"Until 2026, I hadn't posted anything on X, which tells you how introverted I am—probably the last person on Earth to do it."
That was Jensen Huang explaining, in a Y Combinator interview, why his debut on the platform arrived so late. Then the punchline: "But what I had to say was too important to me, too important to the whole industry, too important to the world. So I overcame my shyness."
What he posted was not a leather-jacket selfie. It was a three-page policy PDF: "For my first post, I'm sharing a letter @NVIDIA signed on why open models matter."
Last week in this newsletter, the open-weight story was about token share—Chinese models climbing to 48% of OpenRouter traffic while U.S. models fell to 32%. This week it is about something more consequential: organized political lobbying to stop Washington from restricting open weights at all.
And the tell is in the arithmetic. The "Open Weights and American AI Leadership" letter, shepherded by Microsoft and dated July 24, launched with 25 signatures. Within 48 hours it hit 50. By the following Tuesday, Nvidia's own servers listed 77. Simon Willison's count reached 235.
Manifestos do not gather signatories like that. Coalitions do. A conviction has 25 names on day one; a market position has OpenAI and Google quietly queuing up once they see which way the wind is blowing.
Read the composition of the coalition and the motive becomes legible. Every signatory has identified open weights as the one layer where it holds no moat—and therefore the one layer whose openness costs it nothing while maximizing revenue everywhere else.
Nvidia makes GPUs; every downloaded, fine-tuned, on-premise model runs on its silicon. Meta distributes Llama because doing so eliminates API dependency on rivals. Microsoft backs open weights because Azure hosts them. As Forbes put it, this is "the one layer where almost no signatory keeps its moat."
The proprietary advantages stay exactly where they are. Nvidia controls CUDA. Microsoft controls Azure. Meta controls the social graph. "Openness," in this configuration, is not a concession—it is the precise market shape that turns other people's models into demand for their infrastructure.
This is not hypocrisy. It is rational coalition-building, dressed in the vocabulary of public interest.
And the letter's public-interest argument is, to be clear, mostly correct. "Relying solely on closed models is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect." That sentence is defensible on its merits—all the more so after Hugging Face used a Chinese open model to contain an attack that closed American models blocked it from even analyzing.
The question is not whether the letter is right. It is why it appeared now, in this form, signed by exactly this set of companies.
The sharper answer is that the letter is theater. The lobbying disclosures are the mechanism—and the two companies most conspicuously missing from the PDF are the ones actually writing checks.
Federal disclosures filed the same week show AI firms breaking lobbying records in the first half of 2026. Anthropic nearly tripled its spending to $3.53 million, already past its entire 2025 total. OpenAI nearly doubled to $2.22 million. Together, the two closed-model labs spent $5.75 million in six months.
A Hacker News commenter captured the scale precisely: "Never ceases to amaze me how cheap lobbying is. That's pocket change for these companies." That is the point. Against Anthropic's $2 billion round and OpenAI's $6.6 billion, seven figures of lobbying is a rounding error that buys regulatory architecture.
Against Anthropic's $2 billion round and OpenAI's $6.6 billion, seven figures of lobbying is a rounding error that buys regulatory architecture.
Here is the uncomfortable math. The two most prominent absent names on the letter—Anthropic, and Amazon, its largest investor—are spending more in Washington than most of the letter's signatories combined. The PDF has 235 names. The disclosures have dollars. In Washington, dollars vote louder than names on a PDF.
And what those dollars reportedly buy is the opposite of the letter's position: representatives from both OpenAI and Anthropic have urged federal figures—including Treasury Secretary Scott Bessent and White House adviser Michael Kratsios—to restrict accessible AI models.
Jeremy Howard, the fast.ai founder and one of the open-source community's most credible voices, distilled the contradiction so it could not be misread: "Hey sir. We are not asking you to open source Anthropic. Just don't lobby the government to shut down open source."
That reframes the whole fight. This was never a licensing-philosophy debate. It is a coalition war over who drafts the regulatory framework—and the drafting is happening in meetings the letter's signatories are not in.
Watch how the two sides talk when the cameras are on. Huang, per Senator Mark Warner's readout, pitched open source directly; at GTC he had already rejected the binary—"proprietary versus open is not a thing. It's proprietary and open." The coexistence framing is doing real work: it lets Nvidia sell hardware to both camps while sounding above the fray.
Dario Amodei plays the other position with matching care. He warned about authoritarian regimes building models more powerful than America's, and about misuse for cyber or biological attacks, and called for "a crack down on industrial-scale distillation operations"—while insisting "Anthropic has never advocated for a ban on open-weights models."
Both men have found the same escape hatch: never say "ban," never say "decelerate." When Sam Altman went to the Hill, a reporter asked whether he'd discuss decelerating AI. His answer: "I wouldn't use the word deceleration, but we've talked about the need to pace it."
Pace it, don't decelerate. That is now the axis on which the framework turns—not open versus closed, but safety versus leadership, with each lab positioning its business model as the one that serves both.
Which brings the fight to legislative text. Representatives Trahan and Obernolte have introduced the FRONTIER Act, a federalization of the SB 53/RAISE framework: safety frameworks, model reports, incident reporting, catastrophic-risk definitions, a new emergency shutdown authority, enforcement lodged with an Under Secretary of Commerce for AI Security.
It also permanently preempts the states. That is a steep price, and Zvi Mowshowitz's read is that you'd need a great deal in return to accept it.
Here is the structural trap worth watching. The first analyst to run the bill concluded it does not kill open source—it imposes strong, expensive requirements on closed frontier labs and essentially none on open models.
And yet, in the coalition's framing, a bill that imposes zero hard limits on open models can still be branded as "killing open source." You cannot write a law that both permits anyone to upload any "open" model to Hugging Face regardless of consequences and addresses catastrophic cyber risk. Pick one.
So the question that decides the next chapter is not philosophical. It is whether the FRONTIER Act's permanent state preemption survives the moment the coalition realizes a bill that barely touches open models is still, in their mouths, a bill that kills open source.
The letter has 235 names. The framework has one author, still to be determined. Watch the disclosures, not the PDF.