For most of the AI build-out, the money came from operations. The companies doing the spending were among the most cash-generative businesses ever assembled, and they were buying data centres out of the till.
That has changed, and the change is measurable. Capital spending at the major AI cloud providers is now estimated at around ninety-four percent of operating cash flow, for this year and next. In 2023 the figure was about forty percent.
When capex reaches ninety-four percent of the cash a business throws off, the business has run out of till. Everything beyond that — and the dividends, and the buybacks, and any margin for a bad quarter — has to be borrowed. S&P Global counts about $225bn of bonds issued this year by hyperscalers and related issuers including Nvidia, a record pace. A broader count of AI-related debt, attributed to Goldman Sachs strategist Amanda Lynam, puts the year's issuance nearer $489bn.
The maturity is the interesting part
Debt is not automatically alarming. A profitable company borrowing at a good rate against a productive asset is doing ordinary corporate finance.
What makes this worth examining is that the paper is long dated, and the justification for issuing long paper against an asset is that the asset lasts a long time. That is the standard argument, and for a building it is a good one. Data centres are concrete and steel and they will be standing in 2050.
The thing inside them will not be. The economically productive component of an AI data centre is the accelerator, and how long an accelerator remains productive is not a fact anyone has established. It is an estimate, and — as this desk reported when the depreciation bill first came into view — it is an estimate the companies make about themselves, one of them extending it from fifteen years to twenty-five with no physical change to anything.
Stretch the assumed life and two things happen at once. Reported earnings improve, because the annual depreciation charge falls. And the case for issuing thirty-year paper improves, because the asset it is matched against now nominally lasts thirty years.
What the lender is actually holding
Not a claim on a GPU. Bondholders in these issues hold unsecured claims on some of the strongest balance sheets in the world, and that is why the paper prices where it does.
But the credit case rests on the same forecast the equity case rests on: that the spending produces revenue on a timetable that services the debt. If it does, this is cheap money spent well. If the revenue arrives later than the coupon, the issuer is a very large company with a very large fixed obligation and an asset base whose accounting life it has already extended once.
That is a different risk from the one bond investors thought they were taking when they bought technology paper as a defensive holding. It is closer to project finance wearing an investment-grade rating.
And the price of long money went up
The timing is unfortunate in a way that has nothing to do with AI.
This desk wrote on Wednesday that every long bond repriced at once, in four currencies, for reasons that had nothing to do with any single issuer — the term premium rose, and the cost of borrowing for a long time went up everywhere for everyone. An issuer that decides to fund a multi-decade asset with multi-decade paper is making that decision into a market that has just repriced multi-decade paper.
Each year of extended useful life that justifies a longer bond now costs more than it did a year ago. The two facts are unrelated in origin and compounding in effect.
The number to watch
Not issuance volume, which measures appetite rather than risk.
Watch weighted average maturity of new AI-related issuance against the stated useful life in the same issuer's filings, and watch whether the gap between them narrows. If issuers start funding shorter than they depreciate, they are hedging their own estimate. If they keep funding longer, they are asserting it — and eventually a maturity wall arrives that is taller than it looks, in a year when everyone will finally know how long a 2026 accelerator actually lasted.
The figure of about $225bn of bonds issued this year by hyperscalers and related entities including Nvidia, and the characterisation of that as a record pace, is attributed to S&P Global. The broader estimate of about $489bn of AI-related debt issued this year is attributed to Goldman Sachs strategist Amanda Lynam. The estimate that capital spending equals roughly 94 percent of operating cash flow in 2026 and 2027 against about 40 percent in 2023, the projection that aggregate hyperscaler capex could exceed $770bn in 2026 at around 23 percent above prior expectations, and the observation that the issuance is notably long dated are from analyses reported by TrendForce DataTrack, Investing.com, CNBC and Fortune during 2026. The self-set and recently extended useful-life estimates referred to here are as described in this publication's earlier reporting and in the companies' own disclosures. The analysis is our own.
Topics marketsbondsaicapital spendingcredit
