Early in his second term, President Donald Trump staked out an accelerationist approach to artificial intelligence. Instead of restricting the spread of U.S. semiconductors, as his predecessor did, his administration would promote their export far and wide.
Fast-forward a year, and officials in his administration are sounding the alarm about China's AI advancements. They warn that China's AI labs are quickly closing the gap with their U.S. counterparts. Chinese labs, they say, are stealing intellectual property from U.S. labs and using it to create capabilities that endanger our national security. The administration is now considering banning Chinese AI models.
Both of these approaches to the AI race are logical. Promoting economic dynamism and protecting national security make perfect sense -- but only when considered in isolation. When patchworked together, these two strategies produce convoluted, discordant, and contradictory policies. Such incoherence is giving Chinese labs a runway to catch up in frontier AI.
Take, for example, the administration's efforts to sell Nvidia's powerful H200 chip to China.
The Biden administration restricted Nvidia from selling H200s to Beijing. In January, however, the Commerce Department cleared those chips for sale, and Trump mandated that 25% of the revenue would go into the government's coffers. By May, Commerce had approved 10 Chinese buyers. Among them were Alibaba Group, which the Defense Department subsequently added to its list of Chinese military companies, and ByteDance, the owner of an app, TikTok, that posed such concerning national security risks that Congress forced the sale of its U.S. operations to non-Chinese owners.
Those pushing for the H200 sales to China say we shouldn't be concerned because the chip has nowhere near the capabilities of Blackwells and Rubins, Nvidia's most advanced offerings. This is true, but it is also beside the point. The more salient comparison is between the H200 and the other chips that Chinese firms can legally purchase. The H200 is the best training semiconductor that Chinese labs can directly acquire, either from international or domestic companies. And the administration is fine giving it to them.
That's strange, considering Chinese AI labs' offerings are rapidly approaching the frontier and have capabilities that threaten U.S. national security, including advanced cyber vulnerability detection.
While trying to control China's access to U.S. chips, the administration is also working to sell them internationally. The idea is if you beat Chinese chip providers to global markets, the majority of the world will rely on U.S. infrastructure. So, despite the administration's security concerns, Commerce approved the sale of tens of thousands of Blackwells to Saudi Arabia and the United Arab Emirates.
Policies that encourage widespread exports of cutting-edge chips run the risk of easing Chinese companies' access to them. Chinese AI labs may not be able to legally buy these chips due to U.S. export controls, but they can rent them in third-country data centers. The White House recently accused the Chinese start-up Moonshot AI of training its latest large language model, Kimi K3, on Blackwells in Thailand.
The administration clearly knows about this loophole. It declines to close it. (The House-passed Remote Access Security Act would solve the problem. But, lacking the president's support, the bill has languished in the Senate.)
Why would the administration resist putting limits on which companies can access American-made semiconductors abroad? Because that would institutionalize conditions and audits that would dampen demand for the very technologies the administration is pushing. The White House wants massive sales volumes while outpacing China in AI, a needle that may prove impossible to thread.
The administration is right to be concerned about the security of Chinese AI, and U.S. government agencies should not use it. But restricting the use of open-weight Chinese models, in the absence of viable alternatives, will only harm U.S. companies.
Hundreds of tech firms recently argued in a letter the administration that a ban would raise costs and consolidate the domestic AI market around Anthropic, OpenAI, and Alphabet's Google. They make a good point: A prohibition on Chinese models could create risky market concentration and cede the open-weight model space to China.
A better approach is to support the growth of a vibrant open-weight and open-source ecosystem in the U.S. That was a central tenet of the AI Action Plan the administration set forth in July 2025. The plan encouraged the spread of these models, arguing that they are geostrategically important. It called for Washington to support their growth by making compute capacity available to startups and academics, among other initiatives.
Unfortunately, in the year since, Chinese open-weight models have sprinted ahead of their U.S. competitors. This should push the White House to double down on promoting the development of U.S. open-weight models that would give domestic tech companies more options and make U.S. labs more competitive with their Chinese rivals.
Pushing for economic dynamism, of course, will lead to difficult national security trade-offs. The faster Trump realizes this, the quicker he can formulate a coherent AI strategy.
Guest commentaries like this one are written by authors outside the Barron's newsroom. They reflect the perspective and opinions of the authors. Submit feedback and commentary pitches to ideas@barrons.com .
Sam Bresnick is a research fellow and an Andrew W. Marshall fellow at Georgetown University's Center for Security and Emerging Technology.
Comments