S&P Global says the big cloud companies carry more AI-linked debt than their balance sheets show

The top hyperscalers likely have more AI-linked debt than is now known, S&P Global warned on Monday^1,2. The agency’s Habib said there was not an estimate of how much additional debt the circular financing might add to the firms’ balance sheets, although he said S&P was examining it¹. Debt issued by AI infrastructure providers and so-called neocloud firms such as Nvidia-backed CoreWeave has sunk to increasingly stressed levels².

The warning lands alongside other credit signals. The Wall Street Journal reported that the weak points in AI bonds are becoming clearer³. NPR examined Oracle’s borrowing, noting that the company has taken on a large volume of loans to build out its AI capacity⁴. Morgan Stanley’s projected AI-related capital spending for 2027 is concentrated, with roughly 77% attributed to high-quality hyperscalers and the remaining 23% sitting with weaker borrowers⁹.

Why it matters to investors

The issue S&P raises is not the size of the disclosed debt but the part that is not disclosed, because circular arrangements between suppliers, cloud resellers and end customers can leave obligations off the borrower’s own balance sheet¹. Without an estimate, creditors cannot price the gap¹.

The scale makes small errors expensive. Hyperscalers are spending an estimated $725 billion in 2026, and the analyst Russo expects a significant deceleration in that spending in 2027 and 2028⁷. Wall Street has been warning about overvaluation in what Korea Daily described as a $10 trillion infrastructure boom⁶. BlackRock chief executive Larry Fink told CNBC that a nationwide pause on US data-centre construction could trigger a recession, because of how much of the economy now rests on it^5,6. Tan Su Shan, chief executive of Singapore’s largest bank DBS, sees a change in sentiment, with investors starting to question AI returns⁸.

What to watch

S&P has not quantified the hidden leverage, so the first thing to watch is whether it publishes a number¹. The second is the neocloud bond complex, where prices have already moved to stressed levels and would move first if funding tightens².

The third is the concentration Morgan Stanley describes: if the 23% of projected 2027 capital spending held by weaker names is funded in public credit markets, that is where a repricing would start rather than at the large cloud companies⁹. Fink’s warning and Russo’s forecast of deceleration point the same way, towards capital expenditure as the variable that transmits credit stress into growth^5,7.