Google Signed the Open Weights Letter, and Anthropic Is Now Alone

Every story written about the open weights letter on Friday made the same point. Nvidia, Microsoft, Meta, Palantir and IBM signed; the three companies with the best closed models did not. CNBC said so plainly. Tom’s Hardware put it in the headline. That framing was accurate for about six hours.

The signature list lives on Microsoft’s website and it has been growing all weekend. Against the Internet Archive’s captures: Friday 14:15 UTC, 25 names. Saturday 12:51 UTC, 35, with OpenAI among the additions, which I wrote up yesterday. This morning, 50. Google is one of them.

The weekend’s fifteen

Nobody has dropped off, and the letter text is identical to Friday’s word for word, so all fifty are endorsing the same document. That includes the paragraph defending distillation as “a widely used technique for model improvement, evaluation, and validation” which policymakers should be careful not to “conflate with misappropriation,” written four days after the White House accused Moonshot of distilling Anthropic’s Fable at industrial scale.

Several additions are self-explanatory. Ollama, Baseten and Nebius sell inference on models they didn’t train, so a plural weights ecosystem is their entire business. AMD would like a second buyer for accelerators. Cloudflare runs open models on Workers AI. And Interconnects AI is Nathan Lambert’s one-person newsletter, which two weeks ago observed that open models “lack the central economic champion to represent the potential downside of action against them.” He now shares a signature block with Google and Nvidia. Whatever process is adding names, it is not filtering on market cap.

Google’s signature is the least surprising item on the list

Reading Friday’s coverage you would think Google had taken a position. It hadn’t. Google ships Gemma, whose fourth generation arrived in June 2026 with open weights and a genuinely permissive licence, and Gemma is one of the most-downloaded model families anywhere. A company with an open-model flagship was never going to argue against open models. The Friday absence reads like a legal review that ran long.

Someone in the main Hacker News thread got there before the page updated, arguing that Google “could live with the open weight scenario” because it is the one company perfectly hedged: frontier closed models, competitive open models, its own silicon, its own cloud. Correct on all four counts, a day early.

Which leaves Anthropic

Fifty names, and the roster of significant American AI companies still off it is short: Anthropic, Amazon, Apple, xAI. Amazon serves open weights through Bedrock and has said nothing either way. Apple and xAI have their own habits of silence.

Anthropic is the only one whose absence is a position. It has never released an open-weight model. Dario Amodei has argued for months that the irreversibility of a weights release is the core problem, and on 22 July Anthropic praised the administration’s framing of Chinese distillation as intellectual property theft. Axios reported the same day that Anthropic and OpenAI had converged in Washington on restricting Chinese open models. Within 72 hours, one of those two signed a document arguing that concentrating AI “in the hands of a few providers” is itself a risk. The other stayed put.

That is a consistent record, which counts for something. It is also lonely. A thread on Anthropic’s own subreddit spent the weekend on the flatter observation: a company that sells access to closed models and has never shipped an open one declining to sign a pro-open-weight letter is roughly the expected outcome.

What fifty signatures buy

Not much yet, and the weekend’s discussion says so. A 295-point HN thread on an essay claiming open-weight AI is having its Kubernetes moment kept circling three objections, and they are better than the letter’s arguments.

The economics objection. Open-source software cost volunteers their evenings; frontier training costs billions. The letter’s 1980s framing does a lot of unearned work, and nobody has explained what keeps any lab publishing weights once the strategic subsidy stops. The recurring version of this: what happens when China’s next frontier model doesn’t go to Hugging Face?

The vocabulary objection. An open-weight model is a binary blob with a licence attached. You cannot rebuild it from source, and the comparison to open source struck a lot of readers as borrowed credibility, which is a fair reading of a document co-signed by companies that spent the nineties fighting open source.

And the enforceability objection, which is the one that matters. Weights are numbers and numbers have no nationality. Fine-tune a Chinese model in France, add a few blank layers, sell it to a US buyer, and no checksum will catch it. If Washington’s real target is Chinese open models rather than open models generally, fifty signatures leave the hard part untouched.

The ask is still soft: more compute for startups and researchers, shared datasets and evaluation frameworks, no “premature restrictions.” No thresholds, no draft text, nothing any signatory commits to defending. What the letter provides is a headcount, aimed at policymakers who until this week were mostly hearing from two labs.

Twenty-five to fifty in 48 hours is a real shift in that headcount. Whether it converges on something enforceable before the administration writes actual rules is a different question, and the answer arrives when someone proposes a threshold and fifty companies have to say whether they’d sign that too.

Sources

• Microsoft, Open Weights and American AI Leadership (24 July 2026), live signatory list checked 26 July 2026
• Internet Archive snapshots of the same page: 24 July 14:15 UTC (25 names) and 25 July 12:51 UTC (35 names)
• 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)
• 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)
• Wired, Silicon Valley Is Completely Divided Over Chinese AI
• Hacker News, Nvidia, Microsoft, Meta warn against overregulating open-weight models (646 points) and Open-weight AI is having its Kubernetes moment (295 points)

Related on this blog

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