A public letter that began with about 20 signatory companies grew to 50 in two days. Titled "Open Weights and American AI Leadership," the document was released on July 24 local time, and the early signatories included Nvidia, Microsoft and Meta. Palantir, IBM and Dell also put their names on it. With the venture capital firm Andreessen Horowitz joining as well, it came to look as if the side that makes the chips, the side that makes the models and the side that puts up the money had signed the same sentence.
Weights refers to the mass of numbers left inside an AI model after training is finished. If these are released, anyone can download them, take them apart and modify them. The letter did not deny that point.
It acknowledged that releasing weights takes them out of the developer's hands, while countering that keeping them closed is not inherently safer. The logic is that a broad community of researchers finds vulnerabilities and improves the safeguards.
The letter singled out specific companies and specific practices and demanded answers.
On July 20, US media reported that some cabinet members of the Trump administration were considering what would effectively be a ban on foreign open-source models. The Commerce Department circulated a draft rule last year that would make Chinese open-source AI models subject to regulation, and it also looked at a plan to place a number of Chinese AI labs on a blacklist. The NSA and the White House ONCD prepared recommendations dealing with the threat from Chinese AI labs.
The White House even toyed with an executive order allowing Chinese model services only on the condition that US companies take responsibility if a security incident occurs. All of these attempts ran into objections that innovation would be chilled and were dropped.
As figures including former White House adviser Sriram Krishnan left, the voices of hardliners inside the administration grew louder. OSTP Director Michael Kratsios and Treasury Secretary Scott Bessent hinted at the possibility of sanctions on the grounds that Chinese companies had distilled US models without authorization. Distillation is a technique that takes the outputs of a large model to train a small one, and it greatly reduces development costs.
The letter drew a line here. Its argument is that legitimate development techniques and illicit theft are different, and that the response should be worked out through means such as lawsuits and contracts and should not go the way of shutting down the technology as a whole.
Both parties to this confrontation are US institutions and US companies.
OpenAI and Anthropic have warned that cheap, high-performing Chinese models spreading without regulation would become a serious security risk. Dean Ball, OpenAI's head of strategy and futures, wrote on X on July 17 that a world dominated by open-weight models could lead to "AI communism." On the other side, White House AI adviser David Sacks shot back on X that the closed frontier labs want to borrow the government's hand to clear away open source, their competitor.
Nvidia CEO Jensen Huang introduced the letter in the first post on his X account. The Little Tech Association, a group of small and mid-sized US technology companies, had earlier, on July 22, sent a letter to President Trump, Commerce Secretary Howard Lutnick and Director Kratsios asking that access to Chinese open models not be blocked.
Behind this lies a sense that the performance gap has narrowed. Moonshot AI released its open-weight model "Kimi K3" during July, and Alibaba put out "Qwen 3.8 Max." The two models produced results on a par with US models on some benchmarks, and adoption by US companies is also rising.
If it is close to free and performs about the same, people use it. The side that wants to stop this flow with regulation and the side that sees US developers being tied up the moment it is stopped are divided.
Many South Korean companies and labs have taken open weights and fine-tuned them to build services, and the premise was that they could download the models they needed at any time. If the US actually restricts the distribution of open weights from a particular country, the options change depending on where the blockage falls, from repository access to cloud hosting. Whether the draft rule's scope is limited to Chinese models or widens to foreign open source in general is the fork in the road.
The letter wrote that America's AI edge does not rest on a single most advanced model. It rests on whether an open ecosystem that spreads into every field is built. Whichever way the regulation tilts, the list of models developers can download will change.
