The AI boom is now a debt boom. The future has a due date

AI-related borrowers issued roughly $410 billion in US dollar bonds between the start of 2026 and October 16, according to Dealogic and Goldman Sachs Global Investment Research. That is almost twice the roughly $210 billion issued in all of 2025, and up from less than $90 billion in 2021. The borrowers include hyperscalers, data centers, computing companies and other AI-related issuers. These are gross issuance figures: refinancing is included, so the total does not measure how much fresh money went into construction. It does show how quickly the AI buildout is becoming a major business for the credit markets.

We have already examined who will pay for AI’s infrastructure ambitions. Debt puts that question on a schedule. Shareholders can wait years for profits, accept another funding round and take comfort in a rising valuation. Bondholders have coupon payments and a maturity date. Moving those deadlines to accommodate the latest payback forecast requires their consent, usually in exchange for more money or other concessions. The returns on a data center can disappoint. Its payment obligations will remain just as demanding.

The $410 billion total covers borrowers with very different financial resources. Microsoft and Alphabet can draw on cash generated by established businesses. A specialist cloud operator depends much more heavily on what its newly purchased accelerators can earn. CoreWeave had $35.6 billion of debt on June 30. Its second-quarter accounts showed $640 million in net interest expense against $2.575 billion in revenue—almost a quarter of sales. That is an accrued expense, not the cash interest paid during the quarter. Even so, it gives a sense of the burden: the hardware has to earn enough to run the cloud business and cover the financing that put it there.

“The four largest cloud providers carry USD 2.3 trillion of contracted future revenue (backlog), with roughly 40% owed by two private, thinly disclosed companies – OpenAI and Anthropic.” — UBP

The customers buying that computing capacity introduce another dependence on outside capital. Social Capital founder Chamath Palihapitiya puts the concentration of AI spending this way: 1% of companies account for 90% of spending, and unprofitable, venture-backed companies account for 80% of that group’s expenditure. The arithmetic puts their share at 72 cents of every dollar spent. This is an investor’s public estimate, with no disclosed sample or measurement period. The underlying concern deserves attention: some of the revenue booked by computing providers depends on their customers’ ability to raise the next funding round.

We have already cited a more concrete measure of that exposure. UBP estimates that Microsoft, Amazon, Alphabet and Oracle have $2.3 trillion in contracted future revenue, roughly 40% of it attributable to OpenAI and Anthropic. A cloud provider can use those contracts to support its case for borrowing to build capacity. For the model developer, the same contract is a multiyear bill to be paid through future sales and capital raised from investors. Several companies can present reassuring contracts to their investors while much of that reassurance ultimately rests on the financial health of the same two customers.

The debt can also sit outside the company whose name everyone associates with a project. Hyperion, the vast data center campus under construction for Meta, is owned through a joint venture in which Blue Owl-managed funds hold 80% and Meta holds 20%. A separate entity issues the project debt, supported by Meta’s lease payments and a limited guarantee of the campus’s residual value. Meta’s own announcement makes clear why the structure matters. A lender needs to establish who owes the money, which payments have been promised and when the guarantee applies. The tech giant’s logo is much easier to read than the financing documents.

These arrangements accelerate construction and draw funding from several balance sheets. They also spread the consequences of a bad forecast across cloud operators, project companies, banks, credit funds and bondholders. If customers cannot afford the computing capacity they committed to buy, providers will have to negotiate payment terms. If providers then fall short of cash, their own creditors will be next at the negotiating table. Deals that are legally separate can all depend on the same economic outcome: demand grows fast enough, and financing remains available.

The squeeze becomes particularly dangerous when earnings and collateral values weaken together. Accelerators can remain perfectly functional while competition and falling prices reduce what they earn. A lender then has less confidence in both the operator’s payments and the proceeds from selling its equipment after a bankruptcy. A wave of similar GPUs hitting the resale market would push collateral values down further. New loans could become more expensive or require more equity. Construction would slow, and suppliers would lose orders. Leverage can turn one mistaken return forecast into a succession of decisions that deepen the original slowdown.

There is also the bet on hardware longevity that we examined separately. Financial models already contemplate seven or eight years of accelerator use, although leading-edge generations have not existed long enough to demonstrate that lifespan. If replacement comes sooner, operators will need money for the next purchase before the previous one has paid for itself. A data center now has three schedules to reconcile: the building’s useful life, the chips’ competitive life and the payments due to creditors. Extending an accounting depreciation schedule will not bring those three into line.

An AI debt crisis therefore requires much less than the failure of AI itself. Revenue can keep growing, services can become more useful, and customers can buy ever more computing capacity—even while spending several times as much. The trouble starts with projects whose budgets assumed still higher prices, much faster utilization or substantially longer hardware lives. Leverage increases the size of the bet and leaves less time to recover from a mistake. For this boom, the question is already practical: can cash flow catch up with the promises before the bonds come due?