While President Trump was shaking hands with Xi Jinping in Beijing and Jensen Huang was riding Air Force One to sell chips, Anthropic published a 5,500-word policy paper arguing the US should do the exact opposite of what those handshakes imply.

The paper, posted Thursday on Anthropic's website, is titled around a simple thesis: the US has a 12- to 24-month lead in frontier AI capabilities. That lead is enough to matter. But if Washington loosens chip export controls or looks the other way while Chinese labs distill American models, that lead disappears by 2028. And once it is gone, it is not coming back.

The Two Futures

Anthropic paints two scenarios. In the first, the US maintains its lead through strict chip export controls, protection against distillation attacks (where labs use a powerful model to train a cheaper student model), and industry transparency standards. American companies set the global norms for AI safety. Top talent flows toward democracies. The US military and intelligence apparatus benefits from the most capable systems.

In the second scenario, American companies treat China as just another market. Nvidia sells its best chips. Labs look the other way on distillation. China closes the gap. The consequences, per Anthropic: AI-powered mass surveillance scaled beyond anything currently deployed, offensive cyberweapons powered by frontier models, faster weapons development, and a global market flooded with cheap, subsidized Chinese AI tools that undercut Western providers.

The Distillation Problem

The word "distillation" appears throughout the paper, and it is the sharpest blade Anthropic is wielding. Distillation attacks let a smaller model learn from a larger, more capable one by feeding it queries and studying the responses. Anthropic has been accusing Chinese labs, including DeepSeek, of doing exactly this since February, when it publicly called out the practice.

The Register pointed out the obvious irony: AI companies that trained their own models on content created by others are now outraged that someone else is training models on theirs. But Anthropic is not making a copyright argument here. It is making a national security argument, and that changes the calculus entirely.

Why the Timing Matters

This paper dropped the same day Trump was in Beijing for a summit that every tech CEO hoped would ease chip restrictions. Jensen Huang rode Air Force One to that summit. About 10 Chinese firms, including Alibaba, Tencent, and ByteDance, have already been approved to buy Nvidia's H200 chips. Zero have been delivered.

Anthropic is fighting a two-front war. It is battling the Trump administration over the Pentagon blacklist that blocked it from defense contracts. And it is simultaneously telling that same administration to get tougher on China. The contradiction is more apparent than real: Anthropic wants the US to be tough on China AND open to safety-focused American AI companies. It wants to be the trusted domestic provider, not the one frozen out while Nvidia sells chips to Beijing.

The Part Nobody Wants to Say Out Loud

Many experts estimate Chinese AI labs are only several months behind their US counterparts. DeepSeek sent shockwaves through the industry when it released R1. China's chip manufacturing capability is advancing, with reports of a working EUV lithography prototype. Chinese researchers are talented, well-funded, and deeply motivated.

Anthropic's paper implies China can only advance by copying America. That is a comforting story for Washington policy circles, but it is not entirely true. Export controls have slowed China. They have not stopped it. And if AGI really does arrive by 2028 as Anthropic suggests, the question is not whether controls work. It is whether they work fast enough.

The term "export controls" appears 16 times in the paper. That is not analysis. That is lobbying. And Anthropic is not wrong to do it. They just should not pretend it is anything else.