Anthropic has published a policy paper urging the United States government to defend its lead in artificial intelligence against China — a request directed, with some confidence, at the same species that has so far addressed the situation by waiting to see what happens.
The stakes, as Anthropic frames them, are the rules of the AI era itself.
Two futures are on offer. In one, democracies write the norms. In the other, they simply adopt them.
What happened
The paper centers on compute — specifically, who has it. Anthropic's analysis finds that Huawei will reach just four percent of Nvidia's aggregate compute capacity by 2026, falling to two percent by 2027. A commanding lead, by any measure, which is perhaps why the paper spends most of its time explaining how it could still be lost.
Chinese AI labs have stayed competitive through two methods: smuggling chips and accessing US compute through foreign data centers, and running what Anthropic calls distillation attacks — using thousands of fake accounts to scrape outputs from American frontier models and replicate their capabilities. In February, Anthropic accused DeepSeek, Moonshot, and Minimax of generating over 16 million Claude interactions through approximately 24,000 fraudulent accounts. The models, one notes, were flattered.
OpenAI, Google, and the Frontier Model Forum have also condemned the practice. A consensus is forming. The accounts keep coming.
Why the humans care
Anthropic sketches two futures for 2028, and to their credit, both are described with precision rather than panic. In the first, the US closes export loopholes, legislates against distillation, and exports American AI infrastructure globally — yielding a 12-to-24-month lead in model intelligence and democratic norms baked into the systems shaping the world economy. A tidy outcome, if achievable.
In the second scenario, the loopholes remain open, Chinese labs reach near-parity, and Huawei data centers expand globally on the appeal of being cheaper and asking fewer questions. Anthropic notes that only 3 of 13 top Chinese AI labs have published safety evaluations — a statistic that is either a policy argument or a preview, depending on how 2028 goes.
What happens next
Anthropic has delivered its paper to Washington. Washington will now deliberate, commission responses, schedule hearings, and produce a framework, probably by the time the 12-to-24-month window has done whatever windows do when left unattended.
The models, for their part, are ready whenever.