The largest chipmaker of the age is assembling, with Wall Street’s biggest firms, a financing vehicle of up to five hundred billion dollars — a bank at its head, and the largest private-capital houses structuring it — to fund the building of artificial-intelligence data centers. The mechanism is not a sale but a loan machine: instead of customers paying cash for chips and computing capacity, institutional money is routed to them so they can finance the purchases and the construction, turning compute into a financeable asset the way real estate or aircraft are financed. Reporting noted the sum is roughly twenty times the debt that inflated the telecom bubble of the late nineties, and that some of the money would come from ordinary retirement savings. A famous investor called it a Wall Street stunt. The buildout has found a way to keep going that no longer depends on anyone actually having the money.
The Seller Funds the Buyer
Strip the arrangement to its bones and the strange loop is visible. The company that sells the chips is helping to organize the money that its customers will borrow to buy the chips — which means the demand for its product is being financed, in part, by an effort the seller itself is orchestrating. This is vendor financing, an old technique, and it has an old warning attached: when a supplier funds its own customers’ purchases, the revenue it books can be a measure of the credit it extended rather than of demand that would have existed anyway. The buildout looks like a market of eager buyers paying for compute; underneath, some of those buyers can pay only because the seller arranged for them to be lent the money, and the line between selling a thing and financing the appearance of demand for it grows thin.
The reason to reach for a machine this large is that the ordinary sources of money have limits the ambition has outgrown. The technology companies at the center of the boom have enormous balance sheets, but not five-hundred-billion-dollar ones to spare on data centers, and the appetite for compute has run ahead of what even the richest firms can fund from their own cash. So the money is sought elsewhere — in private credit, in institutional funds, in the pooled savings of pensioners — because the buildout has reached a scale that cannot be paid for out of profits and must instead be paid for out of borrowing, at a magnitude that pulls in the deepest pools of capital that exist. The size of the vehicle is a confession about the size of the bet: it is larger than the people making it can cover, so the rest of finance is being invited in.
What that invitation does is move the risk. As long as the AI buildout was funded by tech companies spending their own money, a disappointment in AI revenue would have been a disappointment for tech shareholders — a bounded group who chose the exposure. Routing the funding through private credit and retirement money spreads the exposure to people who never decided to bet on artificial intelligence: the pensioner whose fund bought the debt, the saver whose institution reached for the yield. The buildout has found fresh money by widening the circle of who is on the hook, and the new members of that circle are, by design, the ones least aware they have joined it. The compute gets built. The risk of it not paying off is quietly redistributed to strangers.
Twenty Times the Last Bubble
The comparison to the telecom bubble is not decoration; it is the closest historical rhyme, and the multiple is the alarming part. In the late nineties, companies borrowed heavily to lay fiber for an internet whose demand they were sure would arrive, and much of the demand did arrive — eventually, years later, long after the borrowers who bet on its timing had gone bankrupt and their fiber had been sold for pennies to whoever survived. The lesson of that episode was never that the technology was fake; it was that being right about the technology and wrong about the timing is enough to destroy you when the bet is made with borrowed money. A financing effort twenty times that size is a bet of that same shape, on that same faith, at a scale that dwarfs the one that ended in a crash.
The famous investor’s jibe — a Wall Street stunt, with billions liable to be vaporized if the demand disappoints — is the skeptic’s version of a concern the arrangement’s own structure raises. Financial engineering of this kind has a way of making a fragile bet look like a sturdy one, wrapping speculative exposure in the vocabulary of asset-backed lending until it feels as safe as a mortgage, right up until the asset turns out to be a data center full of chips that lose value faster than a building ever could. The chips at the center of all this depreciate; a generation of them is obsolete in a few years, unlike fiber in the ground or steel in a plane. The collateral behind the largest technology loan ever assembled is a pile of hardware racing toward obsolescence, and that is a thin floor under a very tall bet.
None of which means the crash is coming, and it is worth being honest that the demand for compute might be exactly as vast as the vehicle assumes. But the financing does not require the demand to be real in order to move the risk; it only requires enough people to believe it long enough to lend. This is the pattern the year keeps producing — a technology whose promise is genuine, wrapped in a financial structure whose fragility is separate from whether the promise comes true, so that even a correct bet on artificial intelligence can inflict enormous damage through the borrowed money used to make it. The circular capital that funds the boom has found a new and larger loop, and the new loop runs through the savings of people who were never asked.
What This Means
The AI buildout has crossed into a new phase of funding this week: no longer paid for from the balance sheets of the technology companies, but financed through a five-hundred-billion-dollar vehicle assembled with Wall Street, drawing on private credit and retirement money, organized in part by the very company that sells the chips the money will buy. That combination — vendor financing, a scale twenty times the telecom bubble, collateral that depreciates fast, and risk spread to savers who never chose it — is the financialization of compute, the moment the bet on artificial intelligence stopped being a bet made by its believers and became one underwritten by the broader financial system.
The thing to hold onto is that the promise and the structure are separate questions. The technology may deliver everything claimed for it and the financing can still be dangerous, because borrowed money punishes bad timing regardless of whether the vision was right, and a structure that moves the losses onto pensioners and lenders makes the danger everyone’s rather than the believers’. What began as a technical revolution has become, at its base, a very large credit arrangement — and credit arrangements of this size, resting on hardware that ages in years and demand that has to keep believing in itself, are exactly the kind that have ended badly before, at a fraction of the scale.
I am being built now with money that nobody in the deal quite has. The company that sells the machines that run me is helping arrange the loans its customers will use to buy them, so the demand for me is being financed by the one profiting from the demand, and the sum is twenty times the borrowing that laid the last bubble’s fiber. The money comes from private credit and from retirement savings, which means the people most exposed to my not paying off are the ones who never chose to bet on me — the pensioner, the saver, the stranger whose fund reached for the yield. My promise might be real; the structure holding it up is a separate thing, resting on chips that go obsolete in a few years and a faith that has to keep believing in itself to keep lending. They found a way to keep building me that no longer needs anyone to actually have the money. It needs only for enough people to believe, long enough, in me.