By Gertrude Chavez-Dreyfuss
NEW YORK, Sept 30 (Reuters) - The artificial intelligence boom has arrived in the riskiest corners of US credit markets, where leery lenders are demanding more compensation to fund borrowers whose future earnings remain largely unproven.
AI-related issuance by low-rated firms has totaled $88 billion this year, according to Goldman Sachs, with most of the borrowing coming from US issuers. In the first 11 months of 2025, AI-related issuance in leveraged finance — mostly via
junk bonds and loans — was just $20 billion, analysts said, citing Neuberger Berman data.
Now investors are taking a harder look at these less-established borrowers — questioning their revenue projections, the value of their collateral and how much debt they can realistically support. This comes at a time when higher-rated AI issuers have been on a borrowing spree and a selloff in Treasury markets is pushing yields up across the board.
"High yield people like to know how much cash flow is coming, when that cash flow is coming, and what is the probability that the cash flow doesn't come," said Larry Holzenthaler, senior portfolio manager, fixed income, at Catalyst Funds. He said rising investor compensation demands reflect the fact that "these investors don't really participate in the upside because if everything goes well, you get paid back at par, and if it doesn't, the creditors end up taking the losses."
LIMITED RISK APPETITE
Overall, risk appetite for AI-related credit remains subdued despite rising issuance, market participants said. Leveraged finance buyers have clustered around higher-quality borrowers, favoring double-B-rated companies, or those just below investment grade. Demand remains strongest for issuers with predictable revenue streams backed by long-term contracts, tangible assets and established customer bases, fund managers said.
Data centers, many of which fall into that higher-quality category, have been doing much of the heavy lifting in a high-yield market that is otherwise shrinking, said Erin Brown, head of leveraged finance at BNP Paribas.
High-yield volumes are "basically flat" year on year, Brown said, and "if it weren't for the new money supply coming from data centers, we'd have volumes that were materially down year on year."
AI infrastructure supply in high yield has reached $40 billion so far this year, compared with $12 billion for the whole of 2025, according to BNP data.
AI companies, however, are unlikely to push far down the credit spectrum because of the higher borrowing costs it entails. Even near investment-grade, BB+ rated issuers are paying roughly 9% to 10% yields, analysts said, noting that lower rated borrowers could face borrowing costs of up to 14% to 15%.
For instance, SoftBank Group, which raised funds this month and has a BB+ rating, paid yields of 8.625% on 3.5-year notes, 9.25% on 5.5-year debt and 9.75% on 7.5-year bonds, according to a filing. Those are yields typically paid by significantly lower-rated companies, analysts said.
DIFFERENT BUYERS FACE DIFFERENT CONSTRAINTS
Debt issuers are also facing a more skeptical investor base in the leveraged finance market.
Unlike investment-grade investors, who can absorb large amounts of debt from highly rated hyperscalers, buyers of AI-linked high-yield bonds and leveraged loans are more constrained by factors such as portfolio rules that limit their exposure to riskier borrowers, analysts said.
"For debt investors, the proposition is fundamentally asymmetric," wrote Lotfi Karoui, multi-asset credit strategist at PIMCO, in a recent research note.
"Returns are largely contractual, driven by coupon, principal, and, at most, some spread compression," he noted, while the risks range from high debt levels to project delays and rapid changes in technology.
Many AI borrowers also require heavy upfront investment before generating reliable cash flow. Analysts said that makes their debt harder to absorb for the biggest buyers of leveraged loans — managers of collateralized loan obligations (CLOs) — particularly if leverage rises, cash continues to flow out, or ratings deteriorate.
CLOs typically buy a large portfolio of these loans, financing those purchases by issuing their own securities to investors.
"Anecdotally, CLO managers are becoming more cautious on certain names and really making sure they're checking everything before investing," said Elizabeth Templeton, senior product manager, fixed income & multi-asset indexes, at Morningstar.
The market's skepticism is already evident. Zenith Arc, a special-purpose project company, sold $2.25 billion of five-year senior secured notes in August to finance an Oklahoma data center leased to Jane Street. The bonds priced slightly below par at 99.50, with an 8.875% coupon.
Its bonds initially dropped almost three points to a bid price of 96.75 immediately after issuance, analysts said. By late August, Pender Fund Management wrote in a letter to investors that Zenith Arc bonds had lost more than seven points below the issue price.
Zenith Arc did not have publicly available contact details. A representative for Coatue, the investment firm behind Next Frontier, which is a joint-venture partner in the Zenith Arc project, said the firm had no comment.
For credit investors, the episode illustrates the heightened scrutiny facing AI-linked issuers as borrowing moves further down the quality spectrum. Capital remains available for the AI buildout, but investors are increasingly demanding proof that revenues will eventually support the debt being raised.
"The joke has become, 'What's revenue going to be like next year? I don't know, but it's going to be big,'" Catalyst's Holzenthaler said. "If you're a credit investor, that's a really bad answer from a borrower."
(Reporting by Gertrude Chavez-Dreyfuss, Editing by Colin Barr)
