Almost a month after creating his X account, Jensen Huang posted his first post.
He doesn’t talk about NVIDIA’s activities, but shares an open letter that the company signed, along with about 80 other organizations. Topic: open-weight models… and their importance for American AI leadership.
The argument borrows heavily from the vocabulary of open-source advocates. In broad terms: openness fuels innovation, which in turn creates opportunities for the economy and society.
The letter presents open-weight as a vector of safety and security. For and by AI. For, in that openness fosters the identification of vulnerabilities and the development of guardrails by the community. On the other hand, these models are also tools to combat cyberattacks.
Regulators are furthermore urged not to conflate distillation with misuse. The letter invites them to consider targeted legal and commercial frameworks rather than “radical restrictions” aimed at this training technique.
Anthropic, Against the Current
The letter initially counted around twenty signatories. Including CrowdStrike, Dell, Hugging Face, IBM, Meta, Microsoft, Mistral AI, Mozilla, Palantir, Perplexity and ServiceNow. OpenAI joined the circle later. As did, among others, Cisco, Cohere, GitHub, Google and Palo Alto Networks.
Amazon did not sign. Anthropic did not either, in line with statements from its chief, Dario Amodei. In late June, before the American Congress, the individual explained that he anticipates a major security risk with open models. With those published so far, the risks are “rather limited,” he conceded. But it will be different “in a horizon of 2 to 3 years for biological risks and probably less for things like misinformation,” he says.
For years, Dario Amodei has been articulating statements of this tenor. At the end of 2023, for example, he had said he favored open AI… but “of small size” (which, according to him, corresponded to the main models at the time). He looked ahead 2 to 3 years. And he estimated that by that deadline, it might no longer be possible to guarantee the safety of open models.
These statements align with Anthropic’s business model. Unlike Google, IBM, Meta or Microsoft, the company does not publish models. Its are closed, and it earns the bulk of its revenue through its APIs.
Distillation, the Nerve of the War
Anthropic has made a specialty of accusing Chinese companies of distilling its models. In early 2026, it pointed to three at once: DeepSeek, MiniMax and Moonshot AI. It banned them—and their subsidiaries—from accessing Claude for national-security reasons. But that ban allegedly was circumvented by going through third-party services.
More recently, Anthropic also accused Alibaba, which would have extracted knowledge from Claude via “thousands of fraudulent accounts.”
OpenAI has also taken shots. Notably against DeepSeek, as early as 2025, after it launched its R1 model.
The White House seems to be adopting the bulk of these accusations. The rise of post-training methods gives them even more weight. At the top of the list, distillation, which appears as an economical alternative to reinforcement learning environments.
OpenAI & Co. Should They Be More Permissive With Their Closed Models?
For a long time, people attributed to the Trump administration the ambition to restrict access to Chinese models. The Moonshot AI release of Kimi K3 revived the rumors. Even more so after Xi Jinping’s remarks. The latter praised the principle of openness at a moment when AI “arrives in the physical world” — that is, primarily, the Chinese industrial production apparatus.
Alibaba, which had at one time renounced publishing the weights of its most powerful models, changed its mind in the wake of these statements. It has at least committed to opening “soon” the weights of its new flagship Qwen3.8 Max. Moonshot AI did the same with Kimi K3, promising its release on July 27, 2026.
Irony or paradox: while Washington cries foul about illegal distillation, the American economy is seizing the open Chinese models… which it uses to generate synthetic data. In this context, voices have risen to propose creating a form of a “distillation pipeline.” It would involve AI labs such as OpenAI and Anthropic relaxing their terms of use. Or introducing more permissive licenses, akin to what NVIDIA pursues with its Nemotron models — albeit with clearly more interest in this area.
