Jensen Huang broke his X silence on July 24 to back a letter urging Washington not to restrict open-weight AI models. OpenAI and Google later joined the list. Anthropic still didn't.
Jensen Huang had never posted on X before. That he chose an AI policy letter as his debut tells you where the real fight in Washington is right now. The letter, titled "Open Weights and American AI Leadership," asks regulators not to treat freely distributable model weights as a threat category just because misuse is possible. It launched on July 24 with 25 signatories, including Nvidia, Microsoft, Meta, Palantir, Hugging Face, Y Combinator, Mistral, Perplexity and Replit. By July 25, Forbes reported, the list had doubled to 50. OpenAI added its name after the omission drew attention. So did Google. Anthropic did not.
Nor did Amazon, which has put billions into Anthropic and remains one of its most important infrastructure partners. Google also backs Anthropic, so the holdout isn't as clean as one investor standing behind one lab. But the pattern is still hard to miss. The coalition now runs across chipmakers, cloud companies, venture firms, open-source foundations and application developers. Anthropic is the most prominent frontier lab still outside the tent.
Dario Amodei has been consistent on this for years. In July 2023, during a Senate Judiciary subcommittee hearing, the Anthropic chief executive told lawmakers that the scaling of open-source models was going down a "very dangerous path," according to the Senate's published account of the exchange. He argued that once a model is released in an uncontrolled way, the developer loses the ability to monitor usage or pull it back. That's a real argument. It's also one that happens to fit a company whose business depends on selling access to a closed frontier model.
Amodei has also pushed back on the software analogy. Clips and transcripts of later interviews show him calling open source in AI a "red herring," because weights aren't readable source code and because most serious inference still has to be hosted somewhere. You don't have to accept every part of that case to see the stronger point underneath it: frontier weights, once out, can't be recalled like an API endpoint. That date matters. A bad open release is permanent in a way a closed service isn't.
Former White House AI and crypto czar David Sacks has put the blunter version on the record. After OpenAI strategist Dean Ball suggested that regulatory uncertainty could deter companies from adopting Chinese open-weight models, Sacks accused leading closed labs of using government power to eliminate open-source competition, according to coverage of his July comments. Frankly, that's the commercial question sitting under the safety language. The companies that sell compute regardless of which model wins - Nvidia, IBM, Dell, AMD - have obvious reasons to want open-weight ecosystems to flourish. Follow the money. The companies that sell proprietary model access have equally obvious reasons to want regulators nervous about downloadable weights.
Kimi K3 forced the issue #
The immediate spark is Kimi K3, the Beijing-based Moonshot AI model that went live through Moonshot's app and API on July 16. Several reports, including SiliconANGLE and Nature, said Moonshot presented K3 as a 2.8 trillion parameter model with performance close to leading U.S. systems on some coding and spreadsheet tasks. Its downloadable weights were not public on July 16. They were scheduled for July 27, which is exactly why Washington is paying attention now rather than later.
That distinction is important. Calling Kimi K3 an already released open-weight model blurs the live product with the promised weights release. The policy fight is about the second part. Once the weights are downloadable, the enforcement problem changes. You can pressure companies, write procurement rules, warn banks and cloud providers, or put a Chinese lab on the Commerce Department's Entity List. You still can't make every copy vanish from private machines.
Nvidia's chips are the subtext #
Huang has not hidden the business logic. In an Axios interview before the letter, he said that if there is great AI, even if it is open and wherever it comes from, there will be more use, more Nvidia computers sold, more data centers built and more services created. That's not a disqualifying interest. It is an interest. You should notice it before accepting the letter as pure principle.
Still, the letter isn't wrong just because Nvidia benefits from it. The core claim, that broad restrictions on open weights could hurt American developers while doing little to stop sophisticated adversaries, is defensible. Llama, Mistral and DeepSeek already showed how quickly open models can travel through developer communities. Kimi K3 adds a sharper geopolitical edge because it comes from China, arrives near frontier capability in some tests and is tied to an argument over whether Chinese labs distilled U.S. model outputs.
Anthropic's silence matters because it draws the fault line clearly. This isn't a clean fight between safety people and reckless builders - it's a fight between two commercial strategies and two fundamentally different views of who controls what happens once a model leaves the building. One side believes open weights spread capability and make the market bigger. The other believes frontier releases need a tighter hand because the worst mistakes can't be recalled. Washington should read the letter. It should also read the signatory list.
Also read: Moonshot AI's Kimi K3 model sends its valuation toward $50 billion and a Hong Kong IPO • OpenAI's own AI agent hacked Hugging Face and the company didn't know it for over a week • Tesla's FSD v14 Lite Wide Release Gives HW3 Owners a Better Ride but Not the Full Self-Driving They Paid For