Open-source AI weights don't solve power concentration, Amodei argues

Anthropic CEO Dario Amodei is pushing back against the notion that open-weights AI models level the playing field for smaller developers. During a public debate on X with investor Gavin Baker, he emphasized that while releasing model weights offers some benefits, it fundamentally fails to address AI's underlying power dynamics.

"AI is structurally a technology that tends to concentrate power, for reasons that have nothing to do with regulation," Amodei wrote, pointing to inherent scaling laws as the culprit. The real constraint isn't code or model architecture—it's compute and hardware access. "Open-weights do help some with this but are nowhere near a sufficient solution because they simply shift the concentration somewhat to those with the most compute and chips," he added.

The debate over distribution versus regulation#

Baker, managing partner at Atreides Management, framed the argument around Meta CEO Mark Zuckerberg's position: concentrating powerful AI systems among a handful of supposedly responsible companies creates its own dangers. Baker suggested that distributing models more widely would spread power and let users access systems aligned with their own values. He also warned that Amodei's repeated safety concerns could sway policymakers to block new data centers needed for AI progress.

Amodei rejected this as a false choice. He disputed the common Silicon Valley assumption that regulation automatically leads to regulatory capture and monopoly power. Instead, he argued that well-designed regulation based on objective standards can constrain incumbents. When institutions "vest power in ideas rather than people, and thereby decentralize that power," they create genuine safeguards without entrenching gatekeepers.

Why compute remains the real barrier#

Open weights let developers fine-tune models and process sensitive data locally—but this advantage shrinks as models grow larger. Smaller teams can rent hardware, yet they still depend on expensive, limited compute resources. The compute shortage is already visible: earlier this year, five European companies committed to purchasing AI compute capacity that doesn't yet exist.

Amodei acknowledged that open models without dangerous capabilities serve as a public good. However, once a system becomes powerful enough to enable serious attacks, he believes it should face mandatory safety testing before release, regardless of whether weights are proprietary or open. Real-world containment failures have shown how hard it is to control capable models even before their weights escape into the wild.