For two years the anxiety about artificial intelligence was whether there were enough chips. As that shortage eased, it became whether there was enough power. Both are the wrong question. The constraint that decides how far and how fast this technology gets built is neither the silicon nor the electrons. It is the balance sheet — and whether it can carry an asset that may be obsolete in three years against debt that runs for fifteen.
Start with the number that frames everything. Over fifteen months, OpenAI has signed compute and data-centre commitments worth around $1.4 trillion — more than the annual output of Spain — against revenue of roughly $20 billion and a loss of about $10 billion. It cannot remotely fund that from its own cash; it is, in effect, building on other people’s balance sheets. By early this year it had quietly trimmed the ambition and begun renting capacity it had promised to own. Hold that gap between what is promised and what is earned. The whole argument lives inside it.
The substation, not the wafer
First, dispose of the chips. They were the bottleneck of 2023; they are not the bottleneck now. The industry’s own spending says so — more than sixty per cent of hyperscaler capital now goes not to processors but to the buildings, the power and the grid connections around them. The wafer has given way to the substation.
And the substation is hard — though it helps to see that “energy” is really three constraints wearing one coat. The first is generation: can you make the electrons? Here the response has been a scramble back to firm power, including a nuclear revival unthinkable a decade ago. Microsoft has contracted the entire output of a restarted reactor at Three Mile Island — the site of America’s worst nuclear accident, now tactfully rebranded — on a twenty-year deal; Amazon, Google and Meta have signed for conventional plants and a coming generation of small modular reactors. Where firms turn to gas instead, they meet a wall: the efficient turbines are sold out, with GE Vernova and its rivals quoting five to seven years.
The second constraint is the connection, and it is the one the headlines miss. You can have the power and the site and still wait. The queue to join the American grid has swollen to some 2,600 gigawatts of projects — twice the capacity that currently supplies the entire country — with average waits of about five years. The equipment is worse: the large transformers that step power down for a data centre, which cost less than a tenth of a project but hold up all of it, now carry lead times of roughly five years, against thirty months before the boom. Last December, for the first time in its history, the grid operator for thirteen eastern states failed to buy enough capacity to meet its own reliability target. The third constraint is simply time — the mismatch between an eighteen-month product cycle and a build that takes half a decade.
The point about all of this is that it is an engineering-and-time problem, not a permanent wall. Given enough money, the turbines get built, the queues clear, the transformers ship. Which returns us to the money — and to whether it is being spent or merely committed.
Whose balance sheet?
The sums have broken through into the without-precedent. The four largest US technology firms plan to spend about $725 billion on capital projects this year, up roughly three-quarters on last, the overwhelming majority on AI infrastructure; Goldman Sachs puts their combined outlay to 2030 above five trillion dollars. Capital intensity — the share of revenue ploughed straight back into plant — now runs at levels no software business has ever sustained. Operating cash flow, vast as it is, no longer covers it, and so the industry has turned to debt: something over a hundred billion dollars raised last year, and, on some estimates, one and a half trillion still to come.
Where it becomes a story for a markets reader is in how that debt is structured — and where the risk finally comes to rest. Much of it does not sit on the technology firms’ books at all. Meta funded its Hyperion campus in Louisiana through a $30 billion special-purpose vehicle, arranged with the private-credit manager Blue Owl and funded by Pimco, BlackRock and Apollo, that keeps the borrowing off Meta’s own balance sheet — while Meta quietly guarantees the assets’ residual value, which is to say it keeps the downside without showing the debt. By the Financial Times’s count, more than $120 billion of data-centre borrowing has already been moved into such vehicles by Meta, Oracle, xAI and others. The money increasingly comes from private credit — a $1.7 trillion industry that has lent the technology sector some $450 billion — and, beneath that, from the pension and insurance portfolios that buy the private-credit funds. That is the quiet part: the risk of the AI build-out is being dispersed, through structures designed to be difficult to see, into the savings of people who have never heard of a gigawatt.
The circle and the clock
Two features of this financing ought to make a markets reader uneasy, because both have been seen before. The first is circularity. When Nvidia agreed last September to invest up to $100 billion in OpenAI, which would spend much of it on Nvidia chips, the chipmaker was in effect financing its own sales; one analyst noted drily that the money would travel from Nvidia to OpenAI and straight back to Nvidia. Nvidia has taken stakes across a web of its own customers, from OpenAI to the cloud provider CoreWeave, and its backing lets them borrow more cheaply. Jensen Huang calls the circularity charge ridiculous. But revenue a supplier has funded looks like demand until the funding stops — and the round-tripping between vendors and customers is precisely what flattered the late-1990s telecom boom before it collapsed.
The second is the clock: the mismatch between how long these assets are financed and how long they actually last. The investor Michael Burry, of Big Short fame, argues that the hyperscalers flatter their profits by depreciating chips over five or six years when the real economic life of a cutting-edge processor is closer to two or three — understating costs across the industry, on his estimate, by some $176 billion over three years. He calls it one of the more common frauds of the modern era, and likens the moment to Cisco at the peak of the dot-com bubble. The firms dispute it, and not without cause: Nvidia and CoreWeave insist older chips keep earning on lighter work for years. But the disagreement is itself the signal, and the companies are not even aligned — Microsoft and Meta have stretched their depreciation schedules while Amazon quietly shortened its. Set fifteen-year debt against a three-year asset and the question is no longer whether the money can be raised. It is whether the returns will arrive fast enough to service it.
Where they land
All of it resolves on a map. And as another piece in this series argued of where a firm places its own people, a map of premises is never accidental — it records where the binding constraints truly lie. A data centre now goes wherever power, water, land and — increasingly — capital and political permission can be assembled fastest, and the geography is revealing. It goes to Texas, whose grid is besieged by connection requests four times the level of a year earlier, and where OpenAI’s flagship campus is rising at Abilene. It goes to the data-centre alley of northern Virginia, until the local grid strains and residents watch their power bills climb to pay for it. And it goes, tellingly, to the money: to the Gulf, where Abu Dhabi’s sovereign funds are both financier and host to vast OpenAI-linked capacity, and to the cheap, cool hydropower of northern Norway. When an industry chases jurisdictions for their spare electricity and their willing capital, it is telling you plainly what its binding constraints are.
The capital cycle
So: discipline or delusion? The honest answer is that this is a capital cycle — the deeper mechanics of which are the subject of a standing programme elsewhere on this site — and capital cycles are only ever legible afterwards. The bull case is strong and should not be waved away — demand for AI is real, the largest firms can fund their bets from operating cash, and, as one analyst tartly put it, much of the bear case is simply garbage. But the structure has every feature that has undone booms before: circular financing, aggressive accounting, a wave of debt raised against revenue that is still largely a forecast, and — most dangerous of all — four giants making the same enormous bet at once, so that if the revenues disappoint, everything re-rates together. An earlier piece in this series argued that a real revolution and a good investment are not the same thing. The lesson holds here, one layer down. The chips can be bought, and the power can, in time, be built. What cannot be manufactured to order is a return large enough, and prompt enough, to justify the debt now being raised against it. And if that return comes late, the people who feel it will not only be in Silicon Valley. They will be in the insurance and pension books whose money quietly underwrites the whole endeavour.