Fourteen Percent to Play: AI's Junk-Bond Era Arrives as Low-Rated Borrowers Tap Credit for $88 Billion

The artificial intelligence boom has officially arrived in the riskiest corners of the US credit market — and the welcome has been frostier than the industry hoped.

On September 30, Reuters published an analysis documenting what strategists had been warning about for months: AI borrowers without proven revenue streams are facing a materially tougher sell. AI-related debt issuance by low-rated firms has reached $88 billion this year, according to Goldman Sachs, with most of the borrowing coming from US issuers. But the money is no longer cheap, and it is no longer easy.

The numbers behind the squeeze

To understand how fast this market grew, compare it to last year. AI-related issuance in leveraged finance — primarily junk bonds and leveraged loans — totaled just $20 billion in the first 11 months of 2025, according to Neuberger Berman data. This year’s $88 billion figure represents more than a four-fold increase in under a year.

Within that, data center and AI infrastructure supply in the high-yield market specifically has reached $40 billion so far this year — up from just $12 billion for the entirety of 2025, according to BNP Paribas. In a high-yield market that is otherwise shrinking, AI has become the dominant source of new supply.

But volume is not the same as enthusiasm. Despite rising issuance, Reuters reports that overall risk appetite for AI-related credit remains subdued. Leveraged finance buyers are concentrating on higher-quality borrowers — particularly double-B-rated companies with predictable revenue streams. Data centers, which often fall into this higher-quality category, have been crucial for keeping the high-yield market alive at all.

What lenders want now

The mechanics of the skepticism are straightforward. “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,” Larry Holzenthaler, senior portfolio manager for fixed income at Catalyst Funds, told Reuters. He noted that rising investor compensation demands reflect an asymmetric risk: investors are repaid at par if all goes well, but bear the losses if it does not.

That asymmetry is now priced in. Issuers near investment grade — BB+ rated companies — are already paying yields of around 9% to 10%. For lower-rated AI borrowers, borrowing costs could reach 14% to 15%.

The benchmark cases are instructive. SoftBank Group, rated BB+ by S&P, recently paid yields between 8.625% and 9.75% on its notes — levels that analysts noted are typically associated with significantly lower-rated companies. This is the same SoftBank that in late September completed the largest high-yield corporate bond deal on record globally, an $11.1 billion multi-tranche sale intended partly to cover its OpenAI funding commitments.

The Zenith Arc cautionary tale

If SoftBank shows what strong-junk AI credit costs, Zenith Arc shows what happens when the market’s patience runs out. The company sold $2.25 billion of five-year senior secured notes in August, a data center project backed by investment firm Coatue, with Jane Street as the lessee of the underlying Oklahoma facility. The notes priced slightly below par — and subsequently fell more than seven points below their issue price, according to Pender Fund Management.

Seven points is a significant move for senior secured paper. It signals that even deal structures with tangible collateral, a blue-chip lessee, and investment-grade-adjacent sponsorship are getting repriced in real time. Investors are scrutinizing revenue projections, collateral value, and debt capacity of less-established AI borrowers — and voting with their bids.

Why CLO caution matters

The most systemically interesting detail in the Reuters report concerns collateralized loan obligations. Unlike investment-grade bond funds, buyers of AI-linked high-yield bonds and leveraged loans face structural constraints — portfolio rules that limit exposure to riskier borrowers. Many AI borrowers require substantial upfront investment before generating reliable cash flow, making their debt harder for CLO managers to absorb, especially if leverage increases or ratings deteriorate.

CLO managers are reportedly becoming more cautious on certain AI-linked names. Lotfi Karoui, multi-asset credit strategist at PIMCO, described the proposition for debt investors as fundamentally asymmetric: largely contractual returns versus significant risks from high debt levels, project delays, and rapid technological change. Erin Brown, head of leveraged finance at BNP Paribas, noted that AI companies are unlikely to push further down the credit spectrum simply because the borrowing costs there are prohibitive.

The implication is a bifurcated market: hyperscalers and strong BB issuers continue to borrow at scale — Goldman Sachs estimates total AI-linked debt issuance across all ratings has already reached roughly $500 billion in 2026, with hyperscaler issuance projected at about $250 billion for the full year — while speculative AI credits increasingly find the door closed or priced at 14%+.

The macro backdrop

None of this happens in a vacuum. A selloff in Treasury markets has been pushing baseline yields higher across the board, compounding the risk premia AI borrowers must pay. The 10-year Treasury has traded near multi-decade highs this fall, and every basis point of risk-free rate increase flows directly into the coupon math of a data center project that may not generate contracted cash flow until 2028.

There is also a genuine open question about what the collateral is actually worth. GPUs depreciate fast and resale poorly in a downturn. An SSRN paper from August 2026 on pricing and securitizing AI compute made the point bluntly: for GPU collateral, recovery is neither constant nor independent of the credit cycle — a wave of defaults among AI borrowers would coincide with exactly the conditions that impair the value of the chips backing the loans.

What it means

Three takeaways stand out.

First, the marginal dollar of AI infrastructure is now debt-financed and price-discovered. Roughly one-third of 2026 AI capex is expected to be debt-financed, per Goldman Sachs. As equity rounds at $1.4 trillion valuations grab headlines, the credit market is quietly performing the more rigorous valuation work — and its verdict is increasingly “show me the cash flow.”

Second, the shakeout will be selective, not systemic — so far. The stress is concentrated in speculative-grade, project-level financing. Investment-grade hyperscaler paper still clears the market, albeit with wider spreads. The system-level risk would materialize if the BB data center cohort began downgrading into single-B territory, forcing CLOs to sell into an illiquid market.

Third, the cost of capital is becoming the binding constraint on the AI buildout. When a lower-rated AI borrower faces 14-15% financing costs, entire classes of marginal data center projects stop penciling out. The AI capex supercycle doesn’t end with a bang; it ends when the weighted average cost of capital exceeds the contracted return on compute — and this week’s price action says that threshold is being tested at the speculative edge of the market.

The AI boom was built on the assumption that capital would always be available. The $88 billion already raised this year proves the demand side of that assumption. The 9-15% yields, the cautious CLOs, and Zenith Arc’s seven-point drawdown prove the supply side is no longer unconditional.