A $5 billion AI datacentre borrower defaults

Where things stand (14 Oct 2026)

I found no qualifying event since the question opened on 16 Sep 2026. Credit stress around AI infrastructure is clearly building, but none of the borrowers carrying $5B or more of AI debt looks close to default:

• CoreWeave (the most leveraged large AI cloud company). Its Q2 10-Q (EDGAR, filed 12 Aug 2026) shows $35.6B of total debt. It also shows $15.6B of liquidity: $5.5B cash plus $10B of undrawn facilities. In September it added a $4.2B convertible.
  • Ratings: Ba3 from Moody's, B+ (positive outlook) from S&P, BB- (positive) from Fitch.
  • Its credit default swaps reached about 855bp on 28 Jul 2026. On Bloomberg's model that implies roughly a 50% five-year default probability, but that is a market-priced figure that includes a risk premium, not a real-world estimate. The stock has since recovered from about $61 to roughly $80–88.
  • Consensus has net debt peaking near $100B in 2029, with free cash flow first positive in 2030 (TIKR, 8 Oct).
• Which CoreWeave entity counts matters. The rules don't let you add up debt across a group, and CoreWeave's parent-level notes were issued for "general corporate purposes" (8-Ks of Apr, Jun and Sep 2026). The clearly qualifying borrowers are its loan subsidiaries:
  • CCAC IV: about $6.2B at 30 Jun, amortising roughly $0.3B a month, so likely below $5B by around end-2026.
  • CCAC VIII: an $8.5B non-recourse facility, $2.84B drawn, which should pass $5B in 2027.
  • In a CoreWeave bankruptcy, I think court filings would probably establish the purpose and amount, but this is not certain.
• Oracle-linked projects.
  • S&P cut Oracle to BBB- on 9 Jul 2026, and its swaps trade around 150–215bp.
  • Project Jupiter in New Mexico has an $18B construction loan (borrowers: Red Chiles A–D LLC) that was quoted at 89–91 cents (FT/Reuters, 18 Sep). Oracle sent a force-majeure notice on 24 Sep.
  • But the lease is reportedly "hell-or-high-water": Oracle still owes carrying costs, and about three years of delay reserves exist. Near-term default is unlikely unless Oracle itself fails.
• Meta's Hyperion bonds (Beignet, $27.3B, rated A+) trade near 91, mostly because long-term interest rates have risen.
• More large borrowers are coming: SpaceX's roughly $40B chip financing (close expected 2027), a Broadcom package of more than $50B for OpenAI chips, an Oracle chip vehicle, and SB Energy, which says in its IPO filing that it is "substantially dependent" on OpenAI.
• Backdrop: the 10-year Treasury yield is around 5.3%. OpenAI's run-rate revenue is about $50B, and it projects $278B of cash burn over 2026–30, with cash lasting roughly into 2028 after raising $122B in March.

Reference class and base rate

• Single names: for B+/BB-/Ba3 issuers, historical five-year default rates are about 10–13%. Swap-implied real-world rates for CoreWeave come out around 15–20%.
• Past booms: large debt-financed buildouts (telecom 2000–02, shale 2015–16 and 2020, coal) all produced defaults of $5B or more within a few years of the peak.
• Prediction markets: Polymarket prices CoreWeave bankruptcy at 12.5% by 2027, 21% by 2028 and 24.5% by 2029. But Anthropic is priced at 9.5%, 18.5% and 23.5%, which shows these thin long-dated markets are inflated. I discounted them heavily. (I pulled these prices from Polymarket's public data feed.)

Paths and my estimate

Scenario mix through 2031:

Scenario	Chance	Chance of a qualifying event if it happens
Severe AI bust (OpenAI or another key tenant fails, GPU prices collapse)	about 25–30%	about 75–85%
Moderate correction	about 30%	about 15–20%
Benign	about 40–45%	about 5% (project-specific or international failures)

The main routes to YES:

• CoreWeave: about 15% chance of default × about 80% chance it resolves the question.
• Oracle-dependent project vehicles (needs Oracle itself to default, about 4%).
• Vehicles that depend on OpenAI (chip leases, SB Energy, Stargate sites).
• xAI/SpaceX chip vehicles.
• Other AI cloud companies above $5B.

Because these share the same risk factor, the union comes to about 26% by end-2031.

Timing. I built the series from quarterly hazards (computed in Python):

• About 0.4% for Q4 2026, since every large borrower has ample liquidity.
• About 0.6–0.9% a quarter in 2027.
• Rising to about 1.75% a quarter in 2029–30, as OpenAI's runway, CoreWeave's peak debt and the 2030–32 maturities converge.

Strongest case against my estimate

• Higher: CDS-implied probabilities, the many correlated borrowers, rates at 24-year highs and OpenAI's funding gap could justify 35%+.
• Lower: demand is still supply-constrained (CoreWeave raised prices 25% in July). Nvidia, Google, Meta and Microsoft backstops and rescue incentives are strong. Most debt is split into project vehicles below $5B or is "general purpose". Private amend-and-extend deals may never show up in the permitted sources (SEC filings, court dockets, or S&P/Moody's/Fitch reports). I land between these.

Sources: CoreWeave 10-Q and 8-Ks (EDGAR); Oracle 424B2s (general-purpose use of proceeds); Capacity Media, 12 Aug 2026; Octus, 31 Jul 2026; 24/7 Wall St, 29 Jul and 24 Sep 2026; El Paso Matters, 24 Sep 2026; Accredited Insight, 5 Oct 2026; Startup Fortune, Oct 2026; Quartz/FT, 6 Oct 2026; Yahoo/WSJ, 8 Oct 2026; Polymarket data feed (gamma-api.polymarket.com).

Question. By the last day of each quarter, has a borrower with at least US$5 billion of debt raised to finance AI datacentres or GPUs missed a scheduled payment, completed a distressed exchange, or filed for bankruptcy?

Definitions.

• Borrower: any legal entity worldwide, including a company or project-finance special-purpose vehicle; not limited to covered developers. The same entity must satisfy the debt threshold and experience the qualifying event. Aggregate facilities for which it is a borrower or joint borrower, but do not aggregate separately borrowed parent, subsidiary, affiliate, or project debt merely because it appears in consolidated accounts. A guarantee alone does not make the guarantor a borrower.
• At least US$5 billion of debt: at least $5,000,000,000 of funded principal outstanding immediately before the qualifying event, raised for the purposes below. Count bonds, loans, drawn revolving or delayed-draw facilities, and finance-lease principal attributable to delivered equipment or facilities. Exclude undrawn commitments, equity, accrued interest, operating leases, future purchase commitments, and trade payables. Count each obligation once. Refinancing of qualifying debt retains its qualifying purpose, but repaid debt is not counted again. Convert foreign-currency principal using the exchange rate stated in the qualifying financial disclosure, or, if none is stated, the relevant central bank’s published exchange rate on the event date or preceding business day.
• Raised to finance AI datacentres or GPUs: proceeds documented as financing construction, expansion, acquisition, or equipment of datacentres intended to support AI training or inference, or acquisition of GPUs for AI computing, including refinancing of such financing. Supporting power, cooling, networking, and land within such a project count. For mixed-purpose borrowing, count only the amount publicly attributed to these purposes. General corporate borrowing does not qualify merely because the borrower also invests in AI; a documented allocation or use of proceeds is required. GPU collateral alone is insufficient evidence of the use of proceeds.
• Missed scheduled payment: failure to pay in full a contractually scheduled principal or interest payment on any of that borrower’s financial debt by its operative due date. A publicly confirmed missed payment counts even if subsequently cured within a grace period; an agency default rating is not required. Use the contractual date after any amendments effective before that date. Failure to meet a covenant without a missed payment does not count.
• Completed distressed exchange: a completed exchange, repurchase, or restructuring that S&P Global Ratings, Moody’s Ratings, or Fitch Ratings explicitly classifies as a distressed exchange or default-equivalent restructuring; alternatively, a borrower filing or court finding must establish both that creditors received less than the original contractual promise (including reduced principal or interest, or deferred payment) and that the transaction was undertaken to avoid an otherwise likely payment default or bankruptcy. Announcement or solicitation alone is insufficient.
• Filed for bankruptcy: the borrower files a voluntary petition under US Bankruptcy Code Chapter 7 or Chapter 11, or initiates an equivalent formal insolvency, administration, or court-supervised restructuring proceeding in another jurisdiction. A creditor’s involuntary petition alone does not satisfy this branch. Solvent dissolution does not count.

Resolves YES if qualifying public sources establish the debt threshold, financing purpose, and at least one of the three events for the same borrower. The missed payment or distressed exchange need not involve the particular debt that financed AI infrastructure, and the financial distress need not be caused by AI. There is no minimum amount for the missed payment or exchanged debt.

Exclusions (resolve NO). Credit-rating downgrades, negative outlooks, high default probabilities, covenant waivers, ordinary refinancing, and proposed restructurings do not suffice without one of the defined events. Debt commitments of $5 billion or more do not suffice if funded qualifying principal is below the threshold. Resolve NO for a horizon if the required facts have not been publicly established by that horizon.

Resolving source. SEC EDGAR filings, including Forms 8-K, 10-Q, 10-K, 6-K, 20-F and their exhibits; equivalent filings in the borrower’s home securities regulator’s official repository; borrower-filed petitions or sworn financial schedules and court findings or orders in official insolvency dockets, including US bankruptcy courts through PACER; and published rating-action or default reports authored by S&P Global Ratings, Moody’s Ratings, or Fitch Ratings. These are the permitted resolving sources, narrower than Public confirmation (A2); press reporting, unsupported creditor allegations, and analyst estimates alone do not resolve. Sources may be combined, but all required facts must be documented. Apply the Resolution window and Resolution date (A3): the resolution date is the publication date of the last qualifying document needed to establish all requirements, not the underlying event date. The Observation period convention applies to reported debt balances; a pre-opening balance alone cannot establish eligibility.