Anthropic CEO Dario Amodei said in a public statement that the company has never advocated banning open-weight models. He did not directly name Nvidia CEO Jensen Huang in the statement, but he said he agrees with a core argument made by Nvidia and other large tech companies: stopping open-weight models developed overseas would not, by itself, resolve national security risks.
Amodei said a more direct and effective step would be preventing advanced chips from reaching China or other authoritarian states. He also argued that the systems that merit the greatest concern may be drone training models that fall into the hands of authoritarian governments, as well as industrial-scale distillation.
Anthropic says it has not opposed open-weight models
Debate around open-weight models has intensified, especially around models from China. Media reports have said some U.S. officials are considering whether to bar American companies from using open-weight models developed in China. In response to that broader debate, major technology companies signed a public letter backing open-weight models.
Against that backdrop, some critics accused Anthropic of trying to ban open-weight models to protect its own business interests. Amodei rejected that claim. He said anyone who has read his earlier writing should know that neither he nor Anthropic has ever opposed open-weight models, and he said he wanted to restate that position clearly to avoid any misunderstanding.
He described open-weight models without dangerous capabilities as a public good. Aside from the compute needed to run them, he said, they come at effectively no cost and can deliver value to companies, developers, and researchers. Opening that class of model, in his view, does not harm U.S. businesses.
Amodei points to drone training and repression models
Amodei said his main concern is that the Chinese Communist Party could build AI models stronger than those in the U.S. and use them to secure a permanent military advantage or to carry out severe repression against its own population. In his view, the most dangerous models may not be publicly released general-purpose systems, but models trained in secret and delivered only to the People’s Liberation Army for drone training, along with AI systems used in national security settings for surveillance and repression.
Open-weight models carry real risks, but that is not a case for a blanket ban
Jensen Huang and other tech leaders have backed open-weight AI while also warning about the risks. Amodei said he agrees with that assessment. He wrote that powerful AI systems could be misused for cyberattacks or biological attacks, and that serious algorithmic matching problems may also exist.
He said open-weight models, whether from China or anywhere else, may present greater risk than closed models because safeguards are harder to impose and their use is harder to monitor. Once weights are released, he noted, they cannot be recalled.
Even so, Amodei said there is no need to prohibit U.S. companies from using these models. Legitimate American businesses, he argued, are unlikely to put them to malicious use.
Supports chip controls and tighter enforcement
On policy, Amodei backed restrictions on advanced chips. He said the U.S. should not sell high-performance chips or chipmaking equipment to China, and should crack down hard on widespread smuggling and other workarounds used to obtain those chips.
His reasoning is that China’s domestic production capacity is limited. Without U.S. chips, he said, China would not be able to build models more powerful than those in the U.S. He described that as the most direct way to stop the threat.
Industrial-scale distillation is another policy concern
Amodei also urged policymakers to watch industrial-scale distillation closely. He said distillation on that scale can sharply increase effective compute, allowing certain countries to build model capabilities close to the frontier even without access to large quantities of chips. In his view, that could weaken existing hardware bans.
He called for a specific enforcement framework aimed at this kind of distillation so authorities can intervene with greater precision.