AI's new balance sheet: $96B in partner debt tied to OpenAI's compute build-out
The AI build-out is now a debt story. Suppliers of data centers, chips, and compute capacity serving OpenAI have piled up about $96 billion in debt, per a Financial Times analysis. Revenues across the stack aren't yet close to covering capex.
OpenAI has committed roughly $1.4 trillion for future energy and compute procurement. It expects about $20 billion of revenue this year. HSBC's math suggests even at $200+ billion in revenue by 2030, the company would still require another $207 billion to fund operations. That funding gap is bleeding into public credit markets.
Where the debt sits
- $30B already borrowed by SoftBank, Oracle, and CoreWeave.
- $28B in loans taken by Blue Owl Capital and Crusoe.
- $38B under discussion with Oracle, Vantage, and their banks.
- $96B total in partner debt tied to OpenAI-related build-outs.
Until this year, Big Tech funded most AI spend from cash flow. Now, leverage is doing more of the heavy lifting.
Hyperscalers are leaning on IG markets
Amazon, Google, Meta, Microsoft, and Oracle have issued about $121B in new debt this year to support AI initiatives, per Bank of America. That's more than 4x their five-year average ($28B).
BofA notes the week before Thanksgiving is usually the last heavy supply window. This time, it ran hot: ~$50B that week, ~$220B over the prior four weeks-around 70% above typical volumes. The analysts add hyperscalers contributed another $63B, effectively explaining the upsized supply.
Credit is blinking yellow
Spreads in the CDS market are widening as investors price higher default risk. Since late September, Oracle's 5-year CDS moved up ~+60 bps to ~104 bps. CoreWeave's CDS widened roughly +280 bps to around 640 bps, according to Deutsche Bank.
CDS are insurance-like contracts that pay out if an issuer defaults. Rising CDS levels often signal stress or uncertainty in forward cash flows. For a refresher on mechanics, see this overview of credit default swaps.
The CoreWeave question
CoreWeave's Q3 snapshot draws attention. The company reported $3.7B in current debt, $10.3B in non-current debt, and $39.1B in future data center lease obligations. It expects around $5B in revenue this year with a stated $56B revenue backlog.
Two tests matter: conversion and coverage. How quickly does backlog convert to cash, and at what margin after energy, networking, and depreciation? And can gross margin plus pre-sold capacity keep pace with interest expense, lease payments, and capex that isn't yet fully underwritten by take-or-pay contracts?
OpenAI's commitments vs. cash generation
OpenAI's commitments are enormous relative to current revenue. If future funding depends on public credit, expect higher coupons, tighter covenants, and more scrutiny of utilization and pricing. Any slip in model demand, unit economics, or energy availability could ripple back into suppliers' ability to service debt.
What finance teams should watch
- Concentration risk: Revenue and contract exposure tied to a single AI buyer (e.g., OpenAI) across lenders, operators, and lessors.
- Contract quality: Take-or-pay vs. best-efforts, termination rights, step-downs, and price-reset clauses tied to hardware or model efficiency.
- Utilization and ARPU: Actual GPU hours sold, pricing trends, and discounting needed to move incremental capacity.
- Coverage metrics: EBITDA-to-interest, FFO-to-debt, and lease-adjusted leverage as capex ramps ahead of revenue.
- Energy certainty: PPAs, interconnect timelines, and cost pass-throughs-key drivers for gross margin stability.
- Refinancing windows: Maturity ladders vs. issuance conditions; sensitivity to 50-100 bps rate moves.
- CDS and spreads: Ongoing repricing as IG supply stays elevated and investors hedge risk.
Scenarios to price in (2025-2027)
- Base case: Capacity ramps, demand absorbs at slightly lower unit pricing, spreads stay wide but orderly, capex funded with mixed cash and debt.
- Upside: Breakthrough use cases boost utilization; longer-dated, take-or-pay contracts improve visibility; spreads tighten.
- Downside: Model efficiency reduces compute intensity faster than expected; price competition compresses margins; refinancing costs rise; selective downgrades and delayed projects.
The takeaway
The AI build is now intertwined with public credit. Funding needs are large, cash generation is catching up, and investors are pricing the gap. If you're underwriting exposure-equity or debt-work from contract quality and cash conversion backward, not from narratives forward.
If your team is pressure-testing use cases and spend, these AI tools for finance can help benchmark adoption and workflow impact across FP&A, risk, and treasury.
Your membership also unlocks:
