Microsoft and Meta want AI you can download. OpenAI and Anthropic don't.

On July 24, 2026, twenty-five US technology companies, including Nvidia, Microsoft, Meta, IBM, Palantir, Hugging Face, and Mistral, published a public letter titled "Open Weights and American AI Leadership" asking Washington to avoid "premature restrictions on downloadable AI models."

"Open weights" refers to AI models whose trained parameters are released publicly, so that anyone can download them, inspect how they work, and run them on their own computers or private servers. The alternative is the closed model used by OpenAI and Anthropic, where the weights stay locked inside the vendor and users access the model only through an API.

OpenAI, Anthropic, and Google are absent from the signatory list, and the omission maps onto a structural split inside the US AI industry that has been hardening for two years. The letter is the first time the two halves have publicly drawn a line on federal policy.

The letter asks the Trump administration not to impose export controls or licensing requirements on the release of model weights. It frames open weights as an extension of the 1980s open-source software movement, a position that argues US AI leadership is best preserved by making American models the global default, the way Linux became the default for servers. The signatories describe restrictions as a self-inflicted wound: in their telling, the United States would cede the global open-model market to Chinese competitors such as DeepSeek and Moonshot AI by tightening its own rules first.

The list is the open-weights business stack: chipmakers (Nvidia), server and PC vendors (Dell), enterprise software (IBM, Palantir, Oracle), open-model labs (Mistral, Hugging Face), open-source foundations (Mozilla, the Linux Foundation), and the venture firms that fund them (Andreessen Horowitz, Y Combinator). These are the companies whose business model assumes customers can download, fine-tune, and self-host models. Closed-frontier labs, by contrast, make money by renting access. Nvidia's interest is also direct: the more places that run models, the more GPUs get sold.

The companies that did not sign have a different business model. OpenAI and Anthropic sell access to their models through APIs rather than releasing the weights, and Google sits between the two camps by publishing some open models while keeping its frontier Gemini line closed. The letter's signatories frame the absence as evidence that the closed-frontier labs prefer narrower licensing rules that protect their API revenue. Neither OpenAI, Anthropic, nor Google has issued a public counter-letter, and none responded to requests for comment by publication time.

Four days earlier, the Trump administration was reported to be reviving a push to ban or restrict Chinese AI models, including DeepSeek and Moonshot AI, over concerns about intellectual-property transfer and benchmark competitiveness. Tom's Hardware and TechCrunch both reported the timing. The benchmark pressure is real, even if the numbers are contested: winzheng.com reports that Moonshot's Kimi K3 model has surpassed leading US models on several industry evaluations, a claim that has fueled calls inside Washington for tighter controls.

By Friday evening, Nvidia CEO Jensen Huang had used his first post on X to promote the letter, and Microsoft CEO Satya Nadella publicly backed it. The coalition is still growing: Forbes reported on July 25 that the signatory list had roughly doubled to around fifty companies, with Amazon and Anthropic still absent. The "doubled to 50" figure is one reporter's count, and the original letter anchored on the initial twenty-five.

The next question is whether the White House treats the letter as lobbying to be heard and then overridden, or as the policy it actually enacts. The administration has not signaled which way it is leaning, and the absence of a public counter-letter from the closed-frontier side means the open-weights camp currently owns the airwaves. That tends not to last.