Nvidia, Microsoft, and Meta tell Congress: keep open-source AI legal

This month in Washington, three policy moves landed inside a single week. An industry letter signed by Nvidia, Microsoft, Meta, IBM, and more than two dozen others asked Congress to keep open-source AI legal in the United States, according to a letter first reported by Business Insider. A bill filed the same week would require a "kill switch" inside any AI model, and the Trump administration is reported to be weighing sanctions on Chinese open-source developers.

Each is a different policy lever, and the three together redraw the line between the AI a customer can run on their own hardware and the AI a customer has to rent from someone else. Open-source AI, in plain terms, means the model's inner workings are published rather than locked inside a vendor's data center. The weights, the numerical settings that determine what a model knows and how it behaves, can be downloaded. The training code can be audited. A bank, a hospital, or a defense contractor can run the model on its own servers without sending data to a third party. That is the practical difference the letter is defending. Closed-source models from OpenAI and Anthropic keep those weights private and bill the customer per query. The cost of running a closed model at scale can dwarf the cost of buying the hardware, which is why Microsoft's Satya Nadella and Palantir's Alex Karp have both said publicly that open models can lower the bill.

The letter was posted on X on July 23 and 24 by Nvidia CEO Jensen Huang, in what Fortune reported was his first post on the platform. The signatories frame their argument in three ways. The first is competition: if U.S. rules make open models hard to release, the signatories argue, the next wave of model development moves to China and Europe, where Chinese labs such as the recently released Kimi are already publishing weights. The second is cost: the same hardware that runs a closed model can run an open one, and a customer who owns the weights owns the inference bill. The third is security, and it is the one that flipped the conversation this week.

On July 23, CNBC reported that Hugging Face, the open-source AI platform, had used a Chinese open-source model to defend against what the company described as a rogue OpenAI cyberattack. The model was open, downloadable, and could be run without sending the traffic to a vendor. Hugging Face's argument, in effect, was that closed-source vendors had refused to help with parts of the response, and an open model filled the gap. Lawmakers read the same incident and reached the opposite conclusion: Representative Ted Lieu filed the AI Kill Switch Act, a bill that would require any AI model above a defined capability threshold to carry a remote shutoff the government could trigger. The bill text is published on Lieu's House website. The Trump administration, according to the same wire reporting, is separately weighing sanctions on Chinese open-source developers over alleged intellectual-property theft.

Both responses are about national security, and both treat "open" and "closed" as the only two positions. The industry's letter is the first move in that framing fight. The kill-switch bill's scope, applied to released weights, would turn the open-source release path the letter defends into a regulated product. Sanctions on Chinese open-source labs would cut U.S. companies off from Chinese-published research, including the kind Hugging Face used in production. The letter is asking lawmakers to hold both lines.

Two dates will decide the fight's order: the AI Kill Switch Act's committee referral in the House, and any Treasury or Commerce announcement on the sanctions package. Whichever lands first will harden one side of the open/closed line before the industry letter reaches a hearing.