Amodei: open weights just move power to chip owners

Owning a model's weights does not deliver independence when someone else controls the hardware: Anthropic CEO Dario Amodei called open weights "nowhere near a sufficient solution" to the concentration of power in AI, in a public exchange on X with investor Gavin Baker reported by Thenewstack.

At a glance

• Amodei traces the concentration to scaling laws rather than government policy, and says publishing weights mainly shifts the advantage toward the organizations that already hold the most compute and chips.
• California's SB 53, backed by Anthropic, treats a model as frontier above 10^26 floating-point operations of training compute, with extra obligations for developers earning more than $500 million a year.
• Baker countered that wider distribution spreads power and gives people systems matching their values, quoting Mark Zuckerberg's warning that extreme concentration of AI is itself the dangerous path.

For teams running open models on their own infrastructure, the exchange sharpens where the real ceiling sits. Weights can be downloaded and fine-tuned, but serving a large model at scale still runs through hardware that someone else owns and prices. Amodei's framing appears to move the limit on independence from the license to the compute bill, which likely matters more to self-hosting plans than the release terms.

Baker framed the choice as regulated incumbents against wide distribution

Gavin Baker, managing partner at Atreides Management, set up the exchange as a choice between concentrating powerful models among a few regulated companies and distributing them widely without the same guardrails. He quoted Mark Zuckerberg: "The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic."

Baker argued that putting more models into circulation would spread power and improve the odds that people could use systems reflecting their values. He also said Amodei's repeated warnings about AI risk could strengthen opposition to new data centers, limiting infrastructure needed to deliver the technology's benefits.

"I know that there's a sort of Silicon Valley shorthand where regulation = regulatory capture = concentration of power," Amodei wrote, calling it an overly simplified picture of the world. He argued that regulation can constrain powerful companies when the requirements rest on objective standards, and that institutions at their best vest power in ideas rather than people.

Anthropic did not sign the July 24 open-weights letter organized by Nvidia

Anthropic was one of the few major AI labs that did not sign the July 24 letter organized by Nvidia, whose signatories argued that open weights encourage competition by giving customers an alternative to proprietary APIs. Amodei later said the company does not support a blanket ban on open models.

He called open models without dangerous capabilities "a public good", though his position changes once a model can help someone carry out a serious attack. Once weights are public, the original developer cannot retrieve every copy or stop users from removing the safeguards, and Amodei wants models at that level to undergo mandatory safety testing regardless of release format.

Open weights let developers adapt a model and keep sensitive data inside their own environment, but that becomes harder as models grow, and teams that rent hardware stay exposed to its cost and availability. Five European companies recently agreed to buy AI compute that does not exist yet.

SB 53 sets the frontier line at 10^26 FLOPs and $500 million in revenue

Amodei disputes the argument that AI regulation will shield incumbents from competition, saying the company has supported frameworks that place stricter requirements on frontier labs while exempting smaller developers: "We try very hard to make proposals that disadvantage (slow down) frontier AI companies while advantaging smaller competitors."

California's SB 53, which Anthropic supported, defines a frontier model as one trained with more than 10^26 floating-point operations, and developers earning more than $500 million a year face additional requirements, including publishing a framework explaining how they test for and respond to catastrophic risks.

Amodei has called for a similar distinction in federal policy, saying the testing process Anthropic has advocated at CAISI and the White House involves more rigorous tests for frontier models than off-frontier ones. He also supports Demis Hassabis' proposal for a FINRA-like standards body that would classify models after they cross regularly updated benchmark thresholds, applying the same rules to open and closed releases.

Where the frontier line gets tested

Training compute gives developers a threshold in advance, but it may not capture capabilities added through fine-tuning or external tools. The U.S. Center for AI Standards and Innovation found Moonshot AI's Kimi K3 behind the strongest closed models in preliminary cyber evaluations.

In a simulated attack on a small corporate network, Kimi K3 completed the full 32-step attack path in one of 10 attempts. Regulators would still have to decide when such a challenger crosses into frontier territory, ideally before its weights are released and modified beyond the creator's control.