Anthropic Doesn’t Want to Ban Open-Weights Models — it Wants a Chip Ban on China
Trending Topics Jakob Steinschaden ● Covered by 72 sources
Anthropic is the lone big AI lab refusing to sign a pro-open-weights letter backed by Nvidia, Meta, OpenAI and 50 others. CEO Dario Amodei says he's not against open models — he just wants tighter chip controls on China and mandatory safety tests instead.
Fifty companies, one holdout. That's the shape of the latest Washington AI fight, and the holdout happens to be one of the industry's most prominent safety-focused labs. When word spread that US officials were weighing restrictions on Chinese open-weights models, Nvidia, Meta, Microsoft, OpenAI, Google, Mistral and dozens of others rallied behind a letter called "Open Weights and American AI Leadership." Jensen Huang liked it so much he posted about it on X for the first time ever. Anthropic and Amazon stayed silent, and Anthropic in particular has taken heat for it, with critics muttering that a company that sells only closed models has an obvious motive to slow down the free ones.
Dario Amodei answered those critics directly on Monday with a blog post that tries to reframe the whole argument. He insists Anthropic has never called for banning open-weights models and never will, because he thinks a US usage ban would be pointless theater. The bad actors he worries about, he writes, aren't legitimate American businesses that would comply with such a rule anyway — they're authoritarian states building weapons or surveillance tools, and those states don't care whether a model was open-sourced or built in a locked basement. A dangerous model handed straight to a military is dangerous either way.
What Amodei actually wants sounds almost boring by comparison: tighter chip export controls on China, a crackdown on distillation (the practice of cheaply copying a frontier model's abilities), and mandatory pre-release safety testing for any sufficiently capable model, open or closed. China, he notes, can't out-build American labs without American chips, so that's the choke point that matters. Startups and academic labs would be exempt from the testing regime, but the big players — including his own company — would not.
Where he actually breaks from the open-weights camp is philosophical. The letter argues that open access mostly helps defenders, letting more researchers spot and patch flaws. Amodei isn't buying it. He points to biology as his worst case: a capable enough model could help someone weaponize a pathogen in weeks using materials anyone can buy, while building countermeasures — even under wartime urgency, as Operation Warp Speed showed — takes years. He'd rather settle the defender-versus-attacker question with actual testing data than assume good outcomes.
None of this happens in a vacuum, of course. Chipmakers and infrastructure companies on the letter benefit handsomely from open models proliferating, and Anthropic makes its entire living from closed ones — it's also publicly accused rivals of using distillation to nearly match its frontier models on the cheap. The debate got fresh urgency from Kimi K3, a Chinese open-weights model that's closing in on Western frontier performance for a fraction of the cost. Washington hasn't decided anything yet, but a chip-focused approach and mandatory safety testing both seem to have more momentum than an outright ban — which happens to be exactly the outcome Amodei has been pushing for all along.
My take
Give Amodei credit for reframing this smartly: arguing for chip controls and safety testing sounds a lot more defensible than arguing to hobble open models, even if the practical effect on Anthropic's bottom line is similar either way. But the bigger tell is that nobody in this fight is disinterested — chipmakers want more compute sold, open labs want their models unrestricted, and Anthropic wants distillation stopped because it's the one being copied. The chip-choke strategy probably is the more effective lever against China than banning downloads, if only because banning downloads was always going to be unenforceable. Watch the safety-testing push closely, though — that's the version of this fight most likely to actually become policy, and it's the one that happens to hand incumbents like Anthropic a regulatory moat.”}]}】。 wait invalid} }}}}} ]}]}
Read more about this at: Trending Topics
