The AI Slowdown Pact Is Private Monetary Policy

Over the weekend of 13 September 2026, Sam Altman, Dario Amodei, Demis Hassabis and Elon Musk loosely converged on something none of them had signed: “pace the frontier”. There was no contract, no regulator, no verification mechanism. There was an essay by Amodei laying out a three-step proposal — external auditors embedded inside the labs, domestic regulation and a global slowdown agreement — plus several public nods. On Monday, Nvidia, Intel and AMD fell, bitcoin rose and traded around $76,900, and Donald Trump phoned Jensen Huang while Huang was on stage at the All-In Summit in Los Angeles to tell him that worrying about AI is “a hoax”.

That is the relevant fact: a verbal agreement between four people moved the price of global compute infrastructure in a single session. It was not moved by a rule, or a fine, or a shift in demand. It was moved by a signal of intent. In any other market that has a precise name: guidance. And whoever issues credible guidance on a scarce asset is not doing applied ethics, they are doing monetary policy.

Compute is this decade’s monetary base

The analogy is worth taking seriously because it is structural, not rhetorical. A central bank sets the price of money by regulating the quantity available and, above all, by communicating the future path of that quantity. Ninety per cent of the effect sits in the expectation, not the operation. That is why Fed minutes move curves before anything is executed.

In frontier AI, the equivalent scarce asset is the combination of committed compute capacity and the reachable capability frontier. Four firms account for a disproportionate share of demand for high-end accelerators, of take-or-pay contracts with clouds, and of the multi-year commitments that underwrite capex across the supply chain. When those four firms hint that they will slow the rate at which they push capability, they are communicating a demand path. The market does not wait to see whether they follow through: it reprices.

The most interesting concrete proposal in the whole episode points in exactly that direction. Daniel Kokotajlo’s AI Futures Project — Kokotajlo is a former OpenAI researcher — argues that labs should open their compute budgets to auditors and commit to cutting them significantly for research. Translated: a verifiable quantitative target on the scarce aggregate, with a third party validating the number. That is not a code of conduct. It is a targeting regime with independent oversight. The institutional architecture is identical to a central bank’s, with the difference that nobody granted the mandate.

Competition law has no doctrine for this

Here is the hole. An agreement among the four largest operators in a market to cap the rate of product improvement and cut R&D spending is, in any antitrust textbook, an immediate candidate for a restraint of competition: a horizontal agreement limiting output and innovation. Classic cartels do precisely this, and they have spent decades dressing it up as an industry standard, a quality code or an environmental commitment.

The problem is that the defence here is not obviously false. If the risk being invoked is real — Amodei identifies recursive self-improvement as the trigger, and Anthropic has said that threshold could arrive in early 2027 — then limiting cadence has a public-interest justification no cement cartel could ever claim. And competition law lacks an operational test to distinguish “we coordinated so as not to cause catastrophic harm” from “we coordinated so we wouldn’t have to compete and, incidentally, to shut out open source”. It is not that the answer is hard: it is that the test does not exist.

The criticisms gathered by The Verge land exactly there. Sacha Haworth, of the Tech Oversight Project, puts it bluntly: any voluntary framework is regulatory capture and “we shouldn’t let the foxes run the henhouse”. Kokotajlo fears the middle scenario, which is the most likely one: “they’ll bring in external auditors, they’ll do a lot of safety paperwork — some of it genuinely good — but at the end of the day it won’t slow them down much”. Nick Reese, former director of emerging technology policy at DHS, compares the move to the social platforms’ playbook a decade ago, when they asked for regulation in order to pre-empt something worse.

And one detail settles a great deal: the calls for a slowdown are aimed exclusively at frontier labs defined by exact size metrics. Compute thresholds, de facto. That is defensible on safety grounds and, simultaneously, it is the cleanest way to build a barrier to entry: whoever is already inside negotiates the threshold; whoever comes next inherits it.

Why capex is the real transmission channel

The financial side is where this stops being a think-tank debate. AI capex is being financed with structures that assume a rising, uninterrupted demand curve: corporate debt, data centre leasing vehicles, multi-year purchase commitments, valuations that discount future capacity already sold. All of that scaffolding is sensitive to duration, not to level: an eighteen-month delay in the capability curve does not reduce total demand, but it wrecks the amortisation schedule of anyone who built for a peak that now arrives later.

Which is why a statement with no contract behind it moves silicon makers and infrastructure operators before it moves the labs themselves. And why Decrypt reported the reaction it did: bitcoin up while Nvidia, Intel and the rest of the semis fell. The simple reading is “rotation into an uncorrelated asset”. The more useful reading is different: the AI trade has become a policy trade, and anyone exposed to it needs a hedge against discretionary decisions by four chief executives and one president. Bitcoin is not a good hedge against AI; it is simply the most accessible liquid asset that does not depend on the GPU deployment calendar.

A pacing cartel only works if the state blesses it

This is the part that closes the loop and that most of the coverage treats as colour. A private agreement to limit cadence is unstable by definition: the first to break ranks captures the frontier. It only holds if someone punishes defection. In practice, that means the state.

And the state answered within 48 hours, in the opposite direction. Trump posted that the only guardrail AI needs is “a STRONG AND SMART president”, and on the phone with Huang he added that the slowdown advocates “are playing into the hands of people who don’t want this to happen: political people, or China”, and that “we’re not going to let it happen”. Huang, live on stage with the phone on speaker in front of the audience, replied: “You’re right, sir. We’re not going to let that happen”.

In other words: the supplier of the scarce asset and the government that controls permits, energy and export controls are explicitly against the pact. A cartel with no coercive power and with the state against it is not a cartel: it is a letter of intent that will last until the first uncomfortable benchmark.

China completes the asymmetry. Foreign Ministry spokesperson Guo Jiakun dismissed the slowdown calls as “alarmism”. The optimists about the agreement — Tyler Johnston, of the Midas Project; Buck Shlegeris, of Redwood Research — argue that the right analogy is nuclear non-proliferation and that no government has an interest in reckless development. The argument is reasonable and it has an enormous practical problem: non-proliferation was verified through inspections of physically locatable fissile material. Here the equivalent would be auditing compute budgets, and nobody has explained how you verify that in jurisdictions that do not cooperate. Until that answer exists, AI safety is a variable of industrial policy, not of ethics.

What would falsify this reading

The monetary-cartel thesis has one weak point that needs saying out loud: it assumes these four firms have effective power over cadence, and they may not. If recursive self-improvement is real and works, the binding constraint stops being aggregate compute and becomes the algorithm; a mid-sized lab with access to open models and one good idea breaks the ceiling without needing a frontier cluster. In that world, the pact fixes nothing: it only penalises whoever signs it.

Second objection, the more uncomfortable one: the motivation may simply be what it looks like. The resignation letter of Jacob Coxon, an Anthropic researcher — “the people building AI sincerely believe it could kill us all before the end of the decade” — has racked up more than 170 million views on X alone, and it comes after a public letter signed in July by more than 1,000 AI lab employees calling for a slowdown. Kokotajlo is explicit: this is not the CEOs’ idea, it is years of outside pressure that they are now “caving to while taking the credit”. Cartels do not usually emerge from a revolt by their own engineers.

And third: social pressure runs along a different vector from the financial one. A Gallup poll puts at seven in ten the share of Americans who oppose building data centres in their area, with more than 50% citing the impact on environmental resources and around 20% the cost of living. That is not an international psyop, as Trump and part of the venture capital world suggest: it is local opposition to infrastructure, the most predictable phenomenon in the world. If a slowdown does eventually arrive, it may be imposed by a planning commission before it is imposed by an Amodei essay.

What to do with this if you build or invest

The operational point: your product’s regulatory risk has stopped being legislative and become contractual and private. If you depend on a frontier lab’s API, the variable that can break your roadmap is not a European law with a two-year legislative process: it is a pacing decision announced on a Sunday via essay, which translates into less available inference capacity, rising per-token prices and more widely spaced capability jumps. That is modelled as concentrated counterparty risk, not as regulatory risk. And it is mitigated the same way: genuine multi-provider setups, not nominal ones; model abstraction at the product layer; and an honest assessment of how much of your competitive edge disappears if the frontier freezes for eighteen months — because if all of it disappears, your product was the model, not you.

And what to watch to know whether the pact is real: a single metric. If any lab publishes its research compute budget with third-party verification — METR, Apollo, Redwood, whoever — this is a targeting regime and the entire supply chain has to be repriced. If what shows up instead is voluntary frameworks, auditors with limited access and oversight committees in the style of Meta’s Oversight Board, then it was safety-washing and capex carries on. The gap between those two scenarios is measured in hundreds of billions, and as of today all that exists is a weekend conversation.