The Bank for International Settlements, the institution that serves as central banker to the world’s central banks, warned Tuesday that AI infrastructure spending is being financed the way history’s worst technology manias were financed. A study published the same day says more than a trillion dollars of hyperscaler capital expenditure now rests on debt, equity swaps, and off-balance-sheet vehicles that obscure real leverage. Bloomberg reported the findings, sourced to the BIS study itself. The BIS did not predict a crash.
Microsoft, Amazon, Alphabet, Meta, and Oracle are on pace to spend more than $1 trillion combined on AI infrastructure across 2025 and 2026, according to the report cited by Bloomberg. That spending is concentrated in chips, data centers, and the power systems needed to run them. It already exceeds what these five companies collectively generate in earnings and free cash flow, which is why several have turned to debt markets to close the gap.
The report’s sharper warning concerns what it calls circular financing. Chip makers and hyperscalers are taking equity stakes in AI labs and neocloud providers, the specialized firms that rent out GPU capacity to other companies. Those same labs and neoclouds then sign multi-year contracts to buy chips or compute from the very companies that just invested in them. Revenue booked on one side of a deal is capital committed on the other, and the BIS says this loop makes the boom’s true scale difficult to measure from public filings alone.
Layered on top are special purpose vehicles, joint ventures, and private credit arrangements that keep debt off the parent company’s balance sheet. Coverage of the study has called this shadow borrowing (a label for financing structures that do not show up on a standard balance sheet). The BIS argues these vehicles understate real leverage across the sector, which matters if the underlying assets depreciate faster than their financing terms assume.
The BIS frames this pattern as a repeat of four earlier booms: the canal mania of the 1830s, the British railway mania of the 1840s, the electrification exuberance of the late 1920s, and the dotcom boom of the late 1990s. Each involved a genuine technological breakthrough. Each attracted capital that outran what commercial returns could ultimately support.
In the BIS model, competitive pressure keeps pushing capital expenditure higher even as the sector’s aggregate economic surplus, total payoff minus investment cost, shrinks. Under an adverse scenario, that surplus turns negative. Disappointment in AI’s productivity payoff could then trigger a sudden pullback in financing, turning the capex boom into a protracted investment bust with knock-on effects on credit conditions.
The report does not name which specific companies carry the riskiest structures, nor does it quantify how much of the trillion dollars runs through circular deals versus conventional corporate debt. That gap matters. A correction concentrated in a handful of over-levered neoclouds looks very different from one that spreads across five balance sheets deep enough to absorb write-downs.
For operators and founders, the practical move in the next ninety days is diligence on counterparty structure, not on model quality. Any team signing a multi-year compute contract should ask whether the vendor’s revenue depends on equity or purchase commitments from the same investor group, since a circular arrangement that unwinds hits committed compute first. Boards approving new data center leases or long-term GPU contracts should stress test unit economics against higher financing costs, because the BIS is describing a risk to how AI infrastructure gets paid for, not a verdict on whether the technology itself works.
Bloomberg reported on the Bank for International Settlements study on Tuesday, July 14, 2026.
