The open-weights letter hits 270 signatories, and Anthropic still won't sign

By August 3 the Open Weights and American AI Leadership letter had passed 270 signatories, with Meta, Microsoft, Nvidia and IBM among them, according to AI News and Microsoft's own post. It started ten days earlier, on July 24, with five names: Hugging Face, Meta, Microsoft, Mistral and Nvidia, arguing against premature blanket restrictions on open-weight models. The list grew from 25 to over 50 in a single day once OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, Block and Ollama joined, per CNBC and Axios. Two names have stayed off every version of it: Anthropic and Amazon.

What the letter is reacting to

Three parallel government moves, all catalysed by Moonshot's Kimi K3 release. The US is considering a federal procurement ban on Chinese models. Treasury's Bessent said on July 21 that sanctions are on the table over alleged distillation and IP theft. And China's MOFCOM is considering its own export controls on model weights. The letter is industry pushing back on the first; the third is the one I would watch, because it cuts the other way. It restricts what leaves China, not what enters the US.

Enforceability is the illusion

The July 26 brief puts it well: once weights are out, a procurement ban decides who may deploy a model inside a company or agency, not who can download it. Kimi K3's weights landed on Hugging Face on July 27. DeepSeek V4, Qwen and GLM-5.2 have been downloadable for months. Nothing in a procurement rule pulls those bytes back. So the real effect of any restriction is compliance exposure for the deployer, not technical availability, and that is a very different risk to plan for.

Once the weights are on a hard drive, regulation cannot reach the model; it can only reach you.

Why Anthropic will not sign

Anthropic's line, as the July 30 brief summarises it, is that weights cannot be recalled, which is consistent with a closed-frontier business and, per WSJ reporting, with a planned IPO. Part of Silicon Valley reads it as safety used as a competitive moat; the WSJ piece on the backlash makes that argument explicitly. I do not think you need to pick a side to draw the practical conclusion: the market has split into open-friendly (Nvidia, Meta, OpenAI, Google, Microsoft) and holdouts (Anthropic, Amazon), and that split is now a vendor-risk input, not just a policy story. A White House frontier framework was expected before August 1; whatever it ends up saying, the split will outlast it.

Risk posture if you run Chinese open weights

Teams running DeepSeek, Qwen or Kimi in production need positions on four questions this quarter. Not answers from me, positions from them.

• Procurement risk: does any customer or contract sit in the public sector or a regulated industry where a ban would bite? If yes, this is a sales problem before it is a technical one.
• Data residency and IP exposure: self-hosted weights keep data in-house, which is the strong argument for them; document it so it holds up in an audit.
• Exit plan: name the Western open-weight alternative you would switch to, run it against your evals now, and keep the adapter layer thin enough that switching is a config change. My router notes cover the mechanics.
• Licence hygiene: MIT (DeepSeek V4 Pro, GLM-5.2) is cleaner than a modified licence; record which one each deployed model carries.

None of this means stop using DeepSeek or Kimi. It means the decision is now a governance decision with a paper trail, which for most engineering teams is new. The good news is that the paper trail is cheap: a one-page register of which open models run where, under which licence, with which fallback, answers most of what a compliance team will ask.

The honest limitation

Everything here is press releases and counts of signatories. No text of a procurement ban exists yet, MOFCOM has considered rather than issued, and the letter is a letter, not a law. If you are betting your architecture on this, bet on the one thing that is certain: the weights you already have will still run tomorrow, and the question of whether you are allowed to run them is going to be asked by someone in a suit, not by the model.