for example, force TSMC not to make AI chips for Huawei, relies on agreement from allies to impose similar controls. The United States has now created the perception that U.S. firms can get controls lifted with a side payment, meaning allies would face Chinese retaliation for tougher controls without any guarantee that U.S. firms would be bound by similar restrictions. As a result, the United States is not likely to find a receptive audience for a blockade in Tokyo or The Hague. Going it alone, meanwhile, would mean watching U.S. firms lose the entire Chinese market to foreign competition.
Policymakers need to have realistic expectations for imperfect tools. Former U.S. Commerce Secretary Gina Raimondo is right that export controls amount to “speed bumps” in the long term, as is U.S. President Donald Trump’s AI czar, David Sacks, that taking export controls too far risks having Chinese technology replace U.S. and allied tech first in China and then around the world.
The more U.S. controls try and freeze China in place, the more they risk forcing Chinese AI firms to go all in on developing a homegrown AI ecosystem, when existing controls already lock in a massive compute gap. A full shutoff would only marginally increase that gap, while risking China applying an even more damaging blockade on U.S. industry.
The post What Would a China Chip Blockade Cost? appeared first on Foreign Policy.
