AI Governance Shifts to Infrastructure Control as Costs Plummet

Overview

The global conversation surrounding artificial intelligence is increasingly dominated by two interconnected narratives: the notion that development is hitting hard technical limits, and the claim that major laboratories are shaping regulation to cement their market dominance. While both perspectives capture elements of reality, they often fail to explain why these trends are occurring simultaneously.

A shifting analysis suggests the convergence is better understood as a dramatic change in the economic gradient of AI deployment, rather than a simple slowdown in capability. Since 2023, the cost of achieving a fixed level of AI capability has fallen by approximately thirteenfold per year. As differentiation becomes harder to monetize at the application layer, the locus of economic value is migrating upward to the infrastructure layer.

Firms are now competing primarily for control over distribution, compute resources, and the governance arrangements that determine access to them. This transition marks a fundamental shift from product-based margins to infrastructure-based rents.

The Collapse of the Capability Premium

The core driver of this structural shift is the rapid commoditization of high-level AI performance. Data from Epoch AI indicates that the cost of achieving a given capability level has dropped roughly 47 percent per quarter since 2023. The decline is steepest immediately after a technology becomes state-of-the-art, starting at 66 percent per quarter and decaying to 32 percent two years later.

This means the premium for frontier capability is remarkably short-lived. Any specific advantage held by leading laboratories is quickly replicated and commoditized by the broader market. Consequently, the frontier’s competitive edge becomes a moving target that must be continuously re-earned at escalating costs, rather than a stable source of long-term profit.

"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 gap between open-weight models and closed frontier models remains distinct. As of mid-2026, open-weight models lagged behind closed frontiers by approximately four months on benchmark indices. This gap has not consistently narrowed, which aligns with predictions that the mechanism favors deployment migration rather than benchmark convergence. Rational actors are shifting strategies from trying to catch the frontier to deploying commoditized capability where it is sufficient.

Traffic Patterns and Market Splitting

Empirical data from the Vercel AI Gateway Production Index highlights a clear bifurcation in how AI is being used in production environments. By October 2026, open-weight models accounted for 71.5 percent of gateway token volume, a significant increase from earlier in the year. Despite this volume dominance, they represented only about 25 percent of estimated spend.

The economics of this split are stark. DeepSeek-v4.1-Flash alone processed nearly half of September’s tokens while accounting for roughly 2.6 percent of gateway spend. This pattern confirms that high-volume, low-stakes workloads are migrating to commoditized open models, while high-value, reliability-sensitive tasks remain with frontier labs.

However, this is not a zero-sum replacement. Sixty-eight percent of teams that increased spending on open-weight models also increased spending on closed-weight models. Frontier spend grew approximately 13-fold since December, even as the average estimated cost per token fell by 47 percent. The market is expanding rather than contracting, but the profit centers are separating along lines of workload criticality.

Token volume figures describe gateway traffic and should be viewed as convergent indicators rather than a complete census. They do not fully establish the economic value of tasks or the profitability of each segment, but they clearly illustrate the direction of travel for enterprise deployment strategies.

From Product Logic to Infrastructure Logic

To understand the strategic implications, it is useful to apply Carlota Perez’s model of techno-economic paradigms. This framework divides technological revolutions into an installation phase, characterized by financial capital flooding into new technologies and speculative investment, and a deployment phase, where productive capital takes over and infrastructure stabilizes.

For AI, the installation period spans roughly from 2020 to the present, marked by massive investments in data centers, GPUs, and training runs. As the industry moves toward deployment, the question becomes who controls the resulting infrastructure and on what terms.

The critical distinction lies between product logic and infrastructure logic. Product logic extracts margin from differentiation; when differentiation declines, the rational response is exit, consolidation, or product innovation. Infrastructure logic, however, extracts rent from guaranteed access. When differentiation declines but demand for the layer remains long-term and general, firms invest in controlling the conditions of access.

Applied to AI, the frontier model operates on product logic, while distribution, compute, and governance layers operate on infrastructure logic. Commoditization of the product layer does not eliminate the infrastructure layer; it makes the infrastructure layer the primary site for rent extraction. This mirrors the energy sector, where firms invest in electricity generation at thin margins because the layer is indispensable.

As leading labs navigate this transition, their 2019 partnerships and subsequent structural moves reflect a growing recognition that controlling the underlying infrastructure is more durable than relying solely on proprietary model performance. The governance regimes being built today are likely to define the economic landscape for decades, prioritizing access control over pure capability leaps.