Microsoft, Nvidia, Meta and 22 others defended open weights. Anthropic and OpenAI didn’t sign.

The argument over open-weight AI just got a lot more explicit. Last week, White House officials accused China's Moonshot AI of stealing American IP through a technique called distillation, and started floating sanctions. Anthropic's policy chief called it industrial espionage. Nearly 200 startups signed a letter to Trump asking him not to cut off the Chinese open models they depend on, and then 25 major organizations, including Microsoft, Nvidia, Meta, Hugging Face, IBM, Mistral, and Y Combinator, put their names on a statement warning against premature restrictions. Anthropic and OpenAI didn't sign it.

The reason startups are panicking is simple: price. Particle founder Shuail Doshi said it bluntly, warning that hundreds of companies would instantly die without cheap Chinese models, adding that it's great for Anthropic because everyone would have to start spending money there instead. A New Stack test backs him up. Kimi K3, Moonshot's open-weight model, produced the same code as Anthropic's Claude Fable 5 across a bug fix, a refactor, and a feature build, for about a third of the cost, at $3 per million input tokens and $15 per million output tokens versus Fable 5's $10 and $50. It was four times slower. For a startup counting tokens, that's a good trade. On OpenRouter, Chinese models have made up more than 30% of U.S. token usage every single week since February, hitting 46% at one point.

Anthropic's response, timed almost too neatly, was Opus 5, launched Friday at half the price of Fable 5 and matching it on most of Anthropic's own benchmarks. The company argues cost-per-finished-task matters more than cost-per-token, which is a fair point. But even after cutting its own price in half, Anthropic still charges more per token than Kimi K3.

The split inside the industry is telling. Nvidia's Jensen Huang told Axios that American companies should absolutely use strong Chinese models, and called distillation fundamental to intelligence, not theft, while agreeing that actual privacy or contract violations should have consequences. Anthropic and OpenAI have aligned in warning Washington about the risks of powerful open-weight Chinese models, according to Axios. Notably, Nvidia sells chips to Anthropic and OpenAI too, and Microsoft is one of OpenAI's biggest investors. These aren't companies picking a side out of loyalty; they signed anyway.

The hypocrisy is hard to ignore. This same week, a court approved Anthropic's $1.5 billion settlement for downloading and retaining pirated books, even as a separate ruling found that training Claude on those books was fair use. Anthropic has also ramped up federal lobbying spending past all of last year's total, all while continuing to support open-source tools like MCP that expand Claude's own ecosystem. Benefiting from material you didn't own is the common thread in both the book case and the distillation fight, even if the legal questions differ.

For developers, the practical lesson isn't abstract. Hosted models, whether Claude or Kimi, can vanish without warning if policy shifts or a provider pulls the plug, the way Fable users found out recently. Weights you've actually downloaded and can run on your own infrastructure are much harder to take away. If you're relying on open models through someone else's API, that's still a dependency. Download the weights. Run them yourself. Know your stack works before a policy fight forces you to find out.