In recent days, a divide has emerged in Silicon Valley over Chinese open-weight AI models. Companies such as Nvidia and Microsoft have supported these models, especially after systems like Kimi K3 demonstrated impressive performance. More than 50 companies have even sent a letter to the White House urging it not to restrict Chinese open-weight models. However, Anthropic was notably absent from the list. Now, CEO Dario Amodei has addressed the criticism, saying the company is not opposed to open-weight models but is concerned about potential risks involving China.
In a blog post, Amodei said Anthropic "never advocated for a ban on open-weights models." He acknowledged that open-weight models offer several benefits, explaining that they "don’t cost anything besides the compute needed to run them, and they provide value to businesses, developers, and researchers."
However, Amodei said he remains concerned about what he described as "two nightmare scenarios" linked to open-weight AI systems.
It is important to note that open-weight models such as Kimi K3 and GLM 5.2 are significantly cheaper to operate compared with closed models like Claude or GPT-5.6. These models can also be deployed locally on personal devices.
Dario Amodei's concerns about China
One of Amodei's biggest concerns is the possibility of authoritarian governments gaining access to highly advanced AI models, particularly China. He said, "The Chinese Communist Party (CCP) is clearly the most capable threat to build AI models that are more powerful than those built by the US."
According to Amodei, if China develops AI systems more advanced than those in the US, they could potentially be used to achieve lasting military dominance or strengthen government surveillance and control over citizens.
He warned that AI could be used for purposes such as suppressing populations or enabling autonomous warfare systems, including machines capable of causing harm with minimal human involvement.
Open-weight models may be easier to misuse
China was not Amodei's only concern. The Anthropic CEO also highlighted broader risks associated with increasingly powerful AI models, including potential cyberattacks and biological threats. He argued that open-weight models carry greater risks because controlling their use after release is extremely difficult.
"Open-weights models—it does not matter whether they come from China or anywhere else—do potentially present a higher risk than closed models, because it is very difficult to apply guardrails to them or monitor their usage, and once weights are released they cannot be withdrawn," Amodei said.
Anthropic itself has previously been involved in discussions around AI safety and cybersecurity following the release of Claude Mythos. The US government had restricted access to Mythos and Fable models over safety concerns, though Anthropic later made them available again after introducing additional safeguards.
Amodei's proposed approach
Rather than imposing a complete ban on Chinese open-weight models, Amodei believes the US should introduce targeted restrictions to slow potentially risky AI development in China.
He argued that the US should prevent the sale of advanced chips and chipmaking equipment to China while also addressing illegal methods used to obtain restricted technology. Despite existing export controls, Chinese companies have continued advancing their AI capabilities.
DeepSeek's CEO recently argued that these restrictions have encouraged Chinese AI companies to collaborate with domestic firms such as Huawei to improve their own semiconductor technology.
Amodei also called for action against large-scale AI distillation efforts. Distillation involves training a smaller AI model using outputs generated by a larger model. Anthropic has previously accused Chinese AI companies of using this technique to develop their own systems.
Amodei said distillation does not allow China to immediately gain equal or superior AI capabilities compared with the US, but it could reduce the gap between the two countries' AI progress to just a few months.
His final recommendation was mandatory safety testing for all advanced AI models, including both open-weight and closed systems.
"The best way to address threats is to just directly test models for cyber, biological, and alignment risks before release," he said.
Amodei added that he was encouraged by recent steps taken by the Trump administration in this direction.
Debate over the industry letter
Addressing the letter signed by several technology companies, Amodei said he disagreed with some of its arguments. He rejected claims that open-weight models automatically make it easier to create safety measures or that wider access to AI capabilities benefits defenders more than attackers.
The controversy comes as technology companies have taken opposing positions in Washington over AI policy. Several firms, including OpenAI, supported the letter arguing that open systems make advanced AI more accessible, flexible and widely available.
Anthropic's decision not to sign the letter attracted criticism from some figures, including former White House AI adviser David Sacks, who accused the company of attempting to weaken Chinese competitors while protecting the interests of closed-model companies.
The debate has intensified after Chinese AI startups released several models that challenged Silicon Valley's assumption that US companies would maintain a clear lead. From DeepSeek and Z.ai's GLM 5.2 to Moonshot AI's Kimi K3, Chinese firms have introduced lower-cost AI models that match or, in some cases, outperform leading American systems on certain benchmarks.
