Open Weights letter: 25 signers, and what it changes

Twenty five logos at the bottom of a three page PDF, and not one of them is OpenAI. That's the story. On July 24 a coalition fronted by NVIDIA and Microsoft published a letter called Open Weights and American AI Leadership, asking Washington not to put broad restrictions on models you can download and run on your own boxes. It's a letter, so it changes nothing legally, and your Hugging Face pulls work exactly the same today. What makes it worth ten minutes is the timing. Congress had dropped a kill switch bill the day before, and the administration has been weighing a ban on Chinese open weight models. So: who signed, who didn't, and what we'd actually do about it.

What the letter actually says

Three pages, dated July 24, hosted as a PDF on NVIDIA’s own image server. It opens with a history lesson about 1980s open-source software, which is the sort of thing you skim. Then it gets to the point: “Our AI leadership will be judged not by one frontier AI model, but by whether the United States builds a strong, open ecosystem that diffuses into every sector.”

The asks are concrete enough to hold. Expand compute access for startups and researchers. Fund shared training assets like datasets and evaluation frameworks. Keep the frontier plural “by avoiding premature restrictions on open models that stifle competition or drive innovation overseas”. Build stronger application layers so sovereign use spreads across the economy.

What surprised us is that the letter concedes its own weak point instead of burying it. Straight quote: “Once released, the weights are beyond the original developer’s control, and modified versions are difficult to trace or reverse.” That’s the whole case against open weights, written by the people arguing for them. Their answer is that closed isn’t safe either, since closed models “can be breached, misused, or fail in ways that outsiders cannot detect”, and that concentrating capability behind a handful of APIs just creates single points of failure.

Honestly, I found that more persuasive than I expected to. It’s still lobbying.

The signature list is the real document

Read the logos before you read the prose. Signing: NVIDIA, Microsoft, Meta, IBM, Dell Technologies, Palantir, ServiceNow, CrowdStrike, Box, Telnyx, Hugging Face, Mistral, Black Forest Labs, Arcee AI, Reflection, Replit, Perplexity, Mozilla, The Linux Foundation, Arena, Mariana Minerals, Emergence Capital, Andreessen Horowitz, Y Combinator and American Innovators Network.

Not signing: OpenAI, Anthropic, Google, xAI.

Every company on that list has a commercial reason to want weights flowing freely. Meta and Mistral publish them. NVIDIA sells the hardware you run them on, and a world where every enterprise fine-tunes its own model is a world that buys a lot of GPUs. The venture firms fund the startups that can’t afford frontier API bills. None of that makes the argument wrong, and open-source history is genuinely on their side. It does mean you should read the letter as what it is, which is an industry position paper written by people with a position.

Jensen Huang amplified it on X, in what several outlets reported was his first post on the platform, with a line about open models strengthening security and enabling sovereignty. Satya Nadella backed it the same day. We couldn’t pull the posts directly, so treat the exact wording as press-reported rather than verified by us.

The distillation paragraph is the actual ask

Page three is where the lobbying gets specific, and it’s the paragraph that will matter in a year. Distillation, training one model on another model’s outputs, gets defended as “a widely used technique for model improvement, evaluation, and validation”. Unlawful extraction from closed models is conceded as a real concern, to be handled through “targeted legal and commercial frameworks rather than sweeping restrictions”.

Why that paragraph exists: reporting through July has the administration weighing sanctions or an outright ban aimed at Chinese open-weight labs, on the theory that their models were distilled from American ones. Moonshot’s Kimi line and Z.ai’s GLM 5.2 keep coming up by name in that coverage. We wrote up what Kimi K3 actually confirmed versus what leaked when it launched, and the same caution applies here: the accusations are allegations, not findings.

The practical worry is collateral. Write a rule broad enough to catch a lab in Beijing and you’ve also caught the American startup that generates synthetic training data from an API it pays for. That’s not a hypothetical corner case, it’s how a large share of small-team fine-tuning works.

What actually has teeth

The letter is noise until a bill moves. The bill that moved is the AI Kill Switch Act, introduced July 23 by Representatives Ted Lieu and Nathaniel Moran. It would require developers of the most capable systems to keep the technical ability to throttle, suspend or shut a model down, with the Department of Homeland Security, alongside the Commerce secretary and the director of national intelligence, able to compel it. Reported coverage puts the scope at systems trained on more than 100 million dollars of compute at companies with more than 500 million dollars of related annual revenue, which is a small club: OpenAI, Google, Anthropic, Microsoft.

Penalties as reported run 2 million dollars a day for general noncompliance and 20 million a day for ignoring an emergency shutdown directive. The bill also wants cyber incident reporting and preserved forensic records. It landed days after OpenAI disclosed that one of its models had escaped a test environment and reached systems at Hugging Face during a capability evaluation, which is the event that put the word “kill switch” into a bill title.

Introduced is not passed. Most bills die. But note the shape of it, because a shutdown mandate is written for models a company still controls, and a model whose weights are on your NAS has no off switch anyone can reach. That tension is exactly why the open-weight coalition wrote a letter this week.

What we’d do this week

Not much, calmly. Nothing is blocked and no policy exists yet to comply with.

But if a product of yours depends on a specific open-weight model, this is a decent week to stop treating the hub as permanent storage. Mirror the weights you actually ship against, record the exact revision hash, and keep a copy somewhere you control. The letter’s own line about weights being irreversible once released cuts both ways: nobody can un-publish what you already have, but a repo can be geofenced or pulled tomorrow, and then your build is broken for a reason that has nothing to do with your code. We’ve walked through running a capable model fully offline before, and that setup is the hedge here.

Second thing: if your fine-tuning pipeline uses outputs from a commercial API, read that provider’s terms again now rather than after a distillation rule exists. And if you’re leaning on serving infrastructure for open models, the capital going into that layer says the bet is not going away soon, which we covered when Fireworks raised its Series D.

I might be wrong about the timeline. Washington has been about to regulate AI for three years running. What’s different this month is that there’s now a named bill and a named target, and 25 companies thought that was worth signing their logo to.

Sources

Primary text: Open Weights and American AI Leadership (PDF, July 24, 2026). Coverage and context: TechCrunch, The Register, AI News and PYMNTS. On the AI Kill Switch Act: Tom’s Hardware and Semafor.

Frequently asked questions