What if the real danger in artificial intelligence is not that it thinks too much, but that it borrows too well? Markets often mistake momentum for strength, yet history says leverage is the crack that turns a story into a system problem. The European Central Bank has now placed AI finance in that older and less flattering tradition, warning that the rapid shift toward debt finance has historical parallels with the dotcom and subprime bubbles and could spill over into euro area financial stability.
The warning matters because the boom is no longer just about chips, software, and cloud capacity. It is increasingly about promises layered on promises, with obligations extending far beyond balance sheets. The ECB estimates big tech’s total off-balance-sheet commitments at more than USD 4 trillion as of Q2 2026, well above their on-balance-sheet liabilities. That is not a normal amount of hidden stretch. It is a cathedral of contingent claims built in the fog, where every new beam looks sturdy until the weather turns.
The ECB’s November 2026 Financial Stability Review argues that the AI buildout is drifting toward a more debt-heavy structure. Its authors, Anne Duquerroy, Oana Furtuna and Lia Vaz Cruz, put the point bluntly: “The AI boom is not just a technology story, but also a financial stability story.” That is the sort of sentence bankers dislike because it strips away the romance. A technology narrative sounds like invention. A financial stability narrative sounds like margin, maturity, and the old lesson that every exuberant cycle eventually meets accounting.
The central bank says cumulative capex by US big tech is forecast to reach about USD 5.5 trillion by 2030, with free cash flow turning negative for most hyperscalers in 2026. That combination should unsettle anyone who has ever watched a seemingly self-funding expansion become a race against time. When cash flow turns negative, the machine no longer feeds itself. It must be refueled from the outside, and outside funding always comes with a price: covenants, refinancing risk, or simply the market’s changing mood.
There is also the matter of time, the most ignored variable in finance. The ECB says hyperscaler debt maturity exceeds the economic life of the physical capital it funds. Market intelligence cited by the bank puts AI equipment life at 1–3 years, while firms raised the accounting life of servers from 3 to 6 years. That gap is not a footnote. It is the classic mismatch that has broken many elegant structures before. Debt wants certainty. Hardware decays. The bond matures later than the machine that justifies it.
The more worrying part is the circularity. The ECB notes mutual commitments between chipmakers and customers that inflate the expected revenues of both. It also points to Alphabet and Amazon “other income” coming from revaluations of Anthropic stakes. That is how bubbles acquire a polished surface: each participant cites the commitment of the other as evidence that demand is real. The market then mistakes mutual reinforcement for independent validation, much as a hall of mirrors can seem like depth.
Game theory helps here. In a boom, each player sees the same incentive: keep building, keep buying, keep funding, because the cost of stopping first is high. No one wants to be the one who doubts the future just as the group is betting on it. This is how coordination traps form. The system can look rational in the short run and fragile in aggregate. The winner is often simply the last participant still willing to extend the ladder after the roofline has already come into view.
erscalers issued over $200bn in bonds in the past year and now compete with banks as the largest single issuers in dollar investment-grade corporate bonds. Separately, the Bank for International Settlements reported bond issuance from the five largest hyperscalers reached US$222bn this year. That is a scale that changes the market’s plumbing. A technology cycle that once lived mostly in equity valuations is now entering the bond market’s bloodstream, where failure is slower but more contagious.
There is a clue in the pricing itself. The ECB’s chart data show hyperscaler credit spreads widened relative to similarly rated issuers, especially at longer maturities, from January 2024 through November 2026. That does not prove panic. It does suggest investors are asking for more compensation to lend over time, which is exactly what they should do when a business model depends on long-dated assumptions and fast-moving hardware. The bond market is not sentimental. It does not care about conference-stage optimism. It asks only whether the borrower will still be standing when the bill arrives.
This is where market psychology usually fails. People like stories of compounding, scale, and inevitability. They do not like stories about decay, replacement, and refinancing. Yet the second set is often the more relevant one. In nature, fast-growing species exhaust their soil. In engineering, a structure’s weakest joint decides the load it can bear. In finance, the weakest link is often not revenue but duration. A company can be profitable on paper and still be trapped if the cost of capital rises or the maturity wall arrives too soon.
The ECB does offer one moderating note: “So far, however, hyperscalers’ earnings have broadly kept up with expectations.” That is important because it prevents an over-simple alarmist reading. This is not a verdict that the boom has failed. It is a warning that success itself can become the source of fragility when it is financed with large, layered, and increasingly opaque obligations. The point is not that AI is worthless. The point is that good outcomes can still be precarious when balance sheets are stretched to accommodate them.
It would be lazy to say AI is simply another subprime bubble. The assets are different, the borrowers are different, and the macro backdrop is different. But the structural rhyme is hard to miss. Before subprime broke, the system believed the collateral would keep rising and the distribution of losses would remain tame. Before the dotcom collapse, capital chased the future faster than the future could justify itself. The ECB is warning that AI finance now carries echoes of both: exuberant capex, questionable duration, and a reliance on future cash flows that are not yet as solid as the market pretends.
The deeper problem is not that investors are irrational. It is that they are selectively rational. They know there is risk, but they assume someone else will absorb it. They know the capital cycle can overshoot, but they assume scale protects the largest players. They know debt amplifies outcomes, but they focus on the upside and treat the downside as a tail event. That mindset works until the tail becomes the center. Probability is cruel in this way: a low-probability event, repeated across a vast network of obligations, is no longer low-probability in its effects.
The immediate documents to watch are the ECB’s full Financial Stability Review and its accompanying euro-area bond-market box by Duquerroy, Furtuna and Vaz Cruz. But the larger lesson does not depend on one report. Every great capital cycle invites the same human mistake: believing that because a thing is productive, it must also be safe. It need not. A machine can be useful and overleveraged at the same time. A revolution can be real and still financed on brittle terms.
That is the uncomfortable inversion in the AI debt story. The question is no longer whether the technology works. The question is whether the financing can survive its own enthusiasm. Markets often discover too late that the most dangerous part of a boom is not the dream, but the spreadsheet that makes the dream look self-sustaining.