The discourse on artificial intelligence is dominated by two claims that are rarely separated: that AI development is hitting technical limits, and that leading laboratories are shaping regulation to protect their market position. Both capture something. Neither explains why these two developments are occurring together. This essay argues that the convergence is best understood as a change in the economic gradient of deployment, not a slowdown in capability — and that the change matters because it alters what firms are competing for. The cost of achieving a fixed level of capability has fallen roughly thirteenfold per year since 2023. Production traffic is migrating to commoditized open-weight models for high-volume workloads, while frontier labs retain high-value, reliability-sensitive tasks. When differentiation becomes harder to monetize, control over the infrastructure layer — distribution, compute, and the governance arrangements that determine access — becomes the site of rent.
The empirical shift: measured traffic and inferred deployment
Epoch AI finds that the cost of achieving a given capability level has fallen roughly 47 percent per quarter — about 13x per year — since 2023, with the decline steepest immediately after a capability becomes state-of-the-art (66 percent per quarter initially, decaying to 32 percent two years later). The premium for frontier capability is therefore short-lived: any given capability is rapidly commoditized, and the frontier lab’s advantage is a moving target that must be continuously re-earned at rising cost.
Vercel’s AI Gateway Production Index, based on traffic through October 2026, reports that open-weight models accounted for 71.5 percent of gateway token volume and approximately 25 percent of estimated spend, up from roughly 5 percent of spend in June. DeepSeek-v4.1-Flash alone processed nearly half of September’s tokens while accounting for roughly 2.6 percent of gateway spend. Yet 68 percent of teams that increased spending on open-weight models also increased spending on closed-weight models, and frontier spend grew approximately 13-fold since December while average estimated cost per token fell 47 percent. The pattern is a split, not a replacement: high-volume, low-stakes workloads migrate to commoditized open models; high-value, high-reliability workloads remain with frontier labs.
The mechanism is straightforward: falling inference cost for fixed capability makes free-riding on open models rational for the commoditized layer. As the economic gradient steepens — frontier costs rise faster than frontier returns — the rational follower strategy shifts from “catch the frontier” to “deploy the commoditized capability”. The capability gradient closes from below, jaggedly, not uniformly. Open-weight models lagged the closed frontier by roughly four months on Epoch’s benchmark index as of mid-2026 — a gap that has not consistently narrowed. This is expected: the mechanism predicts deployment migration, not benchmark convergence.
Caveat: The Vercel figures describe gateway traffic, not the entire AI market, and rely on list-price estimates and a model classification that has broadened over time. They are used here as a convergent indicator, not a census. Token volume does not establish the economic value of tasks, model quality, or the profitability of each segment.
From product logic to infrastructure logic
Carlota Perez’s model of techno-economic paradigms divides each technological revolution into two broad phases: installation and deployment. During installation, financial capital floods into the new technology, building infrastructure and generating speculative bubbles. The transition to deployment requires the disciplining of finance capital and the rise of productive capital, often through institutional reform, crisis, or both. Applied to AI, the installation period corresponds roughly to 2020–present: massive investment in data centers, GPUs, and training runs. The infrastructure will remain after the speculative fever subsides. The question is who will control it and on what terms.
The distinction that matters here is between product logic and infrastructure logic. Product logic extracts margin from differentiation; when differentiation declines, the rational responses are exit, consolidation, or a new product. Infrastructure logic extracts rent from guaranteed access; when differentiation declines but demand for the layer remains long-term and general, the rational response is to control the conditions of access to the layer. The margin is lower, but it is durable. This is why firms invest in electricity generation at thin margins: not because the product is differentiated, but because the layer is indispensable.
Applied to AI, the frontier model is a product; distribution, compute, and governance layers are infrastructure. Commoditization of the product layer does not eliminate the infrastructure layer. It makes the infrastructure layer the site of rent. The conversion can be observed in microcosm. OpenAI’s 2019 partnership with Microsoft — exclusive cloud rights, Azure OpenAI Service as a distribution layer, Microsoft’s 49 percent share of economic returns — is a textbook case of a frontier lab converting a product position into an infrastructure position. Anthropic’s partnerships with Amazon Bedrock and Google Cloud Vertex AI follow the same pattern. The conversion is not concealed; it is disclosed in financial filings.
The Anthropic / METR arrangement
On 12 September 2026, Anthropic CEO Dario Amodei published “Pace the Frontier,” a three-part essay calling for embedded third-party evaluators and a coordinated industry slowdown. Sam Altman and Elon Musk publicly endorsed the call within hours. The timing was coordinated and the substance was convergent. What the statements establish is that the leading laboratories converged, within a short window, on a specific institutional form: the embedded evaluator with privileged access.
The most documented instantiation is the Anthropic/METR arrangement, announced in October 2026. Six METR employees are embedded at Anthropic with company hardware, office space, and access roughly on par with internal risk teams. METR can publish its risk assessments; Anthropic cannot veto publication but can require redaction of predefined categories: customer data, attorney-client privileged information, and information about model architectures. Three access tiers are being built. Accenture is a parallel evaluator under a partnership in which both firms committed approximately $1 billion each over five years to embedded-evaluation work; Anthropic funds Accenture directly. METR is not paid by Anthropic; it is philanthropically and government funded. The asymmetry is the point: the deepest-access evaluator is the one not paid by the firm.
A draft bill Anthropic has circulated would generalize the template. According to reporting by Politico and Inside AI Policy, the bill requires frontier developers to submit to independent evaluation, requires evaluators to register with the Department of Commerce, allows labs to pay evaluators provided compensation is not tied to results, and overrides certain state laws such as California’s. The bill is reported to permit government intervention only when a model poses a foreseeable, imminent risk of more than $1 billion in harm. A federal standard designed by the firm displaces a patchwork of state standards that might have been more demanding.
The eight-dimension audit framework
The empirical case yields a reusable framework. The framework is a multidimensional audit, not a single score or a binary classification. Its job is to expose trade-offs and identify which design features need justification, rather than to assume that every incumbent advantage is illegitimate. For each dimension, the audit proceeds in three steps: observed design feature; mechanism producing a barrier; evidence of competitive effect.
The eight dimensions are: (1) independence of evaluators; (2) appointment, funding, oversight; (3) control over publication; (4) access to evaluation infrastructure; (5) proportionality; (6) treatment of open-weight models; (7) cost distribution; (8) less-restrictive alternatives. Applied to the Anthropic template, the framework finds the arrangement contestable on publication (METR publishes without veto, but Anthropic defines redaction categories) and potentially exclusionary on access, proportionality, and open-weight treatment. The word “potentially” is load-bearing: lack of shared infrastructure is an observed feature; its exclusionary effect is a further claim that requires evidence about actual or expected entry costs.
The framework is then validated against three further cases. The EU AI Act’s conformity-assessment regime scores contestable on six dimensions and mixed on two. China’s algorithm registration regime scores contestable on three, mixed on two, potentially exclusionary on two. The UK AI Safety Institute scores contestable on seven dimensions and not addressed on one — the closest to a non-exclusionary design. The comparison reveals that the Anthropic template is closer to the China regime than to the EU or UK regimes on several dimensions: bespoke rather than shared, firm-funded rather than taxpayer-funded, silent on open-weight rather than addressing it.
What this does and does not claim
The essay does not infer motive. A regime designed with the best of intentions can produce exclusionary effects; a regime designed with cynical motives can produce genuine safety benefits. What matters is what the design does — who is subject to it, who controls it, who bears its costs, who can contest its findings. Three non-exclusive explanations deserve equal consideration: H1 (strategic incentive — firms favor governance designs that preserve their position in the infrastructure layer); H2 (genuine safety concern — the proposals are a response to documented incidents); H3 (regulatory uncertainty — firms prefer a federal standard they can shape to a state-by-state regime). The evidence reviewed here cannot adjudicate among them. The paper’s contribution is not to choose among them, but to show that the design features are consequential regardless of which explanation is correct.
The strongest version of the argument is not that AI safety governance is regulatory capture. It is that legitimate safety governance must be evaluated not only by the risks it mitigates, but also by the distribution of access, authority, and compliance costs it creates.
What to watch
If the draft bill, when introduced, includes government funding, shared infrastructure, proportionality provisions, and an open-weight regime, the exclusionary prediction is disconfirmed on those dimensions. Watch for: the cost structure of the registry (firm-funded, government-funded, or hybrid); whether a shared evaluation facility is included or each lab negotiates bespoke access; whether simplified pathways for smaller developers are written in; whether open-weight models receive a distinct regime or are excluded by design. The design choices are being made now, and incumbent firms are playing a significant role in making them. That is the fact the essay establishes. What follows from it is a matter of design, not destiny.
I will be publishing regular case-by-case audits in this framework. Subscribe if you want them in your inbox.
— Sergei A. Strebulaev, Independent Researcher (ORCID 0009-0006-2710-8910).