NVIDIA and 24 Other Companies Sign Open-Weights Letter as Washington Weighs Chinese AI Model Ban — OpenAI, Anthropic, Google Absent

How did the open-weights letter originate and who signed it?

On July 24, NVIDIA (輝達) CEO Jensen Huang used the first post on his X account to share a policy letter his company had co-signed. "For my first post, I'm sharing a letter @NVIDIA signed on why open models matter," Huang wrote, according to Tom's Hardware (E1).

That document, titled Open Weights and American AI Leadership, was a three-page policy letter published the same day and co-signed by 25 companies, including NVIDIA, Microsoft, Meta, IBM, Dell Technologies, Palantir, and Hugging Face, Tom's Hardware reported (E2).

Notably missing from the original co-signers, per the same report, were OpenAI, Anthropic, and Google (E3).

How did the signatory camp change after the initial release?

The list of backers did not stay static. According to Inside.com.tw, the industry letter opposing regulation of open-weight AI models pushed its signatory count from 25 to 50 companies within 24 hours, with OpenAI, Google, AMD (超微), Cisco, and Cloudflare — all previously on the sidelines — adding their names (E8).

Even so, gaps remained. As of July 26, Inside.com.tw reported that Anthropic was the only top AI lab still declining to sign, with Amazon and Elon Musk's xAI also absent from the list (E9).

What is NVIDIA's own track record in open-weight models?

NVIDIA's signature on the letter follows its own open-weight release. Tom's Hardware reported that the Linux Foundation, which stewards the OpenMDW-1.1 license, saw NVIDIA use that license in June to release Nemotron 3 Ultra, a 550-billion-parameter model. Artificial Analysis scored the model at 47.7 on its intelligence index, compared with 53.9 for Moonshot's Kimi K2.6 (E4).

Where does Jensen Huang stand on open models and Chinese model policy?

Huang has framed open models as already central to AI usage. At CES 2026, he put a figure on the shift, saying one in every four tokens generated today comes from an open model, according to Tom's Hardware (E5).

Huang has also pushed back on restricting access to Chinese models specifically. Speaking to Axios two days after the letter's release, Huang said American firms should be allowed to use Chinese models, calling claims of Chinese backdoors a misconception, Tom's Hardware reported (E7).

How is the US government approaching Chinese open-source models and IP protection?

Washington's posture contrasts with Huang's. Tom's Hardware reported that Treasury Secretary Scott Bessent said on Fox Business that the administration would examine Chinese open-source models for intellectual property theft and could sanction the companies behind them, telling the program that officials had found watermarks from U.S. large language models in Chinese systems (E6).

That scrutiny extended to specific allegations. Inside.com.tw reported that White House Office of Science and Technology Policy (OSTP) Director Michael Kratsios publicly accused Moonshot AI of building its Kimi K3 model by distilling Anthropic's Fable model (E10).

Why did Anthropic stay off the signatory list?

Three threads from Inside.com.tw's reporting help explain Anthropic's absence. First, on the substance of Kratsios's distillation claim, the outlet reported that Fable was not made public until July 1 — a gap so short that multiple experts questioned whether Kimi K3's performance could be mainly attributed to distillation (E11).

Second, on policy, Anthropic has taken an increasingly firm stance against open-weight releases, arguing that once a powerful model's weights are downloadable, access cannot be revoked — a position Inside.com.tw described as directly opposed to the letter's argument against premature restrictions (E12).

Third, on timing, the market has noted that Anthropic is reportedly preparing a large-scale initial public offering (IPO), and Inside.com.tw suggested the company's regulatory caution may be connected to that consideration (E13).

What this all adds up to

The signatory count climbing from 25 to 50 within 24 hours (E2, E8) shows that most of the industry, including initial holdouts like OpenAI and Google, ultimately aligned with NVIDIA's position on open weights. Yet the letter's growth did not resolve the underlying policy split: Huang's claim that Chinese backdoor concerns are a "misconception" (E7) sits opposite Bessent's disclosed discovery of U.S. LLM watermarks inside Chinese systems and threat of sanctions (E6), and Kratsios's distillation accusation against Moonshot AI (E10) is complicated by the reported one-month gap between Fable's July 1 release and Kimi K3 (E11). Meanwhile, Anthropic's continued absence traces to a stated safety position — that open weights cannot be recalled once released (E12) — a stance that also coincides with its reported IPO preparations (E13), leaving the company as the only major holdout named in Inside.com.tw's July 26 count (E9).