The AI Boom Is Built on Debt You Can't See

The AI boom is a story about profits, software, and superintelligence. The Bank for International Settlements — the central bank for central banks, the least excitable institution on the planet — is telling you it is actually a story about debt. The kind of debt you can't see, stretched across a supply chain of companies that have, in aggregate, borrowed far more than their earnings can carry.

It matters to you even if you've never bought a chip stock in your life, because the companies doing the borrowing sit at the heart of every index fund you own. US stocks are now about 64% of the MSCI global index, and households have roughly doubled their equity exposure relative to income since 2010. You are in this trade whether you meant to place it or not.

The plumbing behind the trillion

Here is the uncomfortable number at the center of the BIS's annual report, published in June: the five largest hyperscalers are on pace to spend more than a trillion dollars on AI capital expenditure across 2025 and 2026. That alone would be a story about bold and visionary capital allocation, the kind boards love. The problem is the next line in the same document — spending is outpacing earnings and free cash flow, so the build-out is being paid for with borrowed money.

That's where the plumbing gets interesting, because the borrowing is not a clean, visible corporate bond on a balance sheet you can read. Trace who owes what to whom and it becomes a loop. Chip makers and hyperscalers take equity stakes in AI labs and "neocloud" providers, who in turn commit to multi-year purchases of chips and computing power from the very people who funded them. Data-center construction is outsourced to third-party contractors, who then lease the facilities back to the tech company on long-dated contracts with exit clauses buried in them. The BIS's phrasing is a model of understatement: the terms are "poorly disclosed," with the same asset at risk of being pledged multiple times.

This is the thing to stop and feel. You cannot tell from the outside who actually owes what, because much of the borrowing sits off-balance-sheet or inside circular arrangements where the borrower and the customer are the same circle of people. The cash is real, but so is the leverage — you just can't see the ledger.

Where the hidden leverage lives

The banks mostly got out of this game after 2008, so the money came from the shadows. Direct-lending funds have quadrupled their lending to AI and IT over five years, to roughly 15% of their portfolios — an entire ecosystem now running well past a trillion dollars in private credit, thinly regulated compared with actual banks. The BIS flags that AI already accounts for about half of all investment-grade bond issuance and 87% of venture capital funding. The engineers and contractors building the centers — the EPC firms — are carrying the strain on comparatively weak balance sheets.

Now add the macro layer. The same report notes the Strait of Hormuz shut down after an Iran conflict in late February, cutting more than 10 million barrels a day of oil and sending prices up 67% to an intraday spike around $120, with a bump of about half a percentage point to global headline inflation. Here is the trap, and it's a classic: markets are assuming the energy shock is temporary and the AI boom continues, simultaneously. If inflation proves sticky and central banks have to keep rates high, or even raise them, they pop the AI-financed debt bubble from underneath while pretending to fight power prices.

Who is the forced seller?

Every credit unwind, and this is the durable lesson, has a moment where someone has no choice but to sell. In the AI edifice that actor is hard to name in advance because the structure is designed to hide them. It could be the private-credit fund facing retail redemptions and forced to liquidate. It could be the EPC contractor with thin equity who can't roll its construction debt. It could be the hyperscaler that, under its own borrowing covenants, has to stop buying chips — and thereby strangles the very AI labs whose revenue is the collateral.

The BIS doesn't predict a crash; it argues that the loop is unstable and that a reversal, once started, can unwind faster than a traditional banking crisis because so much of the credit sits outside the regulated perimeter. It reaches for history with intent: the canal mania of the 1830s, the British railway bubble of the 1840s, the dot-com collapse of 2000. Each began with a genuine technological breakthrough that attracted capital beyond any justifiable commercial return, and each ended with overinvestment reversing into recession.

The current BIS chief, Pablo Hernández de Cos, now says flatly that AI's rise is creating new financial stability risks and that spending on AI infrastructure has reached a scale big enough to move the whole global economy.

For someone figuring out what to do about it, the honest answer is that you don't need to forecast AI's future to act. The technology may be entirely real — the railway and the internet both were — and the trade can still be wrecked by the financing. The question is not whether AI will matter; it's whether the debt that funded the build-out can be serviced by the cash flows it is supposed to generate. Watch that plumbing rather than the stock prices. The BIS, the institution whose entire job is staring at balance sheets and finding the hidden leverage, is telling you it looked and couldn't find the bottom of the pile.