Nvidia Rallied the Industry Behind Open Weights. Then OpenAI Joined Anyway.

On July 24, Jensen Huang posted on X for the first time in his life. He used it to publish a three page letter called "Open Weights and American AI Leadership," signed by 25 companies. Within a day the signatory count had doubled to 50, and the new names included OpenAI and Google, both of whom had been conspicuously absent from the original list. The gap between those two facts is the most interesting thing about the whole document.

TL;DR

• The letter argues that restricting downloadable models would cost the US its AI lead, and draws a line back to open source software in the 1980s.
• Original 25 signatories included Nvidia, Microsoft, Meta, IBM, Dell, Palantir, Hugging Face, Mistral, Perplexity, a16z, the Linux Foundation, and Y Combinator.
• OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, Block, and Ollama joined within roughly a day. Anthropic and Amazon did not.
• It landed one day after a bipartisan bill and amid reports the White House was weighing restrictions on Chinese open weight models. This is a lobbying document with a civics framing.
• Its central move is separating open weights from distillation. Treasury Secretary Scott Bessent had drawn the opposite line three days earlier.

What the letter argues

The core claim is short. If the US wants to keep its lead in AI, it should keep supporting open, downloadable models rather than restricting them. The letter defines open weight models as systems anyone can download, inspect, modify, and run on their own hardware, and compares the moment to the rise of open source software in the 1980s, framing that earlier openness as part of what built American tech dominance in the first place.

It pushes back specifically on broad restrictions aimed at Chinese open weight models, arguing that technology theft should be handled with targeted legal and commercial tools rather than category-wide bans. And it separates two ideas that had been getting fused in public debate: open weights themselves, and distillation, the technique some Chinese labs are accused of using to copy American model behaviour. The letter's position is that these are different problems and regulating one will not fix the other.

That separation is the load-bearing argument. Everything else in the letter is context.

The absence that got more coverage than the letter

Nearly every outlet covering the launch led with the same detail: OpenAI, Anthropic, and Google, the three companies with the most invested in closed frontier models, were all missing from the original signatory list. In a letter about American AI leadership, that is a conspicuous hole.

It closed fast. Within about a day the list doubled to 50, adding OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, Block, and Ollama. Forbes tracked the additions as they landed.

Anthropic and Amazon stayed off. Anthropic's absence turned into its own news cycle and eventually its own response from Dario Amodei, which is a separate argument worth reading on its own terms. Amazon has said less.

The speed of the reversal tells you what the letter actually is. Companies do not join a policy coalition inside 24 hours because the argument persuaded them overnight. They join because being visibly absent became more expensive than signing.

The timing was not a coincidence

This did not land in a vacuum. It came one day after a bipartisan bill was introduced in Congress, and amid reports that the White House was actively weighing new restrictions on Chinese open weight models. The administration has separately accused China of large-scale theft of AI technology from American companies, Anthropic among them.

Read against that backdrop, the civic framing about American leadership is also, plainly, a lobbying document aimed at rules that had not been finalised yet. That is not a criticism. It is what industry coalitions do, and the letter is more candid than most.

It is worth being clear-eyed about incentives on the list too:

Signatory type	Examples	Commercial interest in open weights
Chipmakers	Nvidia, AMD	More teams fine-tuning their own models means more hardware sold
Open weight publishers	Meta, Mistral	Already ship open models; restrictions hit their strategy directly
Hosting and distribution	Hugging Face, GitHub, Cloudflare, Ollama	Their platforms exist to distribute and run these models
Venture firms	a16z, Y Combinator	Portfolio companies cannot afford frontier API bills
Closed frontier labs	OpenAI, Google	Least aligned with the argument, joined last

None of that makes the arguments wrong. Nvidia can be commercially motivated and correct at the same time. But a letter where nearly every signer profits from the outcome is evidence about industry alignment, not about the underlying policy question.

Why distillation became the word of the week

If you read any AI coverage that week you saw the word distillation more than you expected, and the conversation got muddled fast. It is worth being precise, because the imprecision is doing real work in the policy debate.

Distillation is not reverse engineering a model's weights. Nobody is extracting the billions of internal parameters that make up a trained model. Those stay locked inside whoever trained it.

What actually happens is closer to an apprenticeship. A smaller student model is trained by sending it enormous volumes of prompts and teaching it to imitate the answers a bigger teacher model returns. Do that at sufficient scale, millions of queries with careful study of the responses, and the student ends up reproducing a meaningful share of what the teacher can do, without ever touching the teacher's internals. It reconstructs behaviour from outputs rather than lifting code or weights, which is close in spirit to reverse engineering while being a technically different process, and legally a much murkier one.

That murkiness is why the fight got heated. In a June 10, 2026 letter to senior members of the Senate Banking Committee, Anthropic alleged that operators connected to Alibaba's Qwen lab ran 28.8 million exchanges with Claude through roughly 25,000 fraudulent accounts over a 44-day window from April 22 to June 5, structured specifically to extract training signal. Anthropic called it the largest known distillation campaign ever run against a commercial model.

Three days before Huang's letter, Treasury Secretary Scott Bessent drew the government's line in public:

"We support open-source AI and the innovation it unlocks. But open source is not open season on American IP. When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table."

He also said officials had found watermarks from US large language models inside several Chinese AI systems, which is a more concrete evidentiary claim than anything in the letter and worth watching as it develops.

So the letter's attempt to separate open weights from distillation was not an abstract taxonomic exercise. It was a direct response to a framing the Treasury Secretary had already committed to in public. The letter's logic is that a model can be open weight without any distillation involved, and a closed proprietary model can use distillation freely, so restricting open weights as a category does not target the behaviour anyone is actually upset about. Regulate the wrong thing, and the actual bad actors keep operating.

What this actually changes

Nothing about this letter changes any law. It is 25, then 50, logos at the bottom of a PDF.

What it signals is that a large chunk of the industry wants regulators to treat openness as the thing worth protecting, right as Congress and the White House decide what the rules will be. Running against that is at least one major lab arguing the real problem is narrower and more specific, and that a general defence of open weights talks past the actual harm.

For anyone building on top of this rather than lobbying about it, the practical exposure is what matters. If restrictions land on Chinese open weight models, any architecture that routes meaningful workload to them needs a migration path, and the cost profile of that migration is not small. I have written about the parallel problem in AI-native pricing, and about the security posture question in the shadow AI governance crisis. Most companies genuinely do not know which models their production systems currently depend on. That is a worse position to be in than having a policy opinion.

Frequently Asked Questions

What is the Nvidia open weights letter?

A three page letter titled "Open Weights and American AI Leadership," published July 24, 2026 by Jensen Huang in his first ever post on X. It argues that the US should keep supporting open, downloadable AI models rather than restricting them, and that concerns about technology theft should be addressed with targeted tools instead of category-wide bans.

Who signed it?

The original 25 included Nvidia, Microsoft, Meta, IBM, Dell, Palantir, Hugging Face, Mistral, Perplexity, Andreessen Horowitz, the Linux Foundation, and Y Combinator. Within about a day the count doubled to 50, adding OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, Block, and Ollama.

Why did OpenAI, Google, and Anthropic skip the original list?

All three have the most invested in closed proprietary frontier models, so a letter defending openness sits awkwardly with their commercial position. OpenAI and Google joined within roughly a day once the absence became the story. Anthropic and Amazon have stayed off.

What is an open weight model?

A model whose trained parameters are published so anyone can download, inspect, modify, and run it on their own hardware. It is distinct from open source software, since the training data and code are often not released, and distinct from an open API, where you can call the model but never possess it.

What is distillation and why is it controversial?

Distillation trains a smaller student model on huge volumes of a larger teacher model's outputs until the student reproduces a meaningful share of the teacher's behaviour. No weights are copied. That is exactly what makes it legally murky: it reconstructs capability from outputs rather than taking anything the law clearly recognises as stolen property.

What has the US government actually said?

Treasury Secretary Scott Bessent said the administration supports open-source AI but that "open source is not open season on American IP," and that covert, industrial-scale distillation crossing into IP theft could bring sanctions and Entity List designations. He also said officials found watermarks from US models inside several Chinese AI systems.

Does the letter change the law?

No. It has no legal force. It is a signal of where a large part of the industry wants policy to land, published while Congress and the White House are still deciding, which is precisely when signals are worth sending.