AI infrastructure is increasingly a credit-market risk, with stress concerns still unproven

AI infrastructure is no longer being discussed only as a stock-market story. It is increasingly a financing story: who lends, who buys the debt, and what happens if the returns from all that data-center spending arrive late.

The clearest alarm signal so far is a Bank of America fund-manager survey in which 48% of respondents named AI hyperscaler capital expenditure as the most likely source of a systemic credit event, meaning a broad market-wide credit shock rather than an isolated company problem. That is a notable shift in sentiment. It is not, by itself, proof that credit stress is already spreading.

The financing is real. Reuters reported that a rush by technology companies to fund AI infrastructure is boosting dealmaking and financing activity for Wall Street, generating fees from capital raising and loans. Goldman Sachs chief executive David Solomon said the buildout is still in its early stages and called it an AI capex super cycle that requires every financing instrument. Reuters also reported that Bank of America extended OpenAI a $520 million credit line and had helped raise nearly $500 billion for AI-related companies since 2025, according to internal data seen by Reuters.

The debt is not sitting only on hyperscaler balance sheets. Reporting points to public bonds, leveraged loans, private credit and project-style structures. PitchBook reported that Blackstone-owned QTS, a data-center operator that builds and operates facilities supporting cloud computing and AI, closed a $3.25 billion seven-year term loan B — a leveraged bank loan structure that is priced off a benchmark and tends to sit in the riskier part of corporate lending — at S+225, with a 0% floor and 99.5 original issue discount, which means it was sold at a small discount to face value so the lender’s effective yield is higher, after the deal was upsized twice from an initial $1 billion launch. The proceeds were meant to repay construction credit facilities, fund a debt service reserve account if needed and support general corporate purposes. That is active financing, not a frozen market.

The same pattern shows up in the public market. Forbes reported that Apollo said hyperscaler bond order coverage — the ratio of investor demand to the amount of bonds being offered, so lower coverage generally means weaker demand at the set price — fell from nearly five times the amount offered in February to below two times by July. Forbes also said Morgan Stanley’s $1.5 trillion figure is a funding gap, not a debt forecast, and that public bond spreads — the extra yield investors demand over a risk-free rate — remain near cycle lows, meaning borrowers are being charged relatively little for credit risk compared with earlier points in the market cycle. In other words, the market looks more like it is charging more for patience than snapping under pressure.

That is why the strongest reading is caution, not crisis. AI capex is increasingly a credit-channel question, but the evidence in hand supports concern and financing activity more clearly than a proven systemic break.

There is also a different bottleneck in view: power and infrastructure. Utility Dive, summarizing Bank of America analysis, said the market is constrained by where power can actually be delivered, not just by demand. It said more than 7.5 gigawatts of data-center projects with on-site generation — that is, projected power capacity tied to announced projects rather than current demand — are already under construction, with another 60 gigawatts-plus in pre-construction. That points to a buildout constrained by grids, permits and equipment availability as much as by borrower solvency.

So the question is not whether AI spending is large. It is. The question is whether the financing stack — banks, bondholders, private credit and project structures — can keep absorbing it without a wider repricing. For now, the evidence says the market is still absorbing the boom, even as it starts to price the risk more explicitly.