A new analytical framework published by independent researcher Sergei A. Strebulaev argues that the convergence of slowing AI capability differentiation and aggressive regulatory shaping is not a coincidence, but a strategic shift in the economic gradient of deployment. The core thesis posits that because the cost of achieving fixed AI capability has fallen roughly 13x per year since 2023, frontier labs are no longer competing on product differentiation alone. Instead, they are securing rents by controlling the infrastructure layer—specifically distribution, compute access, and governance arrangements.

The Collapse of the Capability Premium

Data from Epoch AI indicates that the cost to achieve a specific level of AI capability has dropped by approximately 47 percent per quarter since 2023. This rapid commoditization means the premium for frontier capabilities is short-lived; any advantage a lab gains is quickly eroded as open-weight models catch up. Strebulaev notes that while open-weight models like DeepSeek-v4.1-Flash processed nearly half of September’s tokens on Vercel’s AI Gateway, they accounted for only 2.6 percent of spend. This reveals a bifurcated market: high-volume, low-stakes workloads migrate to commoditized open models, while high-value, reliability-sensitive tasks remain with closed frontier labs.

From Product Logic to Infrastructure Logic

The essay applies Carlota Perez’s techno-economic paradigm theory to argue that the AI industry is transitioning from an 'installation' phase to a 'deployment' phase. In this context, frontier models are treated as products subject to commoditization, while governance and compute access are treated as infrastructure where durable rents can be extracted. The author highlights Anthropic’s recent moves, including a partnership with Amazon Bedrock and Google Cloud Vertex AI, as textbook examples of converting a product position into an infrastructure position. This mirrors OpenAI’s 2019 Microsoft deal, where exclusive cloud rights and Azure distribution secured long-term economic returns beyond mere model licensing.

The Anthropic-METR Governance Template

On September 12, 2026, Anthropic CEO Dario Amodei published 'Pace the Frontier,' calling for embedded third-party evaluators. This led to the documented Anthropic/METR arrangement, where six METR employees are embedded at Anthropic with office space and hardware. While METR can publish risk assessments, Anthropic retains the right to redact predefined categories such as customer data and model architectures. Strebulaev notes that METR is philanthropically funded, whereas Accenture, another evaluator, is directly funded by Anthropic. This asymmetry is critical: the deepest-access evaluator is the one not paid by the firm, yet the firm still controls the redaction categories.

Eight-Dimension Audit Results

The proposed framework audits governance regimes across eight dimensions, including evaluator independence, publication control, and treatment of open-weight models. When applied to the Anthropic template, the audit finds the arrangement 'contestable' on publication and 'potentially exclusionary' on access and proportionality. In contrast, the UK AI Safety Institute scored as the closest to a non-exclusionary design, while the EU AI Act’s conformity-assessment regime scored contestable on six dimensions. The analysis suggests the Anthropic template is structurally closer to China’s algorithm registration regime than to the EU or UK models, particularly in its bespoke, firm-funded nature and silence on open-weight treatments.

Key Takeaways

  • Inference costs have fallen 13x annually since 2023, rapidly commoditizing frontier capabilities and shifting competition to infrastructure control.
  • Anthropic’s governance template involves embedded evaluators (METR) with redaction rights, creating a potential barrier to entry for competitors.
  • The UK AI Safety Institute framework currently scores highest for non-exclusionary design in the eight-dimension audit.
  • The draft federal bill circulated by Anthropic would override state laws and permit government intervention only for risks exceeding $1 billion in harm.

The Bottom Line

The real story isn't that AI is hitting a ceiling, but that the floor is rising so fast that labs are building toll booths. If the new federal governance standards mirror the bespoke, firm-funded Anthropic template, we aren't getting safety—we're getting a moat.