NEW YORK — For most of the past year, the answer to 'who will finance the AI build-out?' was: almost anyone with a balance sheet. According to Reuters, that changed in a matter of days.

Banks and credit investors are scrutinising AI-related debt far more closely and telling market participants they are reluctant to take on more exposure, several banking and credit sources told the agency. The hesitation extends even to deals backstopped by marquee names such as Anthropic and Google. Neither company commented.

The proximate cause is the selloff. On Thursday the Nasdaq 100 erased $2.5 trillion in market value and slid into correction. The Democratic sweep in the midterms added a second worry for lenders: more oversight of data centres. Two sources said the shift in appetite has come only in the past few days.

'Over the past year there was unbridled enthusiasm to provide debt to neoclouds, anyone who was building data centers or some sort of AI asset,' said Chris Poggi, a finance partner at Ropes & Gray. 'Debt investors' books have become full and now they are assessing if they have extended themselves as much as they're going to extend themselves this year.'

The supply side is not slowing down. Goldman Sachs Research projects global hyperscaler debt issuance of $420 billion in 2027, more than 60% above its $250 billion estimate for fiscal 2026.

The exposure numbers explain the caution. At JPMorgan Chase, loans backed by other income-producing properties — a category that includes data centres — reached $29 billion outstanding at the end of the third quarter, the bank's second-largest commercial real estate exposure at 12%, up from 8% three years ago. Bank of America credit analyst Yuri Seliger put the year-to-date widening of 10-year hyperscaler spreads at about 50 basis points, against roughly 20 for high-quality industrial companies with no AI link. Meta bonds issued earlier this year and SpaceX paper are trading below par.

Underwriters are also stress-testing the contracts beneath these projects — service obligations, operational deadlines, and the conditions under which a large technology customer can cut or terminate commitments. Chipmakers typically backstop the senior debt used to buy their accelerators and guarantee payments on long-term data-centre leases, a structure that until recently made the loans easy to sell.

Riskier borrowers feel it first. CoreWeave's stock is down 35% year to date and its bonds have dropped sharply on the secondary market; Baird Equity Research flags CoreWeave and Oracle as exposed to possible construction delays or moratoriums. Nebius explored a high-yield bond in September and decided against it last month, one source said; in October it instead took a $3.5 billion senior secured loan facility from Citigroup.

In Asia, lenders describe a stricter posture. 'The kind of deals we look at, we are very focused on the risk aspect. So we structure the deals around to make sure that the cash flow is there,' OCBC Group CEO Tan Teck Long said on an earnings briefing on Friday.

Not everyone reads it as a turning point. Jay Bala of AIP, a $300 million alternative asset manager, expects most banks and credit investors to 'take a wait and see approach' on how AI growth fares after the midterms.

What the Reuters reporting establishes is narrower and more concrete: the marginal lender is no longer available on the same terms, and the sector's next $420 billion has to be raised from a market that just watched $2.5 trillion disappear in a day.