The AI Boom Just Switched From Cash to Credit — and the BIS Noticed

Every AI infrastructure story fixates on the same numbers: hundreds of billions in capex, hundreds of thousands of GPUs. But the more interesting question was always the boring one. Where is the money coming from? In 2026, that answer changed — quietly, and without a press release. The Bank for International Settlements noticed.

The self-funded era is over

The early AI buildout was, structurally, pretty healthy. Microsoft, Google, Amazon, and Meta were printing cash from search ads, cloud contracts, and social feeds. Data centers got built out of free cash flow. It was the corporate equivalent of buying a house with money you’d saved from your paycheck.

That structure had a built-in circuit breaker. If AI underdelivered, the story ended with disappointed shareholders. The stock drops. Nobody defaults. The loss gets absorbed inside the company and stops there.

Then spending outran cash generation. Even a company minting money can’t cover hundreds of billions in annual capex from operating cash flow alone. So the funding model changed.

Bond markets are the new AI faucet

Big Tech showed up at the corporate bond window in force. These are companies that used to issue debt occasionally — to fund buybacks, to optimize taxes, to be clever about offshore cash. Now they issue bonds to pour concrete and buy silicon. That’s a categorically different thing.

And bonds are only the visible part. AI infrastructure financing has fragmented across several channels:

• Corporate bonds: the transparent option. At least it’s disclosed, and the market prices it.
• Private credit: money lent by funds rather than banks. Thin disclosure requirements. From the outside, you can’t see who’s holding what.
• SPVs: a separate legal entity created to hold one data center project, which then borrows against it. The debt doesn’t show up cleanly on the parent’s balance sheet.
• Vendor financing: the company selling the chips helps finance the company buying them. Revenue and lending get braided together.

This is the part that has the BIS uneasy. Bonds are fine. It’s the rest — nobody has the full picture of where the exposure actually sits. We heard sentences arranged like that before 2008.

Six-year assets, thirty-year debt

This sounds like an accounting footnote. It’s the whole problem.

The core asset in a data center is the GPU, and a GPU is not a building. Nvidia ships a new architecture roughly every 12 to 18 months, which means a three-year-old chip is already losing competitive ground. Companies depreciate servers over five to six years on the books, and a steady drumbeat of skeptics argues the real economic life is shorter than that.

The bonds financing those GPUs run 10 or 30 years. The asset ages out in six. The debt sticks around for three decades.

For a power plant or a port terminal, that math works — the thing spins for forty years and throws off cash the whole time. A GPU cluster is a different animal. Its returns depend on AI demand continuing to compound, and on that demand converting into actual invoices people actually pay. That part is still being proven.

Losses now have an exit route

Here’s the real consequence of debt financing: it changes who eats the loss.

Cash-funded failure hits shareholders. Shareholders signed up for risk; that’s the deal. Debt-funded failure hits bondholders. And who are bondholders? Pension funds. Insurers. Fixed-income funds. The target-date fund sitting inside your 401(k).

Add private credit and the chain gets longer. Private credit funds lever up with bank borrowing, which quietly puts banks in the AI data center business too. If demand forecasts miss, the shock no longer stops at the Nasdaq. There is now a transmission path into the financial system itself.

The BIS is the place where central banks compare notes. When an institution like that starts commenting on one industry’s capital structure, it isn’t saying tech valuations look rich. That’s the vocabulary of systemic risk.

So is it a bubble?

Honestly: unknown. And that’s the point worth sitting with.

If AI demand keeps compounding and the data centers earn their keep, this debt gets remembered as sensible leverage. Insisting on cash-only would look like a failure of nerve. Railroads were built on debt. So was the electrical grid. So was fiber.

Fiber has a sequel, though. In the early 2000s, telecoms borrowed enormously to lay glass, and demand showed up years later than the models promised. Several of them went bankrupt in the gap. The fiber survived — it later carried YouTube and Netflix — but the people who financed it never saw the upside.

The question isn’t whether the infrastructure turns out to be useful. It’s whether the repayment schedule lines up with the revenue schedule.

The takeaway

This isn’t a story about whether AI is a bubble. It’s about the AI boom migrating from an equity game to a debt game. Leverage magnifies returns when you’re right and multiplies casualties when you’re wrong.

Have you ever checked how much data center exposure is sitting in the bond sleeve of your retirement account? Almost nobody has. That’s precisely what the BIS is worried about.