The artificial intelligence build-out began as a story about technology. It became a story about capital spending. It has now become a story about credit, and that is a different kind of story with a different kind of failure mode.
Global debt issuance tied to AI infrastructure is on track to approach $570 billion this year, according to estimates circulating among the large investment banks. Roughly $236 billion of that had already been raised by the end of May. Private bond issuance alone reached about $81 billion through the first five months, the strongest start in data going back to 2016.
These are not the financing patterns of companies funding growth out of operating cash flow. They are the financing patterns of an industry that has outrun its own balance sheets.
The order books tell you what changed
The clearest evidence of a turn is not in issuance volumes, which remain enormous, but in demand for the paper being issued.
In February, orders for hyperscaler bond offerings covered nearly five times the amount on offer. By July, coverage had slipped below two times. That is not a market that has closed. It is a market that has stopped being a formality, and the difference between a five-times book and a sub-two-times book is the difference between an issuer setting the price and an issuer discovering it.
Hyperscalers issued roughly $121 billion of bonds during 2025, more than four times their five-year average. The absorption of that supply was straightforward while investors treated the credits as effectively risk-free proxies for the largest companies on earth. What has changed is that the paper is no longer only being issued by those companies.
The structures that keep the debt off the balance sheet
Much of the incremental financing is not appearing where a casual reader of an annual report would look for it.
A substantial share of data centre development is now funded through 144A bonds, securities sold only to qualified institutional buyers and exempt from registration with the Securities and Exchange Commission. For data centres specifically, this market barely existed a year ago. It is now one of the fastest-growing corners of corporate credit, and by construction it is less visible than the public bond market.
Alongside it sits private credit. Projections from large investment banks suggest private lenders could supply an additional $800 billion of data centre financing over the next two years alone. Much of this flows through special purpose vehicles, joint ventures and lease structures that sit adjacent to the operating company rather than inside it.
The Bank for International Settlements has drawn attention to precisely this pattern, distinguishing between on-balance-sheet and off-balance-sheet borrowing in the AI infrastructure boom. The distinction matters because a leverage ratio calculated from a hyperscaler consolidated accounts may understate the total claim on the same underlying cash flows.
None of these structures is improper. Project finance has funded infrastructure for well over a century, and there are sound reasons to isolate the risk of a single facility from a parent company. The concern is aggregation: when many separate structures depend on demand from the same handful of tenants, the diversification is presentational rather than real.
Why credit investors are treating this as a systemic question
Surveys of institutional fund managers this year have produced a striking result. Close to half of respondents now identify hyperscaler capital expenditure as the most likely source of the next systemic credit event, ahead of commercial real estate, sovereign stress and leveraged lending.
The reasoning behind that answer runs roughly as follows. The assets being financed are long-lived in physical terms but potentially short-lived in economic terms. A data centre shell and its power connection may serve for decades. The accelerators inside it may be uncompetitive for frontier workloads in three or four years. Debt amortising over a longer horizon than the equipment it funded is a structure that depends on refinancing, and refinancing depends on the market still believing in the demand story at the moment the maturity arrives.
Add the concentration problem. A large share of contracted data centre capacity is underwritten by a very small number of creditworthy tenants. If any one of them slows its commitments, the effect does not fall on one lender. It falls simultaneously across the 144A market, the private credit funds, the equipment leasing books and the utilities that built generation capacity against those contracts.
The power constraint that makes this harder
There is a second reason this cycle is unusual. Capital is not the binding constraint on how fast the build-out proceeds. Electricity is.
Grid interconnection queues in the major data centre markets run to years. New generation and transmission capacity takes longer to permit and build than the buildings it will serve. That mismatch has two financial consequences. It pushes developers toward sites with existing power availability, raising land and acquisition costs. And it creates a class of asset that is financed and constructed but cannot yet earn revenue, which is the least comfortable position on any project finance timeline.
Power constraints also explain part of the equity-market anxiety about component costs. When a developer cannot add capacity by adding buildings, the alternative is to add capability inside existing footprints, which means denser and more expensive systems.
What separates this from 2008, and what does not
The comparison to the pre-crisis credit boom is inevitable and only partly fair.
What is different is the quality of the underlying obligors. The largest AI tenants are among the most cash-generative companies in history, with net cash positions and investment-grade ratings. There is no equivalent of the subprime borrower here. The demand for the service being financed is real, growing and revenue-producing today, which was not true of a great deal of 2006 mortgage origination.
What is similar is the machinery. Rapid growth in a lightly observed segment of the credit market. Assets moved into structures that reduce apparent leverage. Concentration disguised by the number of separate transactions. And a valuation model that works only if a demand assumption holds across the full life of the financing.
The historical rhyme that fits better is the late-1990s telecommunications build-out. The fibre laid in that period was eventually used, and used profitably. But the debt raised to lay it defaulted extensively, because the capital structures assumed revenue would arrive on the same schedule as capacity. The infrastructure survived. Its first set of financiers did not.
What to watch from here
- Order book coverage on new hyperscaler and data centre issues. A move back toward the February levels would signal renewed confidence; further deterioration would signal the opposite more reliably than any spread level.
- Spread differentiation between parent-level debt and project or SPV-level paper. If investors begin pricing those tiers meaningfully apart, the market has started distinguishing tenant credit from asset credit.
- Lease and offtake disclosure in the coming earnings reports. Contracted capacity commitments are the collateral for a large share of this borrowing.
- Private credit fund marks and redemption terms. Illiquid lending against assets with uncertain economic lives is the segment least likely to reprice smoothly.
- Any change in assumed useful lives for AI hardware. That single accounting assumption connects the equity story to the credit story.
Outlook
The AI infrastructure build-out is unlikely to stop. Demand is genuine, the tenants are solvent, and the competitive cost of underinvesting is high enough that no major operator can unilaterally pull back.
But the financing conditions that made the first phase easy are visibly tightening, and the shift from equity-funded to debt-funded expansion changes what a slowdown would look like. An equity-funded disappointment costs shareholders a multiple. A debt-funded disappointment transmits through creditors, and creditors are connected to everything else.
Investors watching this week for signs of strain should look past the headline capital expenditure figures, which our analysis of the earnings-week spending question covers in detail, and toward the disclosures on how the spending is funded. The size of the build is now well understood. Who is holding the risk is not.
About the data: Issuance volumes and forward projections in this article are published investment bank estimates rather than official statistics, and should be read as approximations. Discussion of on-balance-sheet and off-balance-sheet borrowing draws on Bank for International Settlements research into financing of the AI infrastructure boom. Order book coverage ratios describe conditions in the primary bond market.
Reader note
This article is general information, not personalized financial advice. Read our Financial Disclaimer.