AI risk moves to credit: the boom is now financed with debt
Until recently the largest technology companies funded AI from cash and earnings. That model has changed. Investment is growing faster than the revenue it generates, and lenders are starting to ask whether it will pay back.
A short assessment of a material new development.
ByDamián Falcone· Signal & Range
- As of
- Analytical horizon
- 6–18 months
- Region
- United States · global technology and credit markets
Approximately 3 minutes.
What is observable
Until recently, the largest technology companies funded AI from cash and earnings. That model has changed. On FactSet's July figures, aggregate capital expenditure by the five largest US cloud providers is expected to exceed US$690 billion across their 2026 fiscal years, growth of more than eighty per cent on the year. Free cash flow is expected to be close to zero or negative at every one of them except Alphabet and Microsoft.
To close the gap, the companies have gone to the bond market. Six AI-focused issuers — Amazon, Alphabet, Meta, Nvidia, Oracle and SpaceX — had sold about US$182 billion of investment-grade bonds by July, on Bloomberg data, against roughly US$13 billion in the same period a year earlier. Incremental debt has risen from nine per cent of capex in fiscal 2024 to thirty-two per cent on a trailing basis, and in July S&P cut Oracle to its lowest investment-grade rating, citing its AI spending and negative free cash flow.
Below the hyperscalers, the specialised AI cloud providers that build and lease capacity depend on high-yield debt and private credit. CoreWeave's unsecured notes, for instance, were placed at a 9.25 per cent coupon in May 2025, with earlier secured facilities reported at eleven to fifteen per cent. That is the cost of money for the layer of the stack that is growing fastest.
Why lenders are pricing this differently
Three structural causes explain the higher perceived risk, and rising interest rates amplify all three.
- Revenue lags investment. Capex is committed now; monetisation is still uncertain and arrives, if it arrives, on a lag of years. A lender is being asked to finance a build whose payback the borrower cannot yet demonstrate.
- Concentration. Many infrastructure providers depend on a handful of customers, and some of those customers are financed, directly or indirectly, by their own suppliers. A circular structure of that kind concentrates default risk rather than spreading it.
- Obsolescence. The chips that make up most of the collateral can lose value before the debt that financed them is repaid. Equipment with a short economic life and long-dated debt is a mismatch lenders have seen before.
The consequence is already visible
Firmus, an Nvidia-backed Australian data-centre operator, withdrew its planned US$5 billion ASX listing on 9 October, two weeks before it was due to price, after investors questioned a valuation built on future capacity. The company had two data centres in operation, with some 46 megawatts live against roughly 865 megawatts in development.
It will now seek private capital, and the terms are likely to be more demanding than the ones it has just walked away from. The episode is not a judgement on one company. It is the first clear instance of the public market declining to finance capacity on the strength of a forward contract book, and that is the test every leveraged provider in the sector will now face.
How it reaches your business
Most companies hold no AI debt of their own. The exposure arrives through suppliers, through pricing and through decisions already taken, in three places.
- Cost. Higher financing costs for the providers of compute and cloud capacity can pass through to the prices their customers pay. A contract indexed to list prices carries this exposure whether or not it names it.
- Continuity. A leveraged AI supplier embedded in a critical process is a single point of failure with a balance-sheet risk attached. Vendor due diligence that stops at service levels has not looked at the right page.
- Capex. AI business cases approved under easier funding conditions may no longer clear the hurdle rate once the cost of the underlying capacity reprices. The question is not whether the project still works; it is whether it still works at the new cost of money.