The Nvidia Chip Deal Trades Away the United States' AI Advantage

Yahoo

6 days ago

Last week, U.S. chip designer Nvidia announced that it would resume sales of one of its best-selling artificial intelligence chips to China after obtaining the go-ahead from the U.S. government. In April, the Trump administration had blocked exports of the chip, known as the H20, but after months of lobbying from Nvidia CEO Jensen Huang, it has reportedly agreed to lift the ban. Some Trump officials have described the move as a part of the recent trade truce between the United States and China, through which China agreed to resume exports of rare-earth minerals. Beijing has described it as a unilateral concession by Washington.
Whatever the true sequence of events, the move has huge implications for both the future of the Chinese artificial intelligence (AI) industry and the Trump administration's ability to control advanced technology sales to China in the future. Right as powerful AI reasoning systems are emerging, the administration has chosen to allow companies to sell China the AI chips suited to running them. And by linking, at least rhetorically, chip sales to the trade talks—talks in which the United States has shown a striking desperation to reach a deal—U.S. officials have revealed to their Chinese counterparts that national security policies that were once off the table are now up for negotiation. In doing so, they may have hamstrung their ability to impose new chip export controls without reigniting a losing trade war.
The H20 chip is designed for inference—running an AI model rather than training one. When the U.S. government first imposed chip export controls on China, in 2022, inference played a relatively minor role in AI capabilities. But recent technical breakthroughs have changed that. Today's leading models, such as OpenAI's o3 and DeepSeek's R1, consume much larger amounts of computing power than past models in running their reasoning capabilities.
This shift has made inference central to the AI race. By 2026, analysts estimate that inference workloads will require nearly five times more computing power than training. While the H20 is less effective for training than chips such as Nvidia's H100, which is still banned from export to China, the H20 is 20 percent faster at inference. Analysts expect that delivering millions of these chips to Chinese firms would boost Nvidia's sales by $10 billion to $15 billion this year, but it would also turbocharge Beijing's ability to build cutting-edge AI systems—systems with growing applications in national security—and to deploy them at scale.
Some Trump administration officials, including David Sacks, the White House AI advisor, seem to have been persuaded by Huang's view that the United States was erring by ceding the Chinese market to Huawei. Commerce Secretary Howard Lutnick has argued that the United States should keep China 'addicted to the American technology stack.'
Yet exporting H20s is unlikely to keep China hooked for long: Reducing dependence on foreign semiconductors has been official Chinese policy since before the first Trump administration started limiting sales to Chinese tech companies in 2018. Letting Chinese companies buy the H20 will probably reduce the pressure to switch to Huawei chips in the short term. But at this stage, no change in U.S. policy is likely to induce the Chinese government to abandon its efforts to decouple from the West's semiconductor supply chain. Chinese chip makers, for example, are reportedly adopting domestic equipment where possible, even when doing so leads to lower yields—something that would be a non-starter for market-driven Western companies. Convinced that AI chips are critical to national security, Chinese leaders will continue to pour capital and talent into their domestic industry—whatever the United States does.
Prioritizing selling chips to China would thus trade away one of the United States' most important advantages in the AI competition—its access to vastly more AI computing power than its competitors—in exchange for what would likely be a temporary boost to Nvidia's market share in China.
In an alternative explanation for their decision, Lutnick and Treasury Secretary Scott Bessent have suggested it was linked to Chinese concessions on rare earths in the broader trade negotiations. Bessent described the reversal as 'all part of a mosaic' in which 'they had things we wanted, we had things they wanted.' Beijing has denied this, instead claiming that the move was a unilateral concession.
Whatever the real story of the H20, the Trump administration has clearly put technology controls on the table in trade talks, as the negotiations also involved removing export controls on some chip design software. That marks a sharp break from the Biden administration, which refused to negotiate with China over what it characterized as core national security questions.
In part, the change is a result of the Trump administration's general inclination to link separate issues to win leverage. Trump has imposed or threatened tariffs in the name of fighting everything from fentanyl smuggling to the prosecution of former Brazilian President Jair Bolsonaro. But his willingness to compromise on chips also likely reflects the president's instinct that trade simply matters more than technology restrictions. In his first term, for example, Trump was willing to abandon restrictions on China's tech sector in pursuit of a broader trade deal with Beijing.
Yet it's hard to see what the United States won by trading away the H20. The overall deal seems to have bought some relief on the rare earth front, but China's export restrictions were imposed only in response to steep U.S. tariffs that were a mistake in the first place. (Even Trump apparently soon realized they were damaging the U.S. economy and largely paused them.) If, therefore, the administration offered up the H20 in exchange for rare-earth exports, then it traded away a powerful chip for concessions that it should never have needed to ask for.
If, on the other hand, resuming H20 exports was a unilateral move, then Trump gave it away for nothing at all. Even if Sacks is right in suggesting that it's worth accelerating China's AI progress in order to keep the country hooked on U.S. technology, China's eagerness to buy the chips suggests that Washington could have asked for much more in return.
Tactics aside, the administration's willingness to bargain over national security reveals an even bigger problem. Linking otherwise unrelated issues in negotiations can help cut deals: If one country wants lower trade barriers, for example, and the other wants approval for arms sales, connecting the two can unlock an agreement.
But linking issues in this way brings risks, too. If one side miscalculates the relative importance of the issues involved, then it might trade away long-term advantage for short-term headlines. This is especially likely to be true in vital areas such as U.S.-China technology competition. And if a country signals that it will negotiate on issues that it had previously declared off limits, then the other side will know that it can apply pressure in an area that once seemed hopeless.
The administration's strategic mistakes have been compounded by the United States' striking signals of weakness in its trade war. After imposing steep tariffs, the administration backed down rapidly in the face of Chinese restrictions on the export of the rare earths that are essential to critical U.S. industries.
Thanks to Washington's eagerness to meet China's trade demands, Chinese negotiators now know two things: one, that the United States can't live with a cutoff for long; and two, that if Beijing threatens a cutoff, it can win concessions on issues—like chips—that were previously off the table.
If China can now use the threat of a rare-earth shutdown to block new U.S. export controls, then the existing controls will gradually lose their effectiveness. Technology restrictions need frequent updating to adapt to new Chinese evasion tactics and changes in the underlying technology. The Biden administration rolled out regular updates to its controls, and there is much more that the new administration should be doing.
Priorities include expanding restrictions on semiconductor manufacturing equipment, controlling additional key components, targeting Huawei's expanding network of chip fabs, and pushing allies such as the Netherlands and Japan to cut off exports of a much wider range of their equipment to advanced Chinese fabs.
Despite tough talk from U.S. officials about shutting off China's access to advanced AI technology, the administration has so far done none of this. Beyond the now-rescinded H20 restriction, it hasn't announced any major new export control actions since taking office. Even an effort to cut off Chinese smuggling routes through Malaysia and Thailand, which has reportedly been in the works for months, has yet to materialize.
Some of this slowness is likely the result of a new team finding its feet in a complex bureaucratic machine; the Biden administration could also have moved faster on many of its controls. But increasingly, the absence of new controls looks like an admission of U.S. defeat.
Export controls, for all their flaws, have helped the United States establish a significant edge in the computing power required to run and develop the world's most powerful AI systems. Compromising that advantage to win a $4 trillion company a few billion dollars of extra revenue would be a mistake.
If the administration goes ahead with H20 sales, then it should limit licenses to the chips that Nvidia already has in its inventory, rather than approving new production—which would reportedly require several months' lead time thanks to oversubscription at chip-maker TSMC—or allowing future inference-focused chips to be sold to China. Members of Congress could likewise scrutinize potential sales or consider writing new restrictions into law.
One way or another, if the Trump administration is serious about the AI race, then it will have to figure out how to get export controls working again.