25 Companies Signed the Open Weights Letter. The Story Is the Three That Did Not.

Jensen Huang has run NVIDIA for 33 years without an X account. Today he opened one, and his first post was not a product, a benchmark, or a photo from a keynote. It was a PDF: a joint letter titled “Open Weights and American AI Leadership,” signed by 25 companies and organizations, hosted on NVIDIA's own image server.

The coverage has settled on the obvious frame. Huang defends open source. NVIDIA, Microsoft, and Meta warn Washington against premature restrictions. All true, and all less interesting than what the document actually is.

I read the letter twice. The first read is a competent, genuinely well-argued case for open-weight models. The second read is a map of who gets richer when frontier models stop being scarce. Those are not the same document, and the second one is the one worth your time.

Every Signatory Sells a Complement

Here is the full signatory list, because the shape of it is the argument: American Innovators Network, Andreessen Horowitz, Arcee AI, Arena, Black Forest Labs, Box, CrowdStrike, Dell Technologies, Emergence Capital, Hugging Face, IBM, The Linux Foundation, Mariana Minerals, Meta, Microsoft, Mistral, Mozilla, NVIDIA, Palantir, Perplexity, Reflection, Replit, ServiceNow, Telnyx, and Y Combinator.

Sort them by what they actually sell and the coalition stops looking like a movement and starts looking like a supply chain.

Not one signatory's primary revenue comes from metered access to a closed frontier model. That is not a coincidence and it is not a scandal. It is how policy coalitions work. But it means the letter should be read as an economic filing, not an ethical one, and the strongest passages are exactly the ones where interest and argument line up.

The Absences

OpenAI did not sign. Anthropic did not sign. Google did not sign. Between them they own the models sitting at the top of every capability index, and all three sell access by the token.

The letter never names them. It does not have to. Read this sentence and tell me who it is about: relying solely on closed models “is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect.” Then the turn, which is the sharpest line in the document. Concentrating advanced AI behind a few closed models “results in a small number of single points of failure, weakens competition, and leaves critical technology in the hands of a few providers.”

That is a safety argument repurposed as an antitrust argument, aimed at three companies that were invited to the industry consensus and are instead its subject. Microsoft signing it is the detail I keep turning over. Microsoft is OpenAI's largest partner and just put its name to a document arguing that OpenAI's market structure is a systemic risk.

The Distillation Paragraph Is the Real Payload

Now the part that has gone almost entirely unquoted, and the reason I think this letter was filed this week rather than any other week. It is on the last page, after the policy asks, where nobody looks.

Policymakers, the letter says, “should be careful not to conflate legitimate model-development techniques with misappropriation.” Distillation, meaning the practice of using one model's outputs to help train or improve another, is “a widely used technique for model improvement, evaluation, and validation.” It reflects “a long tradition of learning from, building upon, and improving existing technologies.” Unlawful extraction from closed models is a real concern, but it should be handled through “targeted legal and commercial frameworks rather than sweeping restrictions.”

Two days ago, White House OSTP director Michael Kratsios accused Moonshot AI of distilling Anthropic's Fable model to build Kimi K3. Anthropic endorsed the statement. Treasury Secretary Scott Bessent raised sanctions. We covered why that charge is the weaker half of Washington's case in the distillation charge is the weak half.

So the sequence is: the administration calls distillation theft on Tuesday, and on Friday NVIDIA, Meta, Microsoft, Mistral, IBM, and Palantir sign a document calling it a legitimate technique with a long and honorable tradition. Whatever else this letter is, that paragraph is a pre-emptive strike against a rule nobody has written yet.

The self-interest is not subtle. Restricting distillation would hurt every open-weight lab that improves models by training on outputs, which is most of them, and it would hand closed labs a legal weapon against exactly the competitors this coalition is built from. It is also, as an intellectual matter, largely correct. You cannot call the technique standard practice when an American lab uses it and industrial theft when a Chinese one does. That argument was always going to arrive. It arrived on page three of a PDF about American leadership.

The Sovereignty Claim Has a Hardware Bill

The letter's cleanest promise is control. Open weights let organizations “control their own data, evaluate and adapt models to their own needs, and deploy them wherever their business requirements demand.” True in principle. The word doing the quiet work is “deploy.”

Downloadable and runnable are different claims, and the gap between them is measured in accelerators. We refreshed our open-weights deployment catalog this morning, and these are the entry costs for the models this letter is implicitly about, at the cheapest quantization where each is still worth running.

Read the last row against the first. Mistral Medium 3.5 is the letter's thesis working: 128 billion parameters, 77.6 percent on SWE-Bench Verified, one GPU, a license you can ship on. Kimi K3 is the thesis straining. A permissive license on a 2.8 trillion parameter model buys a hospital or a county government nothing at all, because the sovereignty on offer costs seven figures of silicon to exercise.

Which is the part where you notice who organized the letter. NVIDIA is entirely sincere that open weights enable sovereignty, and NVIDIA sells the sovereignty. Both things are true at once. We made the narrower version of this argument about GLM-5.2 and the difference between open weights and independence, and this letter is the strongest evidence yet that the industry knows the gap exists and would rather talk about licenses than about racks.

What They Actually Asked For

Strip out the framing and there are four concrete asks, which is four more than most industry letters manage:

Expand compute access for startups and researchers. Fund shared training assets, meaning datasets, tools, and evaluation frameworks. Keep the frontier plural by avoiding premature restrictions on open models that would “stifle competition or drive innovation overseas.” And do not treat distillation as misappropriation.

The first two are subsidy requests and will be read that way. The third is the one with teeth, because “premature” is doing enormous load-bearing work and nobody in the letter defines it. The fourth is the live one.

The best paragraph in the document, and the one I wish had been the headline, is the economic case. Open weights let organizations “match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient, specialized models everywhere else.” That is a description of routing, and it is the correct architecture. It is also what we have been telling builders to do for a year, and it does not require a single policy change to start doing.

Our Take

The letter is right about the thing it is arguing and it is not a neutral document. Both of those can hold. Concentration of capability inside three closed labs really is a systemic risk, transparency really does aid defense, and a plural frontier really is better for everyone downstream. The 25 companies making that case also happen to own every layer of the stack that gets more valuable when models stop being the scarce part.

What actually changed today is that the open-weight position now has an organized industrial lobby with NVIDIA's balance sheet behind it, at the exact moment Washington is drafting rules and accusing a Chinese lab of the technique this letter defends. The policy fight over open models just stopped being a debate among researchers and became a contest between two coalitions with money on the table.

For builders, none of this changes the work. Put a routing layer between your product and any single lab. Benchmark on cost per task, not sticker price. Check the license before you check the benchmark, because MIT and “community, restricted” are not the same product no matter how the press release reads. And size the hardware before you believe the sovereignty pitch: our deployment catalog and models tracker exist so that number arrives before the invoice does.

Three things to watch. Whether OpenAI, Anthropic, or Google answers publicly or lets the letter stand unchallenged. Whether the distillation paragraph shows up in the August White House framework, in either direction. And whether any of the 25 signatories funds the unglamorous half of their own argument, which is not more open models but cheaper ways to run the ones that already exist. The letter asks for compute access for startups. The company that convened it is the company that prices it.