Anthropic CEO Dario Amodei: Open Weights Won't Stop AI Power Concentration
Amodei argues that scaling laws and compute access, not regulation, are the primary drivers of industry dominance.
Anthropic CEO Dario Amodei has challenged the notion that open-weights models can prevent the concentration of power in the artificial intelligence industry. In a public exchange on X with investor Gavin Baker, Amodei argued that the structural requirements of AI development create a natural ceiling that software openness alone cannot bypass.
Amodei asserts that the concentration of power in AI is driven primarily by scaling laws and access to compute rather than government regulation. While open-weights models allow developers to run systems locally, Amodei claims this is "nowhere near a sufficient solution" because it simply shifts the concentration of power toward those who possess the most chips and hardware infrastructure. He dismissed the common Silicon Valley narrative that regulation is synonymous with regulatory capture and power concentration, calling that view an "overly simplified picture of the world."
The Compute Threshold
To address these risks, Amodei is advocating for a tiered regulatory approach that distinguishes between general AI development and "frontier" models. Anthropic has expressed support for California's SB 53, a legislative effort that defines frontier models based on their training intensity—specifically those trained using 10^26 floating-point operations (FLOPs) or more. By targeting models at this scale, the goal is to impose strict safety requirements on the most powerful systems while exempting smaller developers from the same burdens.
Furthermore, Amodei supports the creation of an AI standards body similar to the Financial Industry Regulatory Authority (FINRA). This proposal, which is also backed by DeepMind CEO Demis Hassabis, would involve classifying AI models based on specific benchmark thresholds to ensure safety and accountability as capabilities scale.
Industry Implications
This stance highlights a deepening rift in the AI sector regarding the nature of systemic risk. On one side, proponents of open-source AI argue that concentrating power within a few regulated corporations creates a dangerous bottleneck and invites regulatory capture. On the other, Amodei suggests that the primary threat is catastrophic risk—where unregulated, highly powerful models could enable dangerous attacks.
By focusing on compute-based thresholds, Amodei is signaling that the battle for AI dominance is fundamentally a struggle over physical infrastructure and energy. In this view, the ability to secure massive clusters of GPUs is a more significant determinant of power than the decision to release model weights to the public.
What's Next
As legislative efforts like SB 53 move forward, the industry will be watching whether compute-based definitions become the global standard for AI regulation. The potential establishment of a FINRA-like body would mark a shift toward formal, benchmark-driven oversight of the frontier. However, the tension between the "open weights" movement and the "compute-centric" regulatory view remains unresolved, leaving the future of AI accessibility in question.