The Spot-Price Delusion

America's data-centre boom is being built on the shortest of foundations 

========

Every era of infrastructure mania begins with a plausible story and ends with a balance sheet. The railroad barons of the 1860s believed, correctly, that the railway would remake the American economy. The telecoms financiers of the late 1990s believed, correctly, that the internet would remake commerce and communication. Both were right about the technology and catastrophically wrong about the financing. Investors in artificial intelligence are at risk of repeating the error a third time, at a scale that dwarfs the previous two.

Nobody serious doubts that large language models and the data centres that run them are a genuine technological advance, nor that they may eventually justify a meaningful share of the trillions of dollars now being poured into them. That is not the argument. The argument is that the industry has constructed a financing structure so mismatched to the economics of its own product that it would be surprising if it did not end badly for a great many of the companies and investors involved.

Start with the mismatch of duration. A data centre is a twenty-year asset: poured concrete, cooling infrastructure, transformers, land. The demand it is being built to serve, by contrast, is priced and sold in a spot market for computing capacity that can swing wildly from quarter to quarter as models change, customers churn and rival capacity comes online. Financing a two-decade asset off an assumption about this month's rental rate for a graphics chip is not conservative capital allocation; it is a bet that today's boom prices are the new normal. History disagrees. Nineteenth-century railroads built three lines where one would do, on the strength of freight rates that could not survive the completion of the very track being financed. Telecoms firms laid fibre by the millions of miles on the assumption that bandwidth demand would keep pace, then watched rates collapse by more than 90% within a few years as everyone's capacity came online at once. Data centres, built on shorter technology cycles than either railroads or fibre, are arguably more exposed to this mismatch, not less: the hardware inside them can become uncompetitive within three or four years, long before the underlying real estate and power contracts have been paid down.

Then there is the accounting. A striking number of the companies at the centre of the AI build-out are capitalising their spending as “construction in progress,” a treatment that lets enormous sums of capital expenditure sit on the balance sheet without running through the profit-and-loss statement as depreciation. This is not fraudulent — it is standard practice for assets not yet in service — but it has the convenient effect of flattering reported earnings for exactly as long as the boom continues, while masking how much of the equipment being installed is losing economic value the moment it is switched on. Strip this treatment away and set depreciation against the pace at which chips actually become obsolete, and the profitability of the industry's most celebrated capital spenders would look considerably less impressive.

The returns data, where it can be observed, already point the wrong way. Estimates of incremental returns on invested capital for the largest cloud and infrastructure operators have fallen from roughly 40% a year and a half ago to around 20% today. A trend line, extended, does not need to be dramatic to be alarming: a continued slide toward 10% would make the current pace of spending arithmetically unsustainable, since it implies the industry is spending ever more capital to generate ever less incremental profit. Markets have a habit of ignoring declining-returns data for a long time, right up until they do not.

More troubling still is the circularity that has crept into how this build-out is financed. Chipmakers have taken equity stakes in some of the very companies buying their chips, who in turn use those chips — and the credibility of the chipmaker's backing — to raise the debt and equity needed to keep the whole enterprise afloat. This is precisely the kind of vendor financing that characterised the telecom bubble a generation ago, when equipment suppliers lent money to customers so the customers could buy the equipment, and both sides booked the resulting revenue as if it had come from an arm's-length transaction. It works beautifully until financing conditions tighten, at which point the losses have a habit of surfacing simultaneously at both ends of the chain.

None of this means artificial intelligence is a mirage. It plainly is not, and the productivity gains it is already delivering in some industries are real. But there is an important difference between “this technology has a future” and “this specific set of capital structures, leverage ratios and depreciation assumptions is sound.” The dot-com bust destroyed a great deal of paper wealth built on companies with no revenue and, in retrospect, obviously silly business plans. It was, in the main, an equity bubble: painful, but survivable structurally because losses were concentrated in stock prices rather than in over-leveraged, physically depreciating assets. This build-out is different in kind. It combines dot-com-era enthusiasm with railroad-era physical overbuilding, telecom-era vendor financing, and a level of debt issuance that none of its predecessors carried at a comparable stage. Should demand for compute merely grow more slowly than expected — not collapse, simply disappoint — the assets financed on today's spot prices will not gently depreciate. They will be revealed, all at once, as having been the wrong size, in the wrong place, financed at the wrong price, for a market that no longer wants them at the terms that were promised.

Investors currently pricing this industry as though its most optimistic projections were fact would do well to remember that promises are not a line item. Eventually, someone has to reconcile the balance sheet.