The AI industry’s gigantic debts are making investors nervous
Credit markets are turning on the AI build-out. CoreWeave’s five-year bond has fallen 16% since July and now yields 15% (up from 10%), a distressed level; its credit-default swaps imply a more than 60% probability of default within five years, and its interest bill is about a quarter of revenue. Rival neoclouds Nebius and TeraWulf have seen yields rise, South Korea’s Elice fell by a third on its 19 October listing, and three other listings were delayed or cancelled. Of ten listed neoclouds only two are profitable; their combined debt is $100bn, nine times annual revenue, with $57bn issued this year. Wider borrowing is far larger: Oracle carries $169bn of debt ($25bn raised this year), SpaceX has issued $65bn this year and is seeking $40bn of chip-backed financing, the big four hyperscalers have issued $214bn in 2026 versus $83bn in all of 2025, plus $79bn through special-purpose vehicles and $84bn by chipmakers—nearly $850bn this year, about four times 2023, on top of more than $3trn already on balance-sheets. JPMorgan expects about $5.5trn of AI capex by 2030, three-quarters debt-funded. Off-balance-sheet layers add to it: Meta’s Sopaipilla SPV (80% BlackRock, $13bn of debt, 20-year lease), more than $250bn of lease liabilities plus $1.2trn not yet commenced, $1.5trn of purchase commitments, and Nvidia backstops of up to $300bn of customer debt. Holders are spread across Pimco (which bought $18bn of Meta’s $27bn Hyperion bond), Apollo and Blackstone private credit, insurers such as Athena, and hedge funds. The Economist argues these layers make it hard to see where losses would land (Fitch’s Roelof Steenkamp calls chip-backed financing “infinitely more complicated” to assess) and that the bigger risk is stagnating AI demand: Ramp card data show AI spending flat since July, which could delay Anthropic and OpenAI IPOs. Counterweights it reports: investors are not yet panicking, only about 15% of AI bonds issued this year are junk, Nvidia’s cash pile could absorb shocks, and Jensen Huang says CoreWeave’s cost of capital is actually decreasing. The neocloud debt-to-revenue figure was corrected on 22 October.
Why it matters
This is the financing side of every AI valuation Dealroom tracks. Neocloud credit stress, a cooling IPO window (Elice down a third, three listings pulled) and flat AI spend in Ramp’s data feed straight into private marks, late-stage AI rounds and the exit timetable for Anthropic and OpenAI. It also shows a growing share of AI capital coming as debt, SPVs and vendor backstops rather than equity, which equity-only funding datasets miss.
Executive takeaways
- Stress is concentrated in neoclouds: CoreWeave yields 15% with a CDS-implied default probability above 60%; only 2 of 10 listed neoclouds are profitable, with $100bn of debt at 9x revenue.
- The IPO window is narrowing for AI infrastructure: Elice fell by a third on debut and three other listings were delayed or cancelled.
- Scale: nearly $850bn of AI-related debt issued this year (about 4x 2023) on top of more than $3trn already outstanding; JPMorgan sees $5.5trn of capex by 2030, 75% debt-funded.
- Hidden leverage: $79bn through SPVs, more than $250bn of leases plus $1.2trn not yet commenced, $1.5trn of purchase commitments and up to $300bn of Nvidia backstops make loss paths opaque.
- Demand is the real test: Ramp shows AI spending flat since July; if that persists, Anthropic and OpenAI IPOs could slip.
- Not a panic yet: only about 15% of this year’s AI bonds are junk and Nvidia’s balance-sheet is a buffer—The Economist warns of fragility rather than forecasting a crash.
What The Economist may be missing
Likely thin: Europe barely features—Nebius is the only European-linked name—so there is no read on European neoclouds, data-centre developers or sovereign-AI projects and their financing; no split between debt that funds contracted (take-or-pay) capacity and speculative build; little on how private-credit funds and insurers mark these loans; and Ramp’s flat-spending signal comes from one card provider’s customer base, not the whole market. The equity side is also missing: how credit stress feeds into private-round valuations, secondaries and down rounds for AI startups.
