Wall Street has spent two years arguing about whether the spending on artificial intelligence is a bubble. Dario Perkins skipped that argument and wrote down a date.
In a research note published Oct. 22, the chief macro strategist at TS Lombard said the capital expenditure cycle behind the AI boom will most likely end in early 2028. When it does, he expects the S&P 500 to fall 30% and technology stocks to fall 40%, erasing roughly $20 trillion in market value.
The mechanism he describes is circular, and that is the point. A large share of the revenue growth now reported by AI companies comes from other companies spending on AI. One firm buys compute, its supplier buys more compute, and both book the transaction as growth. The circle looks like demand until someone stops paying.
Perkins puts a number on the concentration. Roughly 80% of revenue at OpenAI and Anthropic comes from about 1% of their customers, he wrote. That is a customer base narrow enough that a handful of decisions can move a revenue line, and it sits upstream of a supply chain that has been built to assume the line keeps rising.
The failure, in his account, starts with the weakest links. Suppliers and developers without balance sheets to absorb a pause begin missing payments. Credit losses run between $125 billion and $190 billion. The drag knocks about two percentage points off U.S. gross domestic product. Equity markets do the rest.
That is not 2008, and Perkins is explicit about it. Households and non-financial companies carry healthier balance sheets than they did eighteen years ago, and the Federal Reserve has room to cut rates if the labor market cracks. A capex bust that stays inside the corporate sector is painful for investors and survivable for the economy.
The place to watch first is not the equity market. Data-center developers have financed a large share of their construction with leased debt and private credit, and the lenders in that market have little experience with an asset class whose collateral loses value as fast as a server rack. A pause in orders would show up there as covenants tripping and refinancings failing, well before it reaches a quarterly earnings call.
Utilities and power producers are the second channel. Several have raised capital-spending plans on the strength of contracts with data-center tenants, and those contracts are typically signed years before the electricity is needed. A developer that walks away leaves a generator with a substation, a transmission connection and no buyer for the output.
The forecast lands in the middle of a live disagreement. Bulls point to order books at the chipmakers, to long-term power contracts signed by data-center developers, and to the fact that the largest buyers are generating cash rather than borrowing against projections. Skeptics point to depreciation schedules that assume GPUs stay useful for five or six years, to lease structures that keep obligations off balance sheets, and to the gap between announced spending and revenue actually collected.
History offers an uncomfortable template. Railway manias, fiber-optic build-outs and the telecom equipment boom all ended the same way: too much capacity, a collapse in the price of the thing being sold, and a wave of defaults among the companies that borrowed to build it. In each case the technology survived and most of the builders did not.
The AI version has a feature the earlier ones lacked. The buyers are among the most profitable companies in the world, and they are funding the build-out from operating cash flow rather than junk bonds. That shifts the risk from lenders to shareholders, which changes the shape of the bust even if it does not prevent one.
Perkins is also careful about what the bust would not do. The companies at the center of the build-out earn money from search, advertising, cloud computing and software subscriptions, and those businesses would keep generating cash even if the expansion stopped. Their share prices would fall. Their ability to fund the next generation of models would not disappear.
What would break the thesis is evidence that AI revenue is spreading beyond the top customers. Enterprise software companies reporting that AI features are being bought by mid-sized firms, cloud providers disclosing broader adoption, and pricing that holds as capacity expands would all point to a wider base. So far the disclosures are thin.
Perkins’s own record argues for caution about the timing. Macro strategists have called the top of this trade repeatedly since 2023, and each time the spending accelerated. A forecast that names a quarter is a forecast that can be checked, which is more than most of the bearish commentary offers.
The practical question for investors is what to do between now and then. The note does not recommend shorting the leaders. It suggests watching the credit markets, where a default wave would appear before it shows up in earnings, and the power contracts, where a cancelation would signal that a developer has lost confidence in demand.
Two years is a long time in a market that has moved this fast. It is also short enough that the spending plans now being signed will still be running when the forecast comes due.
