OPEC for Tokens?

In the space of twenty-four days, the United States built a release gate for artificial intelligence, and almost nobody priced it.

The dates are public. On June 2, 2026, an executive order established a “voluntary” framework in which frontier labs provide the federal government thirty days of pre-release access to new models for cybersecurity review, with a classified benchmarking regime — due August 1 — to define which models are covered. On June 12, the Commerce Department ordered Anthropic to suspend global access to its two most advanced models, days after launch; access was restored beginning July 1. On June 25, OpenAI agreed to launch its newest model only to government-approved customers first, at the White House’s request.

Voluntary in name. Two for two in compliance. A clearance layer for frontier intelligence now exists in the United States — uncodified, non-statutory, and operating.

The order itself insists otherwise, in words worth quoting exactly: “nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models.” Ten days after those words were signed, the Commerce Department suspended two released models by export-control order. The gap between the text and the practice is this ledger’s first entry.

The security rationale is genuine and we do not dismiss it: models that can autonomously find and chain software vulnerabilities are a real concern, and a thirty-day review is a rational option on information before irreversible diffusion. But this is an economics desk, and a release gate has an economic identity independent of its motives. It is a quantity restriction on the supply of intelligence.

What a cartel actually does

OPEC never managed oil reserves; it managed flow. The price of a barrel was defended not by scarcity of oil in the ground but by control of how fast it reached the market. A release gate does to frontier capability what production quotas do to crude: it slows the arrival of supply — in this case, of cheaper, better intelligence — and thereby supports the price of what is already deployed.

Here is why that matters now, and specifically now. For three years, equivalent-capability inference prices collapsed at a median of roughly 50x per year. Our efficiency series flagged an anomaly this week: in 2026, that deflation has decelerated to roughly 6 percent year-to-date. At the same time, the market leader in AI hardware moved to raise H100 rental pricing — into what is, by every physical measure, a growing glut of capacity. Prices firming into oversupply is the signature of managed flow, not of scarcity.

We are careful about causality: the deceleration has several candidate causes — maturing model generations, hardware cost floors, demand mix. But a de facto gate on frontier releases is now on that candidate list, with dates attached, and it is the only candidate that arrived by executive order.

Who the gate protects

Follow the incidence, as an economist would. Roughly $540 billion of committed compute rests on about $35 billion of funded outside equity — the 15.5x recycling loop we measure. That structure has one existential dependency: the price of intelligence must not collapse faster than dollars arrive. Every quarter that token deflation runs slower, existing capacity earns longer, depreciation schedules look more defensible, and the loop’s collateral holds.

A release gate — whatever its intent — is functionally capex-protective. It slows the very force (frontier diffusion) that would strand the installed base fastest. If you wanted to design a policy that defends the economics of the build-out while wearing security clothing, it would look exactly like a pre-release review regime that binds hardest on whoever moves fastest.

The money row

There is a second ledger to read. In 2025, Anthropic and OpenAI spent $3.13 million and $2.99 million respectively on direct federal lobbying — both firms’ first serious years in Washington. In the first quarter of 2026 alone, both posted records: $1.6 million for Anthropic (more than four times its year-ago quarter) and $1.0 million for OpenAI. The disclosed topics read like the executive order’s table of contents: AI procurement, Defense Department procurement, cybersecurity, acceptable-use policy, cloud infrastructure.

George Stigler’s 1971 observation was that regulation is acquired by the industry it governs and operated for its benefit. The prediction of capture theory is not that incumbents resist a gate — it is that they help build it. A thirty-day review is a rounding error for a lab with a Washington office and a procurement pipeline. It is a wall for a small lab, and a moat against open-weight releases. Both incumbents complied within days, publicly and gracefully. That is either good citizenship or Stigler’s textbook, and the observable that separates the two arrives on a schedule.

What decides it

The classified benchmark defining “covered frontier models” is due August 1 — the executive order’s own 60-day deliverable. Its breadth is the tell: if the definition catches the incumbents’ own releases as firmly as everyone else’s, the capture read weakens. If gate reviews shorten as evaluation tooling matures, the precaution read strengthens. If token deflation re-accelerates despite the gate, OPEC-for-tokens fails as a mechanism and we will say so. And if lobbying disclosures keep setting records while the gate’s definition lands just above the smaller labs’ capabilities — the ledger will show it, quarterly, in filings anyone can read.

That is what The Policy Ledger is: the same discipline we apply to 10-Qs, pointed at the Federal Register and the lobbying disclosures. Policy is not politics. Politics is opinions; policy leaves documents. We read documents.

The Policy Ledger and the release-gate timeline: The Policy Ledger · the efficiency anomaly: Jevons or DWDM? · the loop it protects: The Ground Truth Tape