AI Borrowers Now Need 9–10% Yields to Get a BB Done — The Spread Story Behind the $725B Capex Cycle

The Sept. 30 Reuters analysis crystallizes a turning point that has been visible in sub-investment-grade prints since the summer: AI-related debt issued by low-rated firms reached $88 billion in 2026 YTD, per Goldman Sachs, with high-yield supply specifically tied to AI infrastructure already at $40 billion versus $12 billion for all of 2025 (BNP data). The crowding has not been matched by buyers willing to hold paper at flat spreads. Per the Reuters tally, even BB-rated AI issuers (the rung closest to investment grade) are clearing deals at 9–10% yields, and lower-rated borrowers are staring at 14–15% borrowing costs. The repricing has been concentrated in the structures that lack hyperscaler balance sheets behind them — single-tenant project deals and neocloud converts — while the largest tech names keep issuing but pay up on the long end of the curve.

The cleanest read on where the AI HY market is heading sits in a single 5-year deal. On Aug. 13, 2026, Zenith Arc LLC — a special-purpose project company backed by Coatue — sold $2.25 billion of senior secured green notes to fund a 149-megawatt Oklahoma data center leased to Jane Street. S&P rated the deal a preliminary BB with a positive outlook; the bonds went out at 99.50 with an 8.875% coupon. By late August, the issue had already drifted to a 96.75 bid — three points off the print — and by the week of Sept. 20, Pender Fund Management and other investors were marking the bonds at yields near 11.3%, more than seven points below issue price. The data center is leased, the tenant is investment-grade, the rating is BB with positive outlook — and the bond is being treated like a distressed credit.

The split inside the AI bond market is no longer about credit quality — it is about structural seniority. Hyperscalers continue to issue at near-IG tightness, even as 41% of their 2025–2026 supply has landed past 15 years. By contrast, AI high-yield bonds are 4% of the Bloomberg US Corporate High-Yield Index but account for 40% of net new issuance in 2026 — and that is where the indigestion sits. New-issue concessions on hyperscaler deals have widened from a tight 2–3 bps earlier in 2026 to about 20 bps on large prints, but the project and neocloud tier is taking the real hit.

Goldman Sachs now expects hyperscaler gross debt issuance to hit a record $420 billion in 2027, up roughly 60% from 2026's run rate, with about one-third of total AI capex debt-funded. The four-way combined2026 capex from Alphabet, Amazon, Microsoft, and Meta is now guided to $725–$745 billion, with Oracle having issued $43 billion alone since September 2025. The system can absorb that supply — but only if spreads stay close to current levels. When BB-rated borrowers need 9–10% to clear and project-deal bonds trade 7 points off in six weeks, the marginal cost of the next $100 billion of AI infrastructure is being repriced by the day.

The credit-market reset does not stay in the HY index. Business development companies sit directly between AI disruption and private credit, and their software books are large enough to matter. BIS data puts total BDC lending to software firms at roughly $115 billion — about a fifth of all BDC lending and more than 80% of their technology exposure. Ares Capital's portfolio carried $7.9 billion of software exposure across 135+ companies as of mid-2026, with management flagging ~22% of its roughly $30 billion portfolio as software-and-IT-services. By April 2026 Ares had hired an outside consultancy to grade every software investment for AI-disruption risk; it identified ~$1 billion as medium-risk. Blue Owl Capital Corp's parent gated withdrawals from a tech-focused private credit fund earlier in 2026 and reset OBDC's dividend 16% lower in Q2, pushing coverage to 116% — at the cost of a 14.3% drop in net investment income.

The split in the AI credit market flows directly into the supply chain that surrounds it. Hyperscaler-tenant data centers built for Microsoft, Alphabet, Amazon, and Meta keep finding buyers — those tenants write the checks. Single-tenant projects whose offtake is a private trading firm like Jane Street, or a software borrower facing AI disruption, are getting marked. The same divide shows up on the power side: utilities with hyperscaler PPAs and behind-the-meter generation contracts (Constellation Energy, Vistra) command premium multiples, while speculative merchant power depends on a demand assumption that the credit market is now testing. The BDC and private-credit stack picks up the residual — and the residual is where the losses live.

The next four to six weeks settle whether the Sep 30 article is a warning shot or a regime change. The Reuters piece cites a Goldman figure that 15 insurance clients entered30-year AI paper with $50M+ orders in Q1, falling to about half that number by the end of Q2 — the marginal buyer is already stepping back from the longest end. Watch for: (i) follow-on prints from the CoreWeave/Nebius convert tier, where a higher coupon on the next deal would confirm the reset; (ii) BDC Q3 NAV disclosures, where realized losses on software names get crystallized; and (iii) any new 5-year project-finance deals — a single sub-10% print would mark the bottom; a single 12%+ print would confirm the Zenith Arc pattern.