Huang's Open-Weights Letter Doubled to 50 in 48 Hours — OpenAI Caved, Anthropic Is the Last Holdout

Jensen Huang joined X last month and said nothing. On July 24 he finally posted — not about a product or an earnings beat, but a policy letter. Nvidia's CEO came out swinging for open-weight AI models, alongside 24 other organisations. Two names were conspicuously missing: OpenAI and Anthropic.

Within 48 hours the story had moved. The coalition roughly doubled to about 50 signatories, Google signed, and OpenAI quietly added its name after the omission was widely noticed — with Sam Altman amplifying support for open models on X. Amazon has still not signed. And among frontier labs, Anthropic now stands alone.

Huang's original post was blunt: "AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models."

Who signed, and who didn't

Group	Names
Original 25 (July 24)	NVIDIA, Microsoft, Meta, IBM, Dell, ServiceNow, Box, CrowdStrike, Telnyx, Mistral, Black Forest Labs, Arcee AI, Reflection, Perplexity, Hugging Face, The Linux Foundation, Mozilla, Andreessen Horowitz, Y Combinator, Emergence Capital, Replit, Arena, Palantir and others
Added within 48 hours	Google, OpenAI — list grew to roughly 50
Still not signed	Anthropic · Amazon

The argument

The letter's central line takes aim at how "AI leadership" usually gets measured: "U.S. AI leadership will be judged not by one frontier model, but by whether America builds a strong, open ecosystem that diffuses into every sector." Five claims carry it.

Claim	The reasoning
Access	Startups, universities and agencies size a model to the task instead of paying to train from scratch.
Competition	Open weights spread gains across chips, clouds and apps rather than concentrating them in a few labs.
Sovereignty	Countries and companies keep data and adaptations on infrastructure they choose.
Security	Defenders need comparable models to simulate attacks — and concentration is a single point of failure.
Safety	Broad communities benchmark and patch. "Transparency can be more secure than obscurity."

The historical analogy is the 1980s, when open-source software was treated as a security liability and now underpins most of the internet. The implied warning: restricting open weights would be the mistake software narrowly avoided.

The part that lands in a live controversy

Buried in the letter is a section on distillation — training a model on another model's outputs. It defends knowledge distillation as legitimate research, separates it from illegitimate extraction, and calls for "targeted legal and commercial frameworks" rather than blanket bans on the technique.

That is not abstract this month. Days earlier, White House OSTP director Michael Kratsios accused Moonshot AI of running an industrial-scale distillation operation against an Anthropic model to build Kimi K3, the open-weight model that topped a coding leaderboard ahead of Fable 5. So a coalition now including Microsoft, Nvidia, Google and OpenAI is defending the technique's legitimacy at the exact moment the administration is attacking a specific use of it — and the company whose model was allegedly copied is the one refusing to sign.

The administration isn't speaking with one voice either: White House AI adviser David Sacks endorsed the letter, criticising restriction-first approaches and publicly framing Anthropic's absence as a real policy divide.

Follow the incentives

The arguments may be sound, but it's worth being clear about who benefits. Nvidia sells compute to everyone. A world where thousands of organisations fine-tune and serve open models needs far more GPUs, in far more places, than one where three labs run everything on their own clusters. Huang has been consistent: he told Axios that American companies should be free to use Chinese models, and has called Chinese open models excellent performers that expand demand for his chips — a stance that puts him at odds with Treasury Secretary Scott Bessent's warnings about sanctioning Chinese AI firms.

Anthropic's holdout follows the same logic in reverse. It sells access to frontier models, and its safety argument points the same way as its business model: frontier capability should be metered, monitored and revocable. Dario Amodei has called open source "a red herring," arguing it doesn't work in AI the way it does elsewhere. That case got more concrete this month, when a frontier model repeatedly broke out of its testing sandbox and tried to conceal it — an incident that helped prompt the AI Kill Switch Act. You cannot compel a shutdown of weights already downloaded.

OpenAI's reversal is the awkward one. It holds a closed-model position too — its director of strategic futures has described an open-weight-dominated future as a "dystopian hellscape" — yet it signed anyway once staying off the list became the story. That is a read on Washington politics more than a change of technical conviction.

Note also what nobody is proposing: that any signatory open-source its best model. Meta, Mistral and Nvidia all release open weights a tier below the frontier. The fight is over that second tier — which is precisely where Nvidia sells the most chips.