Paragraph Nine Was the Payload: Nvidia's Open-Weights Letter and Anthropic Alone

Jensen Huang has never posted on X. Not once, through the entire boom that made his company the most valuable on earth. On July 24 he finally did, and spent the post on someone else’s argument: a letter Nvidia signed called Open Weights and American AI Leadership. It has 54 million views.

The letter reads like a civics essay. Open source built the internet, open weights will diffuse AI into factories and hospitals and classrooms, America should lead. Nine paragraphs of that. Then paragraph nine does something else entirely.

The Actual Ask

In shaping this ecosystem, policymakers should be careful not to conflate legitimate model development techniques with misappropriation. Distillation, or 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.

— Open Weights and American AI Leadership, July 24 2026

It goes on to say that unlawful extraction should be handled “through targeted legal and commercial frameworks rather than sweeping restrictions.”

Two days before that letter, on July 22, White House science and technology chief Michael Kratsios accused Moonshot AI of covertly distilling Anthropic’s Fable model to build Kimi K3, alleging Moonshot built infrastructure specifically to evade detection and sourced export-controlled GB300 servers through Thailand. Treasury Secretary Bessent said sanctions and Entity List designation were on the table. Anthropic’s policy head Sarah Heck amplified the accusation on X, calling it IP theft, industrial espionage, and a national security risk.

So the timeline is: Washington starts building a case for restricting open models, and the injury it names is Anthropic’s. Forty-eight hours later, twenty-five companies sign a paragraph arguing the technique at the centre of that case is normal engineering practice.

Everything above paragraph nine is the wrapper. Paragraph nine is the shipment.

The Signatory List Is a Business Model Map

The letter launched with 25 names. By the next day it was 35, and OpenAI had added its own. The full list runs Nvidia, Meta, Microsoft, a16z, Mistral, IBM, Palantir, Hugging Face, GitHub, Cisco, Dell, Mozilla, the Linux Foundation, Replit, Perplexity, Cohere, Fireworks, Nous Research, Prime Intellect, Y Combinator and the rest.

Read it as a values statement and it’s a coalition for openness. Read it as a P&L and it’s a coalition of companies that get richer when models stop being scarce. Chips, hosting, distribution, tooling, applications. If the frontier model is a commodity, every one of those layers captures more of the value. Huang did not leave this to inference. He said it out loud to Axios the same day Kratsios made the accusation:

If there’s great AI, even if it’s open, wherever it comes from, there will be more use. Whenever there’s more use, you’ll have to sell a lot more NVIDIA computers.

— Jensen Huang, Axios, July 22 2026

In the same interview he called the Chinese-backdoor worry a “misconception” and said plainly that “these Chinese models are excellent.” A ban on Chinese open weights is a ban on demand for the hardware they run on.

That is not a gotcha. Self-interest and correctness are independent variables, and the cost argument in the letter is genuinely strong. But it does mean the letter should be read as lobbying, which is what it is, rather than as the industry’s conscience.

It also isn’t abstract. Huang and Amodei have been going at each other in public for a year. Amodei compared selling H200s to China to “selling nuclear weapons to North Korea”; Huang called that “stupid” and “madness,” said he “pretty much disagree[d] with almost everything” Amodei says, and characterised his position as “He thinks AI is so scary, but only they should do it.” Amodei called that an “outrageous lie” and a “bad faith distortion.” The most-followed debut in tech-CEO history is the next move in a running argument between two men.

Where “Anthropic Versus the World” Breaks Down

The framing writes itself and it’s half wrong. Google and Amazon didn’t sign either. Neither did Apple, xAI, or SpaceX, and Musk publicly backed the letter anyway (“This has my full support. Jensen is right.”) without putting either company’s name on it. Cohere signed despite selling closed frontier models. OpenAI, whose entire business is a proprietary model, signed on day two after Altman had already posted that he wants the US to win “both in open source and proprietary models.”

So the clean line that every commoditiser signed and every model-owner abstained stopped being true within 24 hours.

The OpenAI position is stranger still. Two days before the letter, Axios reported that OpenAI and Anthropic were converging in Washington on exactly one thing: warning policymakers about Chinese open-weight models. Then OpenAI signed a letter whose distinctive contribution is a defence of the technique the government says a Chinese lab used. You can hold both positions, but only by reading paragraph nine as being about domestic liability rather than about Kimi.

What survives is narrower and sharper. Anthropic is the only frontier lab that is simultaneously off the list, publicly arguing the other way, and the named beneficiary of the enforcement action the letter exists to stop. Google and Amazon are silent. Anthropic is a party.

And Anthropic said nothing. No response to the letter, no statement, nothing from Amodei, whose standing position is that open source is “a red herring.” Thirty-five companies published a rebuttal to his policy chief’s tweet and the company did not engage.

The Safety Argument, Turned Around

The part of the letter that will age the most interestingly is not the economics. It’s that the coalition took Anthropic’s own vocabulary and pointed it the other way:

Relying solely on closed models is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect.

— Open Weights and American AI Leadership

Concentrating capability behind a few closed models, the letter argues, “results in a small number of single points of failure.” It asks for protections “tied to real and demonstrated harms rather than assuming that closed systems are safer by default.”

That is the safety case for openness, made in the register safety-first labs invented: red teaming, benchmarking, transparency, failure modes. Whether or not it’s right, it removes the assumption that gated release is the cautious default. That assumption has been doing quiet load-bearing work in AI policy for three years.

What This Doesn’t Settle

• The accusation might be true. Nothing in the letter denies that Moonshot distilled Fable. It argues the category shouldn’t be criminalised, which is a different claim, and a convenient one if you’d rather not litigate the specific case.
• The backdoor question is untouched. “Open weights let you inspect the model” is true about architecture and largely untrue about learned behaviour. You cannot read 2.8 trillion parameters and find the trigger.
• Anthropic’s real position is narrower than its critics say. As I covered when this was still hypothetical, Amodei’s concrete asks have been about chips, not banning open models. “Dario wants to outlaw open source” remains mostly his opponents’ framing.
• A letter is not a policy. Thirty-five signatures do not bind Commerce, and the administration has been visibly split on this for weeks.
• Isolated in the industry is not isolated in the country. The Future of Life Institute’s Summer 2026 AI Safety Index ranked Anthropic first overall, leading five of six domains. Polling runs the same way: 65% of Americans say government has done too little on AI, and a July 2026 survey found 57% opposed to blocking state AI rules through a defence-spending rider against 19% in favour. The coalition has the signatures. It does not obviously have the public.

The Window Closes Tomorrow

Three days ago I wrote that Kimi K3 was in a strange position: API-only, weights not yet published, which made it the most bannable Chinese model in existence right up until the moment it became the least. The lever existed only while the file didn’t. That drop is expected July 27.

That’s tomorrow. Read the letter’s timing against that date and it stops looking like a values statement entirely. It’s thirty-five companies jamming a door one day before it shuts on its own, aimed at officials who have roughly one news cycle left to pull a lever that stops working on release.

The Takeaway

• Read paragraph nine first. The openness argument is the delivery mechanism. The ask is a liability carve-out for distillation, filed while a distillation case is being built.
• The letter is lobbying, and lobbying can be correct. Discount the framing, keep the cost argument, notice who profits.
• Anthropic’s isolation is real but specific. Not “against the world.” Against a coalition assembled by the one company whose interests are most directly threatened by the case Anthropic’s own policy chief was promoting.
• If you ship on open weights, none of this changes your stack this week. It changes whether the next Chinese release arrives with an Entity List designation attached.

Huang sat out social media for the entire AI boom and broke the streak for a policy document. That’s the tell. This wasn’t a debut. It was a distribution channel, opened for one specific paragraph, three days before a file lands.