Anthropic urges Uncle Sam to kneecap China's AI ambitions before 2028 - The Register
The export control trap: how strangling China's AI ambitions may accelerate them
On May 14, Anthropic published a policy paper arguing that the United States has roughly two years to lock in a decisive lead over China in artificial intelligence — and that the window for decisive action is closing fast. The paper called for tighter chip export controls, legal clarity that AI "distillation" techniques are off-limits, and a global push to adopt American AI systems. It was a serious, detailed document from one of the most influential AI labs in the world.
That same day, Nvidia CEO Jensen Huang was in Beijing, sitting beside President Trump, making the case for exactly the opposite. Nvidia wants to sell chips to China. Lots of them. The two companies had the same audience — the Trump administration — on the same day, with directly opposing financial interests. That collision is the real story in Anthropic's paper, not the policy recommendations themselves.
Anthropic's core argument is straightforward: if AGI-level AI arrives around 2028, as the paper speculates, the country that gets there first gains structural advantages that are hard to reverse. A 12-to-24 month lead in frontier capabilities, the paper argues, would be "enormously advantageous." The solution, in Anthropic's view, is to close the loopholes in the current export control regime and move fast, before Chinese firms close the gap through sheer engineering talent and state resources.
The loopholes are real. Chinese firms have accessed more than 2,300 banned Blackwell GPUs through offshore rental arrangements — routing purchases through Indonesian data centers operated by a company called INF Tech, and through a $1.2 billion contract between Tencent and Datasection, a Japanese firm. The Remote Access Security Act, which passed the House 369-to-22 in January, was designed to close exactly this cloud-compute route. It has not passed the Senate. The cloud loophole remains open.
Trump administration policy has moved in the opposite direction. In December 2025, the White House reversed Biden-era restrictions on Nvidia's H200 chip, approving sales to China under a 25% tariff and a 50% volume cap. Jensen Huang was not in Beijing to thank the administration for tightening controls. He was there to expand the market.
This is where Anthropic's paper runs into an uncomfortable historical echo. Export controls on strategic technology have a poor track record when the goal is permanent suppression of a competitor's capabilities. During the Cold War, the COCOM regime restricted advanced Western technology exports to the Soviet Union, attempting to freeze Soviet technological progress. The result was not a permanently retarded Soviet capability. It was a Soviet Union that built its own chip industry out of necessity — slower and less efficient, certainly, but real. The restrictions became an industrial policy incentive, not a ceiling.
China is not starting from the Soviet baseline. It hasts a deep semiconductor industry, substantial engineering talent, and the most advanced AI research ecosystem outside the United States. Three of thirteen major Chinese AI labs have published safety evaluations; none have disclosed evaluations of chemical, biological, radiological, or nuclear risks — a gap Anthropic rightly flags. But the response to that gap is not obviously more effective export controls. It is just as obviously an acceleration of China's own capability development.
Anthropic has a commercial interest in strict export controls that the paper does not acknowledge. Nvidia, AMD, and OpenAI have equally clear commercial interests in relaxed ones. Lisa Su of AMD testified to Congress that overly restrictive controls could "push other countries toward China AI technology." Sam Altman has made similar arguments. These are not disinterested technical assessments from any direction.
The most intellectually honest reading of Anthropic's paper is that it is a lobbying document dressed as policy research — one with internally consistent logic, credible supporting evidence, and a clear beneficiary. That does not make it wrong. The AGI timeline is speculative. The 12-to-24 month lead target may be politically impossible regardless of policy choices. But the paper is right that the current moment is consequential, and right that the cloud loophole needs to be closed.
What it misses — or declines to say — is that each closure tightens the noose, and the noose is historically one of the most reliable accelerants of domestic innovation. China will not accept a permanent second-place position in AI any more than the Soviet Union accepted permanent technological inferiority. The question is not whether China develops world-class AI capabilities. It is whether the path is through American chips, which keeps the relationship commercial and inspectable, or through domestic Chinese chips, which does not.
Anthropic wants the noose tighter. Nvidia wants it looser. The Trump administration is, characteristically, doing both at once — reversing H200 restrictions while hosting Jensen Huang in Beijing and not moving the Remote Access Security Act through the Senate. This is not a strategy. It is a negotiation between two interest groups with the same administration, and the administration has not decided which way it wants to fall.
The narrow window Anthropic identifies may be real. But the history of strategic technology restrictions suggests that the act of strangulation is often the thing that makes the subject stop needing to breathe.
Sources:
Anthropic policy paper, May 14, 2026
Nikkei Asia: Anthropic urges tough chip controls as Nvidia CEO joins Trump in China
The Register: Congress votes to close GPU cloud loophole
Introl: Remote Access Security Act and the cloud loophole
TechPolicy.Press: Technology restrictions as economic statecraft
Courthouse News: Tech leaders testify to Congress on China AI competition