Anthropic CEO Dario Amodei published a statement on July 27, 2026, explaining where his company stands on a hot question in the AI industry: should companies share the inner workings of their AI models so that anyone can download and use them?

When a company builds an AI model, it trains the model on huge amounts of data using powerful computers. The result is a set of numbers, called weights, that determine how the model behaves. If a company releases those weights publicly, anyone can run the model on their own computer, change it, and use it however they want. This is called "open weights." If instead the company only lets people use the model through its own website or app, it can watch what people do and block harmful requests. Amodei says Anthropic has never wanted to ban open weights.

He does draw a line, though. Models that are not capable of dangerous things are a public good, in his view, and should not be restricted. The real question is where to set the bar for what counts as dangerous.

Amodei names two main worries. The first is about countries, not individuals. Authoritarian governments could build AI models more powerful than what the United States has, and then use them for military advantage or to control their own people. This connects to something Amodei has spoken about before: he supports U.S. restrictions on selling advanced computer chips to China, because those chips are what make it possible to train the most powerful AI models. He also wants to stop the smuggling of chips that are meant for China through back channels.

His second worry is more familiar in the AI safety world. Powerful models could be misused for cyberattacks or biological attacks, or they might behave in ways their creators did not intend. Open-weights models are riskier on this front, Amodei says, because once the weights are out in public, the company that made the model cannot monitor how people are using it or stop them from doing harmful things.

The statement comes during a push by the chip company Nvidia to rally support for open weights. A letter organized by Nvidia's Jensen Huang had gathered 50 signatories as of July 25, 2026, according to Forbes. Amazon and Anthropic were both missing from the list. The Next Web reported on July 27 that Anthropic is the only major AI lab that has not signed the letter, which makes Amodei's statement the first clear explanation of why.

What stands out is what Amodei chooses not to do. He does not sign the letter, and he does not call for a ban on open weights. He picks a middle path: open weights that are not dangerous can help the public, but releasing the weights of the most powerful models, especially when rival nations can get the chips needed to train them, creates risks that should be handled through export controls and careful evaluation of what a model can actually do.

The connection to chip export restrictions is worth understanding. The U.S. has been limiting sales of advanced chips to China for years. Amodei's support for those limits puts Anthropic on a different side from Nvidia, which has fought against the restrictions because selling chips is core to Nvidia's business. The two companies disagree not just about open weights but about a bigger question: who should be allowed to access the most powerful AI technology, and how should that access be controlled?

For people watching the industry, the practical takeaway is that Anthropic is not against open weights in principle. The company is saying that the most powerful models should face careful testing before their weights are released, and that controlling who can get the chips to train such models is a necessary part of keeping AI safe. Whether Anthropic can hold this middle position as pressure builds to share more powerful models is what the industry will be watching next.