Amazon, Google, Meta and Oracle have borrowed against compute the world may not need. The International Monetary Fund’s Global Financial Stability Report, released 13 October 2026, named an AI share price crash as one of three “salient risks” to global financial stability, sitting alongside rising government borrowing costs and a major cyber-attack, and it described the transmission in plain terms: high valuations, rising leverage, concentrated exposures and circular financing arrangements, such that a correction would “sharply reduce investment, trigger equity market corrections, and tighten financial conditions, with global spillovers.” This is the same fund whose managing director named the AI boom alongside energy shocks and sovereign debt as threats to global growth six days earlier. The institution that lends to governments in distress is now describing the American technology sector in the vocabulary it usually reserves for sovereigns.
It is true that the IMF is not a technology assessor, and a fund that models balance sheets is not entitled to a verdict on model capability. That concession is worth making precisely because the most consequential sentence in the report is not financial at all. The fund said there is “growing evidence” that successful AI deployment “may not require investment on the massive scale being undertaken by ‘hyperscaler’ firms,” and that “potential economic gains may not be as large as expected.” Strip the hedging out of that and it says something an engineer can check: the returns to scale on this build may be breaking down. The trouble for the capital structure is that $650 billion of debt was raised on the assumption that they are not.
What the money buys is worth being precise about. A frontier training run is not a rack in a back room. It is tens of thousands of accelerators, Nvidia’s parts almost entirely, bound together with high-bandwidth interconnect — NVLink, InfiniBand — tightly enough that one model trains across the whole cluster as if the cluster were a single machine, for weeks at a time, at full draw. The engineering constraint that follows is not the chip count but the power and the heat: a campus of this kind is planned in hundreds of megawatts, some approaching a gigawatt, with switchgear, cooling loops and multi-year silicon orders committed years before a single token reaches a customer, and the dense racks now run hot enough that liquid cooling has stopped being optional. The capital goes into concrete and electrical capacity first. Silicon second. Software somewhere after that.
The economics then split into three phases, and the IMF’s sentence lands exactly on the seams between them. Pre-training is the capital-intensive phase — a fixed, largely sunk cost measured in cluster-weeks, where capability improves with compute along a power-law curve that has not stopped holding but whose slope has been paying less per dollar for some time, each successive frontier costing multiples of the previous run’s training compute for gains an ordinary user can perceive. Adaptation — fine-tuning a released model on a domain, distilling it into something smaller — is cheap by comparison. Inference, the phase that actually runs in production, is a marginal-cost business: the serving fleet is sized to query demand rather than to prestige, distilled and open-weight models now carry much of the traffic that recently required a frontier system, and serving capacity is measured in megawatts and tokens per second, not in frontier jumps. If the durable value sits in adaptation and inference rather than in the next pre-training run, then the marginal data centre is capacity without corresponding capability. That is a technical proposition before it is a market one, and it is roughly what our own reporting described in September when Wall Street kept buying the hyperscalers while their executives were already describing a slowdown.
Which brings the report to its most uncomfortable table. The fund counted $300 billion in transactions between the hyperscalers and weaker AI firms over three years, supplemented by $650 billion in debt, in exchange for $1.4 trillion in future financial commitments. The mechanism deserves its own sentence, because it is not demand in any ordinary sense of the word. A weaker AI firm can pay for cloud capacity because the hyperscaler holds its equity, extended it cloud credits, or underwrote the debt that funded the capacity purchase. Revenue books upstream. The exposure that produced it books downstream. The same dollars appear on both sides of the sector’s ledger as though they had been earned twice, when what has actually happened is that a supplier has arranged its own customer and then reported the sale. Cory Doctorow’s chokepoint framing is the useful instrument here — an intermediary positioned between compute and demand, collecting on every transit, including the transits it arranged itself — and the arrangement is legible in the filings if you read the footnotes rather than the capex line. We reported the shape of it in September: $132 billion in datacentre debt already stacked against a build-out whose customers were, in part, the borrowers’ own portfolios.
The leverage is layered rather than located. The IMF identified three sources: debt contracted by the AI companies themselves; hedge fund positions in the sector; and retail investors borrowing to buy AI stocks. The third changes the mechanics of a correction, because borrowed retail money does not wait — a margin call is an order, not an opinion, and forced selling arrives into a book where seven stocks now carry more than 35 percent of the S&P 500, up from 12 percent a decade ago, and where their combined market capitalisation has risen more than 50 percent in two years. Concentration that behaves well on the way up behaves predictably in reverse: index weight of that size means nearly every diversified American retirement account is already long this trade, whether or not anyone chose it, and the selling, when it comes, comes from every direction at once.
It is worth saying plainly that none of this makes the systems useless. The models work. Hospitals, laboratories and firms use them daily, and a correction in AI asset prices would neither un-ship the software nor un-write the papers. At the mill the distinction was routine, and it was the whole of the discipline: you bought iron against orders that had been signed, and the question in the gate meeting was always who had signed for the tonnage, and what happened to the order if he changed his mind. What the IMF has documented is capacity bought against commitments arranged by the same balance sheet that will book the revenue from it. That is not an order. It cannot be inspected like one.
The fund’s exposure map is honest about where the bag sits. The United States is most exposed, Asian markets next, the UK and euro area “potentially more insulated” — which is a technicality for any pension fund holding US equities, and the IMF says as much. The remedy does not require anyone to be right about the ceiling of the technology. It requires the exposures to be legible: stress-testing the concentrated positions under a correlated AI drawdown rather than under the standard market scenarios; disclosing take-or-pay capacity commitments and vendor-financed arrangements with the same prominence as the capex line, so that a reader of the filings can separate booked demand from arranged demand; and treating the leverage embedded in retail margin balances as the systemic quantity it has become.
Galbraith’s bezzle — the interval in which the money is gone and the loss has not yet been felt — is the right instrument for the quarters ahead, and the IMF has now told you what ends it: not a research finding about scaling, but a revenue line that fails to arrive on schedule. The report does not say AI is a bubble. It says the balance sheets rest on an engineering assumption that is now under test, and that the test is being run against savings accounts nobody in the sector has met.
A commitment is not a customer. That is the whole of it.