- Split. OpenAI joined a 32-company industry letter on July 25 defending open-weight AI models, breaking ranks with Anthropic and separating itself from the position both companies had privately lobbied in Washington.
- Fact. Anthropic alone among major US AI companies declined to sign, with its policy head Sarah Heck calling Chinese model distillation “IP theft and industrial espionage that supports adversary military and intelligence capabilities.”
- Stake. The split exposes a fracture inside the US AI industry at the exact moment Washington is deciding whether to restrict Chinese open-weight models — a question with consequences for global AI competition that the government will not be able to defer indefinitely.
On July 25, OpenAI quietly signed a 32-company industry letter organised by Nvidia, Microsoft and Meta that urged policymakers to avoid “premature restrictions” on open-weight AI models. Sam Altman signalled the shift publicly, writing: “i want the US to win in AI both in open source and proprietary models.” The sign-on placed OpenAI alongside IBM, Palantir, Hugging Face, the Linux Foundation, and, ultimately, Google — and separated it from Anthropic, which remains the only major US AI lab to have declined the letter.
Anthropic’s Position
Anthropic’s stance has been explicit and consistent. Sarah Heck, the company’s head of policy, described Chinese AI model distillation practices as “illicit, adversarial distillation” amounting to “IP theft and industrial espionage that supports adversary military and intelligence capabilities.” The company has privately lobbied Washington to consider restrictions on the distribution of Chinese open-weight models, citing the controversy around Moonshot AI’s Kimi K3 release — a 2.8 trillion-parameter model that White House officials alleged was built by distilling Anthropic’s own Fable model at industrial scale. Moonshot’s K3 reached first place on web-development benchmarks and third on a major intelligence index within days of release.
The Government’s Position
US officials have sided rhetorically with Anthropic. Treasury Secretary Scott Bessent said on Wednesday: “open source is not open season on American IP.” But the government’s working posture appears to favour case-by-case national security reviews over blanket restrictions. US-China AI talks are scheduled for September, giving both sides reason to avoid drawing formal lines before then. The White House’s pre-release security review framework for frontier models, still being finalised, is seen by some observers as a bridge between the two camps — strict enough to satisfy Anthropic’s security concerns without locking Chinese open-weight models out of the US market entirely.
Regulatory Capture or Real Risk
Critics of Anthropic’s position, including some Trump administration advisers, argue that restricting Chinese open-weight models would entrench the largest US AI companies by making it harder for competitors — including domestic ones — to release models, a form of regulatory capture dressed as security policy. Defenders of Anthropic’s stance counter that the distillation problem is real: if a frontier proprietary model can be replicated cheaply by a foreign competitor in weeks, the commercial return on that investment — and the incentive to make it — erodes. That tension, unresolved by the coalition letter, is what Washington will have to arbitrate before the US-China AI talks in the autumn.