An Eight-Dimension Audit of AI Governance

Summary

Sergei Strebulaev argues that AI is not necessarily approaching a technical ceiling; instead, the economics of deployment are changing. Citing Epoch AI, he says the cost of reaching a fixed capability level has fallen about 47% per quarter, or roughly 13-fold per year, since 2023. Vercel data through October 2026 reportedly show open-weight models generating 71.5% of gateway token volume but about 25% of estimated spend, while frontier-model spending also grew, suggesting workload separation rather than replacement. High-volume, lower-stakes work is moving toward commoditized open-weight models, while reliability-sensitive work remains with frontier providers. The essay therefore treats distribution, compute, and access rules as increasingly important infrastructure layers. It uses OpenAI-Microsoft and Anthropic's cloud partnerships as examples of product positions being converted into infrastructure positions. The central governance case is Anthropic's reported arrangement with METR: six METR employees are embedded at Anthropic, with access to company resources and publication rights subject to defined redactions. The essay also discusses a proposed Anthropic-backed federal evaluation framework, including independent evaluators, registration, compensation rules, and a reported threshold for government intervention. Strebulaev proposes an eight-dimension audit covering evaluator independence; appointment, funding, and oversight; publication control; evaluation infrastructure access; proportionality; treatment of open-weight models; cost distribution; and less-restrictive alternatives. Applied to several regimes, the framework finds the Anthropic template potentially exclusionary on access, proportionality, and open-weight treatment, while emphasizing that these are claims requiring evidence rather than conclusions about intent. The essay considers strategic incentives, genuine safety concerns, and regulatory uncertainty as three possible explanations, without choosing among them. It concludes that governance should be judged both by the risks it mitigates and by how it distributes authority, access, and compliance costs. Future design choices to watch include funding, shared evaluation infrastructure, simplified paths for smaller developers, and rules for open-weight models.