25 Tech Giants Just Signed a Letter Defending Open AI. The 3 Biggest Labs Are Missing.

Jensen Huang has run Nvidia for more than three decades without once posting on X. On Friday morning he finally did, and he spent the debut on a policy document.

"For my first post, I'm sharing a letter NVIDIA signed on why open models matter," he wrote, repeating the message on LinkedIn. "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."

By the afternoon the post had cleared eleven million views. Satya Nadella had promoted the same document hours earlier. The letter, titled Open Weights and American AI Leadership and hosted on Microsoft's corporate responsibility site, carries 25 signatures. Most of what it argues will sound familiar to anyone who has followed this fight for the past two years. The signature page is where the new information is.

The case

The letter starts in the 1980s. Open-source software won an argument that once looked unwinnable, the story goes, and code that anyone could read and modify now runs most of the internet along with systems used by the U.S. military and federal research agencies. What open source produced, the signatories write, was a shared base of knowledge that American engineers have built on ever since.

Their claim is that AI has arrived at the same fork. American leadership, they argue, will be measured by whether the country builds an open ecosystem that spreads into every sector, and not by which company holds the single best model at any given moment. Four arguments follow.

The first is about access. Startups, universities, hospitals and public agencies can build on advanced models without training one themselves or paying frontier rates for routine work. The economic version of this claim is stronger than the usual pitch: the signatories say that matching model size to task is what will keep AI affordable once usage reaches billions of daily operations. Save frontier capability for frontier problems and run something smaller everywhere else.

The second is about competition. When many organizations can adapt and deploy advanced models, rivalry spreads across chipmakers, clouds, applications and services rather than settling among a few firms at the top.

The third is about customer control. Enterprises investing heavily in AI want assurance that they will not be locked into one vendor or lose the capabilities they build. Open weights let them keep their data in house, adapt models to their own needs and deploy wherever their requirements point.

The fourth argument is about security, and it is the one likely to draw fire. The letter concedes the danger up front. Once weights are released they are outside the developer's control, and modified versions are hard to trace or reverse. But it rejects the leap from that risk to prohibition. Closed models can be breached, misused, or fail in ways nobody outside the company can see, and stacking capability behind a handful of them creates a small number of very large single points of failure. Openness, the signatories write, may be one of the most important paths to AI safety and security. That sentence quietly inverts the assumption underneath most current AI policy.

There is also a passage on distillation, the practice of training one model on another's outputs. The letter calls it a legitimate and widely used technique for improvement, evaluation and validation, and warns policymakers against confusing it with theft. Unlawful extraction from closed models is a real problem, the signatories allow, but the fix is targeted legal and commercial remedies rather than broad restrictions on a technique the whole field depends on.

The roster

Read the signatures and the shape of the coalition becomes clear: Nvidia, Microsoft, Meta, IBM, Dell, Palantir, CrowdStrike, ServiceNow, Box, Andreessen Horowitz, Y Combinator, Hugging Face, Mozilla, the Linux Foundation, Mistral, Replit, Perplexity, Arcee AI, Black Forest Labs, Reflection, Telnyx, Arena, Emergence Capital, Mariana Minerals and the American Innovators Network.

Chips, clouds, enterprise software, open-source institutions, venture capital. Almost none of the labs at the top of the closed-model market.

OpenAI, Anthropic and Google all sat this one out. Those three sell access to the frontier through APIs, and all three would benefit if Washington restricted foreign open-weight competitors. Sam Altman said afterward that he was glad to see support for open-weight models, which is a mild thing to say about a letter you chose not to sign.

The incentives on the signing side deserve the same scrutiny. Nvidia sells GPUs to everyone, and a world where thousands of companies fine-tune and self-host their own models buys considerably more silicon than a world with five API providers. Microsoft sells closed models on Azure and hosts open ones alongside them. Mistral is French. Mozilla and the Linux Foundation are nonprofits with no commercial stake at all, which makes the letter's American framing looser than the title implies. None of that makes the argument wrong. It does mean this is a coalition of interests rather than a disinterested finding.

Why now

The timing is not accidental. Axios reported around July 20 that parts of the Trump administration have been looking at ways to restrict access to advanced Chinese open-weight models inside the United States, with AI czar David Sacks among those pushing the concern. The trigger was Kimi K3, the very large open-weight model Moonshot AI released in mid-July, which posted near-frontier results and became a Washington talking point within days.

That sits on top of a steady escalation. The State Department warned American companies about Chinese models on July 8. Congressional committees opened inquiries into firms using them. Treasury Secretary Scott Bessent told Fox Business on July 21 that the U.S. has the ability to sanction overseas models found to be stealing from American companies. Anthropic published its own account of an alleged large-scale distillation campaign against its models back in February. Seen against that record, the letter's distillation paragraph looks less like a technical footnote and more like a fence being built around the exact technique now being used to justify restrictions.

Then there is the enforcement problem, which nobody in Washington has solved. Export controls work because chips move through ports and fabs. API restrictions work because someone owns the server. Published weights have neither property. They sit on company hardware, already downloaded, already copied, with no switch left to flip. By some counts Chinese open models already handle close to half the enterprise tokens routed through platforms like OpenRouter. A ban would fall hardest on the American startups least able to absorb frontier pricing while leaving the weights themselves available to everyone outside U.S. jurisdiction.

What Friday actually changed

Legally, nothing. This is advocacy. The letter asks for compute access for startups and researchers, investment in shared datasets and evaluation tools, and restraint on early restrictions, but it does not say what release thresholds should look like, what evaluations should be required, or what protecting national security involves in practice. Its safety argument also carries a bill attached. Community scrutiny only works when someone funds the people doing the scrutinizing, and the letter says nothing about who pays. Any hearing worth the name will press on both points.

What did change is that a split inside the industry became public and easy to read. The hardware layer, the enterprise software layer, the open-source institutions and a large slice of the venture ecosystem have now put their names to a position. The three companies with the most valuable closed models in the world did not.

Huang's own line is the least maximalist thing anyone said all day, and probably the hardest to argue with. The world needs both kinds of frontier model. The fight in Washington is over whether the United States ends up with both.