Jensen Huang had never posted on X before July 24. His first post on the platform wasn’t about chips or Nvidia‘s earnings or the next GPU architecture. It was a letter. An open letter signed by more than 20 tech companies, including Nvidia, Microsoft, Meta, Google, OpenAI, Hugging Face, Mistral, and Palantir, urging Washington not to place restrictions on open-weight AI models.
Anthropic wasn’t on the list.
For days, nobody said anything official. Then OpenAI signed on, and suddenly Anthropic was the last major American AI lab standing outside the letter. The longer it stayed quiet, the more the silence said something. On July 27, Amodei published his answer, CNBC reported.
What Dario Amodei said about open-weight AI and why Anthropic didn’t sign
“Anyone who has read my past writing should know that I don’t regard such bans as a useful measure, but let me state it clearly so that there is no doubt,” Amodei wrote on Anthropic’s website. “Anthropic has never advocated for a ban on open-weights models.”
He went further. “Open-weights models that don’t have dangerous capabilities are a public good,” he wrote. “They don’t cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers.”
That’s a significant statement from a company that had just spent several days being publicly accused of the opposite.
White House AI adviser David Sacks said Anthropic was using safety concerns to protect its own business model. An Anthropic researcher had taken a swing at Nvidia and Microsoft over the letter and received a pointed rebuttal asking why Anthropic hadn’t signed at all. Amodei’s post is a direct answer to all of that.
Why Anthropic still won’t sign the open-weight letter despite agreeing with parts of it
Amodei said he agrees with a lot of the letter. Open weights expand access, strengthen competition, give customers more control. He’s on board with all of that.
Where he parts ways is on two specific claims: that open-weight models necessarily make it easier to develop safeguards, and that broad access to capabilities helps defenders more than attackers.
More AI:
- Workers just sent AI companies an ultimatum
- Palantir CEO has a blunt verdict on OpenAI and Anthropic
- Elon Musk pulls no punches with AI rivals as Grok 4.5 debuts
His counterexample is biology.
A sufficiently capable AI model could potentially help bad actors weaponize pandemic-level viruses using widely available materials. Defense against those agents is a multi-year operational task at best, as the Covid pandemic showed.
In Amodei’s view, that’s an attacker-defender asymmetry that open-weight release makes worse, not better. Whether that’s actually true should be determined by testing, he says, not assumed in advance.
The Kimi K3 situation is part of the backdrop here. The shock release of the Chinese open-weight model in July, approaching US frontier performance at a fraction of the cost, is what triggered the entire debate. Amodei’s post specifically addresses Chinese frontier models and distinguishes them from Western open-source releases, according to Axios.
The three things Anthropic actually supports on AI policy
Instead of a ban, Amodei named three things Anthropic has been pushing for.
Keep advanced chips and chipmaking equipment out of China, and go after smuggling networks that get around export controls. Crack down on industrial-scale distillation, which lets Chinese labs close the gap to the US frontier without needing matching compute. And require safety testing for all sufficiently capable models before release, open or closed.
The safety testing piece is the most interesting. It doesn’t pick open or closed. It just says if a model is capable enough, it gets tested before it ships. Amodei even floated the idea that China might go along with a global testing regime because keeping AI away from bioweapon development is something Beijing doesn’t want either.
He also made a distinction that the open letter doesn’t. The letter treats all open-weight models the same. Amodei is specifically worried about frontier models with dangerous capabilities. Startups, academia, smaller labs: he’s not talking about them. That’s a much narrower target than a categorical ban, TechCrunch reported.
Anna/Getty Images
Why the business model criticism of Anthropic has some truth to it
The accusation that Anthropic is using safety to protect its closed-model business isn’t entirely unfair. Nvidia sells compute. The more open models there are, the more companies self-host them, and the more GPUs they buy. Huang’s interests align perfectly with open-weight expansion. Anthropic sells API access to Claude. A world of freely distributed capable models is more threatening to that business model than it is to Nvidia’s.
Amodei doesn’t address this directly. His safety concerns go back years and are well-documented. But the incentive alignment is what it is. Sacks isn’t wrong to point it out, and the safety concerns being genuine doesn’t make the business angle disappear. Both things are real.
Amazon also didn’t sign the letter. Like Anthropic, it sells cloud AI services. Like Anthropic, its absence speaks for itself without needing an explanation.
What Anthropic’s open-weight position means for AI safety and regulation
The argument Amodei is making isn’t really about open versus closed. It’s about what a model can do. A capable enough model gets tested before release, regardless of format. A model that doesn’t hit the capability threshold gets left alone. That’s the framework he’s pushing Washington toward.
That framing, if it gains traction in Washington, would change how AI regulation develops. It moves the debate away from a binary choice between banning Chinese models and allowing everything, toward a regime based on what models can actually do. Whether policymakers find that framework more workable than a simpler ban is the next question.
Anthropic is still the only major US AI lab that hasn’t signed the letter. Amodei’s post explains why without apologizing for it. Whether that holds as the political pressure builds is the next question nobody has an answer to yet.
Related: Mark Zuckerberg, Microsoft CEO just made major AI decision
