The AI boom now runs on borrowed money

The Nasdaq Composite fell 5.52 per cent on Thursday, to 24,315, its worst day since April 2025. The S&P 500 lost 4.36 per cent and the Dow more than 1,800 points. Chipmakers led the rout: Nvidia fell 9.3 per cent, a loss of almost $450bn in market value, AMD 17.3 per cent and Micron about 17.5 per cent. An exchange-traded fund of semiconductor shares dropped 12 per cent, its worst day since March 2020. The Nasdaq ended the day 11.9 per cent below its record close of October 6, which puts it in what traders call a correction.

The trigger was mundane. Qualcomm gave a weaker forecast than expected. But a disappointing outlook from one phone-chip maker does not normally take $3trn off the S&P 500, as Reuters estimated. Something larger changed in how investors price the build-out of artificial intelligence, and it is visible in the accounts of the companies doing the building.

From cash to credit

For most of this boom the biggest spenders paid for data centres out of their own profits. That made the question for investors a simple one: would the spending earn a decent return? It did not much matter what interest rates were, or who controlled Congress, because no one needed to borrow.

That is no longer true everywhere. Alphabet spent $51.3bn on capital projects in July to September, up from $44.9bn three months earlier, according to its results on October 28. Its free cash flow, the cash left after that spending, was minus $9.1bn, the second negative quarter in a row. To fill the gap it issued $28.7bn of senior unsecured debt and raised $12.4bn by selling shares into the market. It expects capital spending of up to $205bn this year and a significant increase in 2027. Meta, which reported on October 21, also had negative free cash flow, of $2.12bn.

Microsoft shows the other path. It spent $45.2bn in its latest quarter and expects more than $50bn in the next, yet still generated $18.6bn of free cash flow. Its backlog of signed commercial contracts rose 73 per cent to $677bn. It is building capacity that customers have already agreed to pay for.

The difference matters because a boom financed by borrowing behaves differently from one financed by profits. Once the spending depends on debt and new shares, it depends on the price of debt and shares. The ten-year Treasury yielded 4.95 per cent on November 3, close to levels last seen in 2007. The Federal Reserve raised rates in September and markets now put the chance of another rise in December at about three in four. Every point on the cost of capital is a point off the return on a data centre that will take years to fill.

Why the election mattered

Tuesday's vote added a second risk. Democrats won both houses of Congress, and in Texas, Gina Hinojosa, who campaigned for a pause on new data centres, won the governor's race. Data centres need land, power and water, and local opposition to all three has been growing. Investors began to ask whether permits and grid connections would arrive as quickly as the spending plans assume.

That explains the pattern of the losses. The biggest falls were not at the companies with the largest AI revenues but at their suppliers: makers of chips, memory and chip-making equipment, and GE Vernova, which sells power equipment to data centres and fell 7.5 per cent. A supplier's sales depend on the timing of its customers' capital spending. If a project slips by a year, the customer's revenue barely changes, but the supplier loses a year of orders. Intel showed how little good news helps: its data-centre revenue rose 78 per cent, yet its shares fell on a weak forecast.

Not a collapse in demand

None of this means AI demand is fading. Google Cloud's revenue rose 103 per cent to $30.71bn in the quarter. Microsoft's Azure grew 48 per cent. Customers are buying. The question the market has started to ask is narrower and more financial: whether revenue will arrive fast enough to service the debt that now funds part of the build-out, at interest rates that are higher than when the plans were drawn up.

That is a question about duration. A company that borrows today to build capacity that pays off in five years is making a long-dated bet, and long-dated bets lose value when rates rise or the payoff is delayed. Shares of chip suppliers had been priced as if the spending path were fixed. Thursday was the market admitting that it is not.

Prices recovered a little on Friday, when the Nasdaq rose 1.1 per cent. The next tests come quickly. CoreWeave, which rents out AI computing capacity, reports on Tuesday. Cisco and Applied Materials report on Thursday. Investors should watch less what these companies say about AI demand and more what they say about free cash flow, financing and delivery dates. The boom's biggest risk is no longer that nobody wants the product. It is that the bill for building it arrives before the revenue does.