Anthropic urges compute controls now

Anthropic published a policy paper on May 14 arguing that U.S. officials should move now to preserve what it called America’s “compute advantage” before 2028. The company said access to advanced chips remains the key input for building the most capable AI systems and warned that China is still narrowing the gap through loopholes in existing controls. The paper framed the issue as a near-term policy choice, not a distant forecast. It said the window for action is limited because Anthropic expects transformative AI systems to arrive by 2028. ### Why is Anthropic focusing on “compute” rather than models or talent? Anthropic said on May 14 that “the most important ingredient” for developing advanced AI is access to the chips used to train models. The company’s argument rests on the fact that the highest-end AI chips are designed by U.S. companies and can still be constrained through export rules. Anthropic said those controls have been “incredibly successful,” but added that Chinese labs have remained close to U.S. peers through talent, loopholes and distillation attacks. (anthropic.com) Anthropic has been making versions of that case for more than a year. In a March 6, 2025 submission to the White House Office of Science and Technology Policy, it called for tighter semiconductor restrictions, including controls on Nvidia’s H20 chips, government-to-government agreements for countries hosting large chip deployments, and lower no-license thresholds. In an April 30, 2025 filing on the Commerce Department’s diffusion rule, it said older chips could leave some countries facing AI training costs 10 times higher by 2027 than users of cutting-edge U.S. chips. (anthropic.com) ### What loopholes does Anthropic want Washington to close? Anthropic’s May 14 paper pointed to three pressure points: chip smuggling, access to U.S. compute through foreign data centers, and distillation attacks that copy the capabilities of stronger models. The company said Chinese labs have stayed close to the frontier partly by “exploiting loopholes” around export controls and by extracting capabilities from American systems. (anthropic.com) The company has separately described distillation in operational terms. In a February 23 post, Anthropic said DeepSeek, Moonshot and MiniMax generated more than 16 million exchanges with Claude through about 24,000 fraudulent accounts to improve their own models. Anthropic said distillation is a legitimate method when companies use it on their own systems, but argued that illicit distillation lets competitors acquire capabilities “in a fraction of the time, and at a fraction of the cost” of developing them independently. (anthropic.com) ### How does Anthropic connect distillation to export controls? Anthropic said on February 23 that distillation attacks “reinforce the rationale for export controls” because restricted chip access limits both direct model training and the scale of illicit copying. The company argued that, without visibility into those attacks, outside observers may wrongly conclude that export controls are ineffective when some advances instead depend on capabilities extracted from U.S. models. (anthropic.com) That line matters because Anthropic is asking policymakers to treat model outputs and compute access as linked parts of the same control regime. Its May 14 paper said the United States stays ahead in one scenario by tightening export controls further, disrupting distillation attacks and accelerating AI adoption across democracies. In the alternative scenario, it said policymakers fail to close loopholes and Chinese firms catch up or overtake the frontier by 2028. (anthropic.com) ### Are U.S. trade officials treating chip controls as a live China-talks issue? Jamieson Greer, the U.S. trade representative, said on May 15 that semiconductor export controls were not a major subject of recent discussions with Chinese officials in Beijing. “We did not talk about chip export controls at the meeting,” Greer told Bloomberg TV, according to Reuters. Reuters said the comments indicated a breakthrough on sales of Nvidia’s H200 chips to China remained distant. (anthropic.com) President Donald Trump said on May 15 that he and Chinese leader Xi Jinping had discussed artificial intelligence and possible cooperation on AI safety, Reuters reported. Greer said any decision to allow H200 imports would be a “sovereign decision” for China. ### What happens next in Washington? Anthropic’s latest paper puts its recommendations directly in front of U.S. policymakers as the administration weighs how hard to press on semiconductor controls and related AI restrictions. (money.usnews.com) The company’s earlier submissions were filed with the Office of Science and Technology Policy and the Commerce Department, and its May 14 paper reprises the same core demand: move before the gap closes. The next concrete marker is whether U.S. agencies translate those asks into rulemaking or enforcement. For now, the public record shows Anthropic escalating its case on May 14 while Greer said on May 15 that chip controls were not a major topic in the latest Beijing talks. (anthropic.com 1) (anthropic.com 2)