Jensen Huang is lending his customers the money to buy his chips.

It sounds like a joke about a bad business model, and for three years the AI buildout could afford to let the joke stand, because the money kept moving. It has stopped moving for the marginal borrower, and the pattern underneath the joke is now sitting on the face of the filings. Since 2024, Nebius — the Amsterdam company that used to be the parent of Yandex and now sells AI computing capacity — has sold $2.7 billion of its own stock to Nvidia and other investors and issued billions more in convertible bonds for its AI expansion. Nvidia has announced that it intends to invest up to $100 billion in OpenAI, the customer at the centre of the demand story, with the money to arrive as OpenAI deploys the systems — which is to say, as OpenAI buys. This week SpaceX announced the largest AI-chip financing ever assembled, $40 billion, for the stated purpose of purchasing the latest lineup of Nvidia processors; the record survived a few hours, until Broadcom sealed a $63 billion arrangement to lease chips to Anthropic, a deal whose banks had to be steadied at the last minute by the personal intervention of Broadcom’s own chief executive, Hock Tan.

Follow one dollar around that system and watch what it does. Nvidia invests in the cloud company; the cloud company spends the investment on Nvidia systems; the cloud company signs long contracts to sell that capacity to the labs and the hyperscalers — the giant platform clouds, Google, Microsoft, Amazon, Meta; the labs and the hyperscalers raise equity and debt on the strength of the AI demand they are supplying to one another; and every pass of the dollar is booked somewhere as revenue and pledged somewhere as collateral. The arrangement has an old name. It is vendor financing — the seller lending the buyer the purchase price so that the sale can be booked — and it is the trick the late-1990s telecom buildout ran, the one that ended with Lucent, Nortel and WorldCom. Nortel, which was ours, financed its customers’ purchases on terms it spent the following decade regretting. The trick does not require anyone to lie. It requires only that each pass around the circle be counted as though a new dollar had entered the economy, and the counting, for three years, has been enthusiastic. It is true, in the narrow sense in which the word applies, that this is rational on Nvidia’s part; financing your customers is how a company with monopoly power locks in its market. The trouble is that a boom financed by its own suppliers is not a boom that has found outside believers. It is a boom circulating its own sales receipts.

For most of that time the story was told as a technology story with a financing tail. It was always a financing story with a technology tail. The bottleneck was never intelligence, and it was never chips. The bottleneck was the price of the next dollar, and October is when the price arrived. You can see the arrival at the cheapest bearing, where a tradesman would look first. Bonds CoreWeave issued earlier this year at a coupon of 9.75 per cent now change hands in the secondary market — the after-market where existing bonds trade, as distinct from the primary market where companies first sell them — at yields above fifteen per cent. For the reader who does not live in this world: a buyer of the paper today demands five and a quarter points more per year than the coupon, a premium of more than half, because the price of the bond has fallen far enough to make the arithmetic work. The company did not get worse at anything like this rate, this fast. The market repriced the only question that matters in credit — not whether the story is exciting, but who gets paid first when the refinancing comes due — and it repriced the answer violently. The reporting is that CoreWeave, Nebius and Nscale — the firms the industry calls neoclouds — have effectively been shut out of the bond market in recent weeks, and bankers say the high-yield market, where borrowers too risky for investment grade sell their paper, has closed for them too. Lenders have not become unreasonable. Lenders have read the same filings everybody else has.

The scale of what is being repriced deserves writing down, because the numbers have been arriving so quickly that they have stopped registering. AI-linked issuance ran to $394 billion across U.S. corporate debt markets in 2025 and $655 billion through mid-October of this year, on a Bank of America tally, with analysts forecasting close to $1 trillion of fresh paper next year. That is not a stock-market bubble in the shape of 1999; it is a debt boom in the shape of something older, and debt booms have a property worth respecting: the losses do not land on the people who bought the story; they land on the people who bought the paper. The past week offered a preview. Alphabet shed roughly $500 billion of market value on Thursday, its largest single wipeout ever, after quarterly results showed its search revenue slowing while its AI capital spending surged; Intel fell twenty-nine per cent on Friday on weakness in personal computers, the demand crack under the financing crack; and the Nasdaq’s 2.9 per cent drop was its worst session since June. The volatility we have been tracking for a fortnight is not noise around the story. It is the story arriving in prices.

What actually broke is worth precision, because the coverage has the habit of calling this a change in sentiment, and sentiment is not a thing that exists in the sense the phrase implies. In September the Federal Reserve raised short-term rates for the first time in three years. The tap did not close. It stopped being free. Then Oracle notified the developer of a large New Mexico data centre it had previously agreed to lease that it was pushing back the date at which full rent would begin, citing difficulties with power supply and permitting — difficulties now compounded, across the sector, by a growing local political resistance to the buildings themselves. That one scheduling letter repriced construction risk across the entire asset class, because it told every lender something the prospectuses had obscured: the binding constraint on the buildout was never capital — capital was the abundant input — and the scarce inputs were electrons and permits. A data centre is easy to count and very hard to finish, which is the kind of fact a tradesman checks before the banker does, and lenders everywhere began asking what their collateral actually was. In more cases than anyone had been pricing for, it was a building waiting on a substation.

Charles Kindleberger built the standard account of financial manias on an observation that fits this month precisely: manias expand through credit, and they do not end when the story stops being true — they end when the marginal borrower loses the market. The neoclouds were always the marginal borrowers. These are the companies that lease or own data centres, fill them with AI servers, and sell the capacity under long contracts to the hyperscalers and the labs; many of them began as cryptocurrency miners — the same buildings, the same grid connections, the same appetite for electricity, with the mining rigs swapped for racks of AI servers — which means a meaningful share of the sector’s institutional memory was formed selling compute into the last mania and pivoting when it ended. Most have histories in AI infrastructure so scant they can be dated. Nebius began life as the Amsterdam company behind Yandex, divested its Russian assets after the invasion of Ukraine, and pivoted to AI cloud services a little over two years ago; it holds contracts with Meta and Microsoft. The compute is real — Nebius’s machines run genuine workloads for customers who are not buying them for sport. The trouble is that the compute is financed like a boom: debt priced for demand curves that do not bend, layered onto operating histories you could fit on a restaurant napkin. Manias do not die of refutation. They die of a Fed meeting and a rent letter.

It is true that a market repricing risk is a market doing its job in public, and a bond paying fifteen instead of nine is the job being done. The trouble is that the repricing arrives after the money has been placed. None of the $655 billion was priced by buyers expecting the door to shut behind them, and the buyers were pension funds, insurers and bond desks. The institutions that lend to the hyperscalers have been feeling increasingly tapped out and concerned about their concentration in one theme, after a year in which Meta, Amazon, Alphabet and Microsoft flooded the investment-grade market and signalled their intention to keep flooding it. Dan Ivascyn, Pimco’s chief investment officer, put the new discipline plainly: the bar keeps rising for proving that the spending turns into revenue, and the question of the week is “how full can investors get on one name, or one theme?” His other line was blunter. “The weak links are going to suffer first.” The counterparties of last resort for the AI buildout are other people’s retirements. The market is not discovering the price of artificial intelligence. It is discovering the price of the paper — after the fact, in whatever quantity was already sold.

The neoclouds’ adjustment to all this is itself a document worth reading. Nebius’s chief executive, Arkady Volozh, said in Berlin last week: “From last week, probably some of the capital sources are temporarily closed, like the asset-backed or debt market… But we’re working now with customer advances and several other instruments.” The company’s chief operating officer has since said the asset-backed market remains open to Nebius, and both statements can be true at once. Neither changes what the pivot is. To be fair, at the level of a single company, taking prepayments is careful treasury management, the kind of thing any finance department should do when the window is closing. The trouble is what the prepayment is, once you look at it plainly. A customer advance is your customer paying you before you deliver, and when the customer is profitable and solvent, that is demand arriving early. When the customer is itself raising tens of billions in the same season, the advance is not demand. It is a loan from a borrower, wearing demand’s clothes — the accountants book it as a liability, the press releases book it as revenue — money that arrives before the product does, carrying worse disclosure than a bond and no secondary market at all, and a loan from a borrower belongs on the liability side of the story no matter which line of the filing it sits on. This is vendor financing with the polarity reversed, which means the boom’s revenue lines and its financing lines now run through the same pipe.

The rest of the instrument drawer reads the same way. Asset-backed financing, the market Nebius says remains open, means the loan is secured by the servers themselves — which sounds reassuring until you notice what the servers are: silicon on an annual refresh cycle, pledged against paper that matures on a five-to-seven-year one, worth whatever the next buyer will pay in a market that may by then be selling the same capacity at distress prices by the rack. Convertible bonds — loans that convert into shares on terms fixed in advance — let a borrower be debt or equity, whichever flatters the quarter.

John Kenneth Galbraith called this stage of affairs a bezzle: the interval between the taking of the money and the discovery of the loss, when the embezzler has his gain and the victim feels none of it. On the published schedule, the AI buildout’s bezzle is growing on time: $394 billion, then $655 billion, then, on the forecasts, nearly a trillion next year.

There is one piece of collateral with a blast radius wider than the bond funds. Anthropic’s IPO prospectus, filed this week, names as a risk factor the financing constraints facing the electrical-power providers its data centres depend on. Read that plainly: the buildout has grown to the size where its own suppliers’ ability to borrow is a listed risk to its shareholders, and those suppliers stand in front of the North American grid. When a power developer cannot finance an interconnection, the interconnection does not arrive. When it arrives at a cost the contract did not assume, the gap does not stay inside the contract. Utility commissions exist for precisely the moment when a private financing problem becomes a public rate problem. When the builders of the machines begin disclosing, in the documents they use to raise money, that the money for the electricity is uncertain, the bezzle has entered its candid phase. The bezzle does not create wealth. It borrows serenity, and this month it is charging fifteen per cent for the loan of it.

It is true that a great deal of real infrastructure gets built during episodes like this one, and Cory Doctorow’s distinction holds: this bubble is more WorldCom than cryptocurrency — overbuilt, but built. The substations will be real. The fibre will be real. The liquid-cooling plants and the buildings will be real, and someone will get years of discounted use out of them after the repricing. WorldCom’s fibre was real too. It was no comfort to the bondholders, because the people who financed the fibre were not the people who inherited it, and the income they had bought was priced off traffic that never came. The glut left fibre and left losses, and the two were distributed to different people, and the structure currently in place — debt on the asset, equity with the sellers, paper with the institutions — distributes them the same way. The residue argument is an argument about who benefits from the wreckage. It is not an argument that the wreckage was cheap.

I have watched this choreography at scales where you can see the faces. The mill in Selkirk where my father worked thirty years was bought in 1995 by a Brazilian steel company the way assets of that era changed hands — bought against the asset’s own future cash flows, the acquirer keeping the upside, the workforce and the town carrying the payments. The debt went onto the mill, a third of the line went out the gate, and the men who stayed worked another sixteen years beside the absence. Two decades later I watched the same performance from inside a Toronto software firm that an American holding company took apart over eighteen months — leverage first, fees second, the operating company last — and nobody called it a bubble either time; the first was modernization and the second was restructuring. The operators who acquired Canadian heavy industry between the mid-eighties and the mid-two-thousands ran the same structure so consistently that it stopped being a strategy and became a weather system. The choreography is now staged on a floor where the asset being levered is the credit market itself, which is to say the thing every pension and every insurance annuity in the developed world is standing on. My father’s trade had a rule about borrowing against equipment. You never do it to prove the machine works. The proof is the machine running on its own power. Everything else is the loan.

It is true that Meta, Alphabet, Microsoft and Amazon carry top credit ratings and could, on paper, carry a great deal more debt. The trouble is that the price of the marginal dollar is set by the marginal borrower, and the marginal borrower has just been shown the door; when risky yields move from nine to fifteen inside a season, the investment-grade desk feels the draft, and by Friday the hyperscalers’ own bonds had dropped toward record lows, their yields sitting well above typical investment-grade paper. That is why the largest names are hurrying. Meta, whose disclosed future spending commitments stood at roughly $825 billion when this publication noted its plan to raise up to $97.5 billion in new equity last week, is selling shares instead of borrowing, despite borrowing being the cheaper instrument on any honest cost-of-capital accounting; a fixed-income portfolio manager told the reporting, in so many words, that the equity raises are happening because the fixed-income community pushed back. Meta’s securities filing does the quiet part out loud: Alphabet, SpaceX, Anthropic and OpenAI are all lined up at the same capital window, and the institutions that buy hyperscaler bonds have begun pushing back on how much of one theme they will hold. A company issuing stock when debt is its cheaper money is not being bold. It is being told no by its natural lenders, and complying. The equity market is where credit-market fatigue goes to be absorbed, and the absorption has been visible all week.

The rest of the queue is adjusting in the ways available to it. Nscale, the neocloud whose board features the former Meta executives Sheryl Sandberg and Nick Clegg, has heard from its own bankers that the $30-to-35-billion valuation it wanted is unlikely to clear, and has slid its listing from November toward December while Nvidia and the other would-be backers of its $45-billion data centre in West Virginia for Anthropic demand tougher terms. Lambda cut the valuation it is seeking in its final pre-IPO round from $14.5 billion to $13 billion after the tech investment firm Coatue and other investors declined to participate. Anthropic is aiming for up to $100 billion in a mid-November offering. “Everybody is trying to raise money for their build-outs before the liquidity dries up,” Ben Connard of Carnegie Investment Counsel told the reporting, before adding the sentence that should be printed above the door of every data centre being financed this month: “Ultimately, there might not be enough money to fund this build-out.”

Notice what the loudest voices in this industry were talking about while the bond market closed. For three years the rhetorical apparatus around the buildout has pointed at 2040 — machine gods, alignment, existential risk. The pivot is on the record: in 2023, the same executives who signed a public statement about the risk of extinction from AI were the ones raising capital on the capability the statement implied, and the move has run ever since — relocating every concrete present question (who holds the paper, who pays for the power, whose pension is concentrated in one theme) behind a hypothetical future catastrophe. The pivot has always had two sets of customers: the builders, who prefer the attention forward, and the largest incumbents, who benefit when the regulatory conversation is about licences for the few instead of disclosure rules for the many. The question this month is almost embarrassingly concrete. What is a stack of AI-linked debt approaching a trillion dollars worth when its collateral is a circle of companies lending each other the money to keep the circle turning? The apparatus has no answer. The question is not on the slide.

If a public backstop is proposed for any of this — a federal loan window, a grid surcharge dressed up as reliability investment, a quiet guarantee under the data centres — the millwright’s answer is the right one: the people who took the spread on the way up do not get to hand the downside to ratepayers and bondholders on the way down. What the situation actually needs is boring and checkable, and the honest policy position is the opposite of a rescue. Issuers should disclose, prominently and as financing, revenue that flows from counterparties who are also their investors or their suppliers, and customer prepayments along with it — related-party paper and advance payments are not revenue, and the distinction is not subtle enough to require tenderness, so it belongs at the front of the filing where it cannot hide. Financial-stability authorities should map and stress-test AI-linked credit concentration with the same enthusiasm they eventually displayed for subprime, before the pension funds discover their weighting at fifteen per cent. And the genuinely binding constraint is public works, not finance: utility commissions should treat data-centre interconnection as the ratepayer question it is, and if Ottawa and Washington want this compute built, the grid interconnections should be built in public and charged for honestly; if they do not, the paper should carry its own risk all the way down to its last holder. What should not happen is what happened last time — the shortfall discovered publicly and settled asymmetrically, the losses mutualized, the gains retained, the invoices for the rescue arriving on the same desks that just finished explaining they were full.

Markets have forgiven circles before; there is a version of this in which the paper rolls over at better terms and the circle keeps turning on borrowed time. But what arrived in October is not sentiment. It is the discovery phase of a bezzle, and discoveries do not negotiate. The machines, it should be said plainly, will be real. The electricity contracts will be real. So will the losses, which are the one part of a debt boom that never fails to materialize on schedule. The next instalment of the buildout is already planned — close to a trillion dollars of fresh paper next year — and deadlines are the only part of financing that borrowers have ever respected. This one falls due whether or not the revenues do. A data centre that only runs while the money runs is not infrastructure. It is a debt with a building around it, and fifteen per cent is the market’s opening bid on finding out whose problem it turns out to be.