The Open Weights Letter Grew Ten Signatures Overnight, and One of Them Was OpenAI

The letter went out Friday morning. Nvidia’s Jensen Huang fronted it with his first-ever post on X, which is the kind of detail that tells you how much the signatories wanted this noticed. Politico got it exclusively. Satya Nadella shared it. Elon Musk quote-tweeted Huang with “this has my full support,” though SpaceX did not sign.

And then, per CNBC’s Friday piece, the headline fact: “OpenAI and Anthropic did not sign the letter.” The Next Web ran with the same framing. Tom’s Hardware too. The absence of the three closed-model leaders was the story.

That framing lasted about fourteen hours.

The list is a live document

The PDF Nvidia hosts still carries the original 25 names. So does the Internet Archive’s capture of Microsoft’s page from Friday at 21:54 UTC. The live page this morning carries 35. The ten additions:

A Hacker News commenter caught it at around the same time, replying to a thread that had spent the evening puzzling over Sam Altman’s position with a flat three words: “openai just signed the letter.”

GitHub signing separately from its owner Microsoft is a nice touch. But OpenAI is the one that matters, because on Wednesday Axios reported that OpenAI and Anthropic had converged in Washington on warning about Chinese open-weight models, a story I wrote up on Thursday. OpenAI has now signed a letter arguing that restricting open weights would concentrate AI “in the hands of a few providers.” Both things are on the record within 72 hours of each other.

The public trail is consistent, at least in tone. Greg Brockman told reporters in New York on Thursday that AI usage “is something that is actually very important to democratize” and that he had not been in any conversations with the administration about banning Chinese models. Altman posted Friday that he wants the US to win with both open and proprietary models, and that he was “glad to see this.” Glad to see it, then on it. What happened in between is the interesting part, and nobody involved has said.

What the letter argues

Strip the 1980s open-source framing and three claims do the work.

First, openness is itself a safety argument. “Relying solely on closed models is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect. And concentrating advanced AI capabilities behind a small number of closed models compounds that risk.” This is aimed squarely at the case OpenAI and Anthropic have been making in Washington, and it lands harder than it would have a month ago, because Hugging Face spent last week reconstructing a live breach after Anthropic’s Fable 5 refused the forensics and Z.ai’s open GLM-5.2 did the job.

Second, the risks are real and the answer still isn’t prohibition. The letter concedes the hard part in one sentence: “Once released, the weights are beyond the original developer’s control, and modified versions are difficult to trace or reverse.” That is Dario Amodei’s argument, restated accurately and then set aside.

Third, and most pointedly, distillation is a legitimate technique. Policymakers “should be careful not to conflate legitimate model-development techniques with misappropriation,” and unlawful extraction should be handled through “targeted legal and commercial frameworks rather than sweeping restrictions.” Michael Kratsios had accused Moonshot on Wednesday of industrial-scale distillation of Anthropic’s Fable while building Kimi K3, an allegation whose evidence I picked through and found thin on the weights side. Scott Bessent had floated sanctions the day before. The paragraph is a rebuttal with a two-day turnaround.

The ask is soft

What the signatories actually request: more compute for startups and researchers, investment in shared datasets and evaluation frameworks, and avoiding premature restrictions. No thresholds, no proposed rule, no line anyone commits to defending. The top Hacker News reply on the Microsoft submission went straight there: “I like the message of open weights, and even support it, but I don’t see the point of this letter if there’s no action to back it up.” Another user asked the obvious follow-up: “Now apply that logic to operating systems and office software. Open source Windows and Office when?”

Fair. Nvidia sells chips to whoever trains models. Microsoft rents the inference. Meta, Mistral and Hugging Face ship open weights already. A cheap, plural model ecosystem is straightforwardly good for all of them, and the letter costs its signatories nothing.

It still changes the board. On 12 July, Nathan Lambert wrote that “open models lack the central economic champion to represent the potential downside of action against them.” Thirteen days later they have 35, plus the nearly 200 startups the Little Tech Association organised on 22 July, whose executive director Harry Godfrey asked for “a scalpel rather than a sledgehammer.” Suhail Doshi put the startup case less diplomatically: ban the Chinese models and “there’ll be hundreds of companies that instantly die. It’s great for Anthropic.”

Anthropic and Google remain off the list. Amazon, Apple and xAI too. Anthropic is the one with a coherent reason to stay off, having argued for months that irreversibility is the whole problem with open weights, and it is at least holding a consistent position while the industry organises against it.

Whether the list is still 35 by Monday is the thing to watch.

Sources

• Microsoft, Open Weights and American AI Leadership (24 July 2026), and the Nvidia-hosted PDF with the original 25 signatories
• Internet Archive, snapshot of the Microsoft page, 24 July 2026 21:54 UTC
• CNBC, Nvidia, Microsoft, Meta warn against ‘premature restrictions’ of open-weight models (24 July 2026)
• Politico, Big Tech companies defend open-weight AI models (24 July 2026)
• Politico, Startup founders urge Trump not to shut off Chinese open weight AI (22 July 2026), and the LTA letter PDF
• The Next Web, The AI giants just broke with OpenAI and Anthropic on open weights (24 July 2026)
• Axios, OpenAI and Anthropic unite against China’s open models (22 July 2026)
• Nathan Lambert, 6 months to live for open models (12 July 2026)
• Hacker News, discussion on the Microsoft page and on the Axios report

Related on this blog

• Every Three to Five Months, Someone Asks Washington to Ban Open Models
• Moonshot, Fable and the distillation accusation
• How far behind the frontier are open-weight models on cyber?
• Hugging Face’s breach, and the model that would do the forensics