On July 24 Jensen Huang made the first post of his life on X. He used it to share “Open Weights and American AI Leadership,” a letter urging Washington not to restrict downloadable AI models. Nearly every outlet that covered it reported the same striking detail.
At that time, OpenAI had not signed. That is no longer true.
Huang’s X post carried the original 25 signatures.
The letter now carries 50, exactly twice as many. OpenAI is among the 25 names added, along with Google, AMD, Cisco, Cloudflare, GitHub, Block and Ollama.
Anthropic and Amazon are absent from every version, and that pairing is the most interesting thing missing from the list.
How The Open Weights Letter Doubled in a Day
The sequence matters more than the roster. This was not a coalition assembled over months and unveiled when it was complete. It went out with 25 names, Huang’s post cleared 11 million views, and the signatures kept arriving while the coverage was still being written.
Roughly two dozen companies joined after publication. Two of the hosted copies carried different counts on the same afternoon.
That tells you something about how technology policy gets made now.
The document was the easy part. The distribution was the event, and the signatures followed the distribution. Any company weighing whether to sign a public letter should read this as the new tempo: the window between a position becoming visible and becoming crowded is now measured in hours.
The Two That Did Not Move On Open Weights
Amazon and Anthropic are absent from the list today July 25, 2026.
Amazon is Anthropic’s largest investor, and Anthropic trains and serves on Amazon’s Trainium silicon among other hardware. Anthropic just took over the #1 position in the Enterprise. Google, also an Anthropic backer, signed anyway, which makes the Amazon alignment the more specific one.
Neither has explained the absence, and it would be a mistake to assume the reason. The readings run from plain commercial interest, since Anthropic sells closed frontier access, to a substantive safety position that weights once released cannot be recalled. Anthropic has argued the second in public for years. Both can be true at the same time.
What is not in doubt is the shape.
The coalition now spans chipmakers, cloud providers, security vendors, application companies and the venture firms behind them. Amazon and Anthropic are the absences that matter, because both sell frontier AI capability and both have a direct stake in how this is decided. If you buy AI from either, that is a question worth raising in your next review, not because they are wrong, but because you are entitled to know your vendor’s position on a policy that will shape what you can run.
What the Open Weights letter settles for your business
Sovereignty is the word carrying the most weight in the letter, and most coverage reads it as a trade term. For a business it means something plainer. Can you keep running if your vendor changes the deal, and can you prove what your systems did.
Open weights answer the first half. They do not touch the second. It comes down to four questions.
Can you look inside the model and change it? That is what open weights buy you. You download it, examine how it works, and retrain it on your own data.
Can you run it? Weights are inert without chips, and chip access is set by export policy and allocation queues. Nvidia, which organized the letter, sells that layer.
Where does your data go? Running a model yourself turns data residency from a contract clause into an architectural fact.
Who is acting in your name? When an AI agent moves money, signs something or opens a customer record, can you prove it was authorized and reconstruct what happened afterward? Open weights leave this one untouched.
The Question The Letter Leaves Open
That last one is what shows up in an audit. A model you can read line by line still cannot tell you whether the agent that approved a payment last Tuesday was yours, or whether the version running in your Singapore subsidiary is the one your security team signed off on in January.
The letter concedes the mechanism. It admits that once weights are public the developer loses control and altered copies are hard to trace. Its answer is that enough outside researchers will catch problems. That may hold for the industry over years. It is not what a CIO needs on the day a regulator asks who checked this specific copy, when, and against what standard.
What to do about Open Weights
Whatever Washington decides, the letter describes a stack most companies are not architected for.
Five moves worth making this quarter.
- Map model dependency by workload. Name the processes that stop if one vendor changes pricing, terms or availability. Most leadership teams cannot answer this, and the exercise usually surfaces more exposure than expected.
- Put one real workload on an open weight model, with a real service level, not a pilot. A fallback you have never run is not a fallback, and it is your only credible position in a renewal negotiation.
- Price the liability transfer before you self-host. Running weights yourself means inheriting the safety filtering, red teaming and audit obligations your vendor carried. Ask legal and risk to put a number on it.
- Decide agent identity architecture separately from model choice. Provenance, authorization and audit trail are not properties of the model. Build them once and they survive every model swap you make.
- Write the restriction scenario now. If open weight access narrows in 2027, name which roadmap items break and what substitution costs. Do it while it is a planning exercise rather than a fire drill.
Watch the signatory list rather than the headlines.
It doubled in a day and it will move again, and each name tells you something about where that company thinks the money goes.
The harder work is yours either way. Weights you can read are a start. They are not the same as agents you can prove. Consider the Open Weights news carefully.
