CoreWeave’s Debt Explained: GPU‑Backed Loans, Bonds and CDS

$35.6 billion of principal at June 30, 2026, if you use the 10-Q’s own definition of debt; $35.1 billion if you use the balance sheet’s carrying value; about $51.6 billion if you also count lease obligations. All three are correct for what they include, which is why different articles quote different totals. The table rebuilds each from the filing.

Scope: this page explains CoreWeave’s borrowing as reported in its 10-Q for the quarter ended June 30, 2026 and later 8-Ks, and how bond yields and CDS are read. It does not estimate a default probability, rate the bonds or say whether to own the stock or the bonds. Market levels quoted are as reported by the press on the dates shown.

Debt has grown fast: the carrying value was $21.4 billion at December 31, 2025 (20,726 recourse + 647 non-recourse). For scale, our arithmetic on the same balance sheet puts debt at 6.98 times stockholders’ equity of $5.0 billion. The company in context, with its customers and suppliers, is in what is a neocloud.

Two very different layers: loans secured by specific GPUs and the customer contracts they serve, held in separate subsidiaries, and unsecured bonds and convertibles owed by the parent company. Which layer a lender sits in decides who gets paid first, from which cash, and on what schedule. Principal by type, from the Note 10 table (our grouping):

It is a construction-style loan for computers: lenders commit a sum, the borrower draws it as it buys GPUs for a named customer contract, and the loan is repaid monthly out of that customer’s payments. CoreWeave’s 10-Q says these loans are secured by the financed assets and pledged contractual cash flows, generally from investment-grade counterparties, and that they amortize as those cash flows arrive, with excess cash made available to the company. At June 30, 2026 the special-purpose subsidiaries’ debt was secured by $18.2 billion of non-current assets, mostly equipment, and $2.6 billion of current assets, and CoreWeave was in compliance with all covenants under these facilities.

“Non-recourse” means the lender can take the subsidiary’s assets but cannot generally pursue the parent, except for narrow carve-outs such as fraud. The newest example, DDTL 4.0, is an $8.5 billion facility signed in March 2026 with an effective rate near 7%; the oldest, DDTL 1.0 from 2023, carries an effective 15%. Rates differ by vintage and by the contracts and guarantees behind each loan. The August 2026 DDTL 5.5 ($2.6 billion, a floating benchmark rate plus 5.50%, maturing September 2031) is guaranteed by the parent. Who those customers are is in who are CoreWeave’s customers.

Because their market price fell, not because CoreWeave pays more on them: the coupon is fixed when a bond is issued, and the yield is an annualized calculation assuming contractual payments are made; realized returns depend on default, sale price and coupon reinvestment. At June 30, 2026 the five senior-note issues had a fair value of $10.0 billion, about their $10.0 billion principal. By October 15, CNBC reported the 9.75% notes due 2031 at a 15.78% yield to maturity, and on October 19 the 9.625% notes due 2032 at 14.9%. Worked example for the 9.75% notes, standard semiannual pricing, about five years left:

Why it matters to shareholders: the coupons on existing notes do not change, but new unsecured borrowing would have to be priced near the market yield, and CoreWeave keeps borrowing to build. Rates moved too: the Federal Reserve’s policy rate stood at 3.75%–4.00% (raised on September 16, 2026, held on October 28) , and the 10-Q estimates that each 1 percentage point move in rates changes interest on its floating-rate debt by about $30 million a quarter. For rate sensitivity in a whole portfolio, see portfolio rate sensitivity.

A credit default swap is insurance against a borrower failing to pay; a five-year spread of 1,000 basis points means protecting $10 million of CoreWeave debt costs roughly $1 million a year (10%) for five years, in annual-cost terms. CNBC reported CoreWeave’s five-year CDS at about 1,038 basis points on October 15, 2026 after trading above 1,140 intraday, about 1,082 on October 19 and 1,145 on October 21. A credit-pricing executive quoted by CNBC said 1,000 or more is generally a sign of credit distress. For comparison, Oracle’s five-year CDS was 315 basis points on October 19.

What it does not mean: a known chance of default. Some headlines converted the CDS level into “odds of default”. That conversion is a pricing convention, not a forecast. It assumes a recovery rate, and the price also includes a risk premium and the cost of trading a thin market, so the implied number is typically higher than any real-world estimate. This page reports the levels and leaves the conversion out. For a balance-sheet screen that does not depend on market prices, see the Altman Z-score.

Not from operating profit yet; so far from new financing, prepayments and contract cash flows. In Q2 2026 net interest expense was $640 million against an operating loss of $49 million. On the company’s adjusted EBITDA of $1,510 million interest looks covered about 2.4 times (1,510 ÷ 640), but that measure adds back $1,393 million of depreciation, the cost of the GPUs wearing out. In the first half of 2026 operating cash flow was $3.7 billion (after interest paid), capex $14.1 billion, new debt $16.7 billion, repayments $5.2 billion and new equity $3.0 billion.

Most of the early payments are installments on GPU loans that amortize as customers pay; the bonds and convertibles fall due from 2030. Against these, CoreWeave reported $15.553 billion of total liquidity: $5.524 billion of cash, $15 million of marketable securities and $10.014 billion of commitments available under its revolving facility and delayed-draw loans. Undrawn GPU loans can generally be drawn only to buy equipment for specific contracts, so they are not spare cash. Debt classified as current ($7.5 billion) exceeded unrestricted cash ($5.5 billion) at June 30. Management said existing cash and available borrowing would be sufficient for obligations due over the following year.

The alarming reading overstates the risk when it treats $35.6 billion as one lump: $13.6 billion is secured, amortizing debt repaid from named contracts, and lenders priced the March 2026 non-recourse loan near 7% while the unsecured bonds trade near 15%. Bond prices also fell across AI-related high-yield debt in October 2026, partly because Treasury yields rose (CNBC, citing Janus Henderson and Goldman Sachs). The reassuring reading understates it when contracts slip or a large customer does not renew: in a pledged structure the secured lenders are paid first, so any shortfall lands on the unsecured bondholders and shareholders. The three largest customers were 36%, 26% and 10% of Q2 2026 revenue.

Who should be careful with this story: anyone reading a 15% yield as a safe 15% income. The yield is high because the market sees real risk of not being paid in full. To watch next: the Q3 2026 10-Q (refinancing terms, draws on DDTL 4.0, 5.0 and 5.5, the 2033 convertibles), interest expense against operating income, and how revenue backlog turns into delivered capacity (see backlog, RPO and ARR explained). How CoreWeave’s debt compares with Nebius, IREN and the data-center landlords is in neocloud stocks compared, and the stock’s price history is on CoreWeave’s stock page.