The argument that artificial intelligence is overvalued has been made many times and in several languages, and it is largely the wrong argument. Equity overvaluation is a painful event for shareholders and a survivable one for the system. Shares fall, paper wealth evaporates, portfolios rebalance, and the economy absorbs it. The case now being advanced is different and considerably more serious: the risk is not in the valuations, it is in the debt.
The distinction matters because of who holds the exposure. When a company funds expansion by issuing equity, the people who bought the shares carry the downside, and they knew they were carrying it. When it funds expansion with borrowing, the exposure sits with banks, with private credit funds, with insurers reaching for yield, and with the structured vehicles that repackage all of it and sell the pieces to investors who have not read the underlying documentation. Equity losses are contained by design. Debt losses propagate by design.
The capital intensity of the current build-out is the part that distinguishes this cycle from the last technology boom. The dot-com era burnt enormous sums on marketing and headcount, which are expenses rather than assets. This era is spending on data centres, power infrastructure, cooling, land and silicon, which are assets, and assets can be pledged. A chip fleet or a campus with a grid connection is collateralisable in a way that a banner advertising campaign never was, and collateral is the fuel on which leverage runs. The ability to borrow against the build-out is precisely what allows the build-out to exceed what cash flow would support.
Here is the espresso problem. A café that buys a second machine out of retained profit has made a bet with its own money, and if the footfall does not arrive it has an idle machine and a bruised ego. A café that finances three machines, a roastery and a second site against the projected revenue of the second site has made a different kind of bet, and the bank is now in the business with it. The coffee may be excellent. The equipment may be the best available. If the queue is shorter than the model assumed, the schedule of repayments does not adjust out of sympathy, and the machines get sold at auction to someone who paid nothing for the optimism.
The rate environment is what makes the current structure fragile rather than merely aggressive. A great deal of this borrowing was conceived when money was close to free and refinancing was a formality. In a world where rates are higher and staying there, every refinancing becomes a genuine negotiation conducted at a worse price, and the interest coverage ratios that looked comfortable at origination tighten each time a maturity comes round. Depreciation schedules make it harsher: accelerator hardware has an economic life measured in a handful of years, so the asset backing the loan is losing value on a clock while the principal is not.
The insight is that systemic risk is a question about plumbing rather than sentiment. The relevant questions are who lent, against what, with how much recourse, and how many of those loans have been sliced and distributed into vehicles whose holders believe they own something diversified. If the answer is that a concentrated group of borrowers has pledged rapidly depreciating hardware to a lightly supervised private credit sector, which has sold the exposure onward, then a slowdown in revenue growth becomes a credit event rather than a disappointing quarter.
None of this requires the technology to fail. It requires only that revenue arrive later than the repayment schedule assumed, which is the most ordinary outcome in the history of capital investment. The machines will keep working. The question is who owns them when the first repayment is missed, and how many institutions discover simultaneously that they were closer to the trade than they thought.
Source: Robin J Brooks, “Financial Stability Risk as AI Debt Rises”
