CoreWeave's Debt Machine: How It's Financed and What It Means for AI

The short answer

CoreWeave's balance sheet at September 30, 2026 shows about $46.9 billion of debt (the author's sum of reported line items) against $4.0 billion of stockholders' equity, and its Q3 net interest expense of $846 million was about four times its $207 million of operating income. Much of the debt is tied to specific customer contracts, and the company reports a $132.5 billion backlog.

The same pattern is spreading: in the first week of November Oracle disclosed $10 billion of stock and $10 billion of notes at 7.0% to 8.65%, and Meta completed a $15 billion bond offering at 5.85% to 7.5%. The post explains the risks, CoreWeave's defense, and what to watch.

Hey everyone,

CoreWeave reported third-quarter results on November 10, and the numbers show both sides of the AI buildout in one place: revenue is up almost 170% from a year ago, and so is the amount of money the company owes. This post walks through how CoreWeave pays for its data centers, why that structure creates risks that its income statement alone does not show, and why the same pattern is now spreading to far bigger companies. I am describing a financing model here, not accusing anyone of anything. Where CoreWeave has a good argument, I give it.

What is CoreWeave, and how does it make money?

CoreWeave rents out computing power built around Nvidia chips to AI developers and big tech companies. Customers sign multi-year contracts, and CoreWeave builds or leases data centers, fills them with chips, and delivers the capacity. According to its third-quarter release, revenue was $3.66 billion, up from $1.37 billion a year earlier, and its contracted revenue backlog was about $132.5 billion. Its 10-Q for the second quarter shows how concentrated the customer base is: three customers accounted for 36%, 26%, and 10% of revenue, together 72% by my math. The 10-Q names OpenAI, Meta, Microsoft, and Jane Street as significant customers or commitments, including an OpenAI order form of up to roughly $6.5 billion through 2031.

How is CoreWeave financed?

Building this capacity costs far more than the company earns in a year, so it borrows. From the company's own filings, the structure has four main layers:

1. Contract-backed delayed-draw term loans (DDTLs). A special-purpose subsidiary signs a loan facility with banks and draws it down as chips and equipment are delivered. The loan is secured by the hardware and by the customer contract that will pay for it. The 10-Q describes, for example, an $8.5 billion facility entered into in March 2026, and in the third quarter CoreWeave closed a $2.6 billion facility designed for shorter-duration customer contracts.
2. Unsecured senior notes and convertibles. These are bonds sold to investors at the parent level, including euro-denominated notes, plus a convertible note offering in the third quarter and a new at-the-market stock program for up to 35 million shares.
3. Leases. Data center space is often leased, which adds to what CoreWeave owes. The balance sheet shows roughly $18.7 billion of operating lease liabilities.
4. Customer prepayments. CEO Michael Intrator said in the release that financing is supported by "customer contracts, operating cash flow, and approximately $9 billion in customer prepayments and deposits expected in the next three months alone."

The key design feature is that loans are tied to specific contracts. Lenders are betting on a customer's promise to pay (such as a large, highly rated company) more than on CoreWeave itself. That can lower the cost of borrowing and is a legitimate technique, similar to financing an airplane against a lease.

How much does CoreWeave owe?

Adding up the balance sheet in the Q3 release, recourse debt is $41.6 billion ($9.2 billion due within a year) and non-recourse debt is $5.3 billion, for total borrowings of about $46.9 billion. That is roughly 2.2 times the $21.4 billion at the end of 2025 (up about 119% in nine months, my math). Against that, stockholders' equity is only $4.0 billion, and the accumulated deficit is $4.7 billion. Cash and equivalents are $4.1 billion, plus $2.3 billion of restricted cash. All of these totals are my sums of reported line items.

The income statement tells the same story. In Q3, CoreWeave earned $207 million of operating income but paid $846 million of net interest, producing a net loss of $709 million. For the fourth quarter it expects interest expense of $1.05 billion to $1.15 billion. Annualized, that is more than $4 billion of interest for a company guiding to $13.0 billion to $13.3 billion of 2026 revenue. It also expects capital spending of $37 billion to $39 billion this year, nearly three times its revenue.

What are the risks of this structure?

• The cash arrives later than the bills. CoreWeave pays for chips and data centers up front and gets paid over several years. Growth consumes cash, so it must keep borrowing or selling stock to keep growing. If financing markets close, growth stops, and fixed obligations remain.
• The customer is the real collateral. A contract is only as good as the customer behind it. With a few customers making up most of revenue, trouble at one of them, such as a funding problem at an AI lab, would hit CoreWeave's lenders too. Many of those customers are themselves unprofitable AI companies or are spending heavily.
• The chips may not last as long as the loans. A loan can run six years while the chip generation turns over every year or two. If the hardware loses value faster than the loan is repaid, the collateral is worth less than the debt. Whether GPUs hold their value is an unresolved question that my earlier post on AI compute as an asset class discusses.
• Interest rates are not low. Fixed obligations are painful when growth slows. A higher-for-longer rate environment, with a Fed rate hike on the table, raises the cost of rolling over debt.
• Circularity. Nvidia sells chips to CoreWeave, backs the AI ecosystem with investments and financing platforms, and those customers buy more chips. When money flows in a circle, it is harder to tell real end-demand from financing-driven demand.

What is CoreWeave's defense?

The company has a serious case. First, the contracts are real, and the backlog of $132.5 billion is about ten times expected 2026 revenue. Second, it says Q3 adjusted operating margin reached double digits ahead of plan, that revenue grew by more than $1 billion sequentially, and that it is raising full-year guidance. Third, a large part of the debt is secured against signed contracts with big customers, and the roughly $9 billion of expected prepayments reduces how much it needs to borrow. Fourth, if demand for AI compute remains scarce, a company that owns capacity in a shortage can look very good in hindsight. Investors who own the stock are betting on exactly that. Adjusted EBITDA margin was 62%, so the underlying data center business earns a lot before interest and depreciation. It is the financing and the depreciation that turn that into a loss.

Is this just a CoreWeave problem, or an AI sector problem?

This is what I think matters most to everyday investors. CoreWeave is the extreme case, but the pattern is spreading to the largest companies in the world, and the filings from just the last week show it:

• Oracle disclosed in an SEC filing that it sold $10 billion of common stock on November 2 and also issued $10 billion of notes on November 4, with interest rates from 7.000% on the 2029 notes to 8.650% on the 2046 notes. The weighted average is about 7.8% (my math). It also has a mandatory convertible preferred stock outstanding.
• Meta completed a $15 billion bond offering on November 5, with rates from 5.850% (2031) to 7.500% (2056), about 6.6% on average (my math), and exercised in full an option to sell additional common shares tied to forward sale agreements announced October 26.

Two points. First, companies that used to fund AI spending out of free cash flow are now raising tens of billions in the capital markets within the same week. That is a sign of how large the spending is, not necessarily of weakness. Meta and Oracle are far bigger and more diversified than CoreWeave. Second, the interest rates are high for companies of this size. Mid-to-high single-digit coupons on long bonds mean lenders are charging real compensation, and the cost of the AI buildout is showing up in interest expense, not just in capital spending.

Sector-wide totals for AI-related borrowing in 2026 vary a lot by source, so I am not going to quote an industry-wide number. What the filings do show is a clear direction. The whole structure assumes that today's spending will earn enough in the future to pay back the debt, and that assumption has not been tested through a downturn.

What does this mean for investors?

This is not an argument that AI is a bubble or that CoreWeave is doomed. A debt-funded buildout can work out very well if the demand is real, which is how railroads and telecom networks were built. It also has a history of ending badly for the weakest borrowers when demand or financing conditions change, which is how those same industries produced some famous bankruptcies. The important thing is to know what kind of bet you hold. If you own CoreWeave, you own a highly leveraged bet on AI demand, and the stock's moves will be large in both directions. If you own an S&P 500 index fund, you already own Oracle, Meta, and Nvidia, so the sector's financing health matters to you indirectly.

What I would watch: whether interest expense keeps growing faster than operating income, whether customers keep prepaying, how long the new chip generations hold their value, and whether bond investors keep demanding rising yields from AI borrowers. Those signals will tell you more than any single quarter's revenue.

Key takeaways

• CoreWeave owes about $46.9 billion against $4.0 billion of equity, and its Q3 interest expense ($846 million) was about four times its operating income ($207 million).
• The debt is mostly tied to customer contracts, which is a real strength, but it concentrates risk in a few big customers.
• The company's defenses are not trivial: a $132.5 billion backlog, raised guidance, and about $9 billion of expected prepayments.
• The pattern is bigger than CoreWeave. Oracle and Meta raised $20 billion and $15 billion of debt, respectively, in the first week of November alone, at rates up to 8.65%.
• The unknown is whether the future returns arrive on time. If they do, the debt looks smart. If they slow, fixed interest does not.

More soon,
Learn to Love Money

Sources

• U.S. SEC EDGAR: CoreWeave Form 8-K with third-quarter 2026 results (Nov. 10, 2026)
• U.S. SEC EDGAR: CoreWeave Form 10-Q for the quarter ended June 30, 2026
• U.S. SEC EDGAR: CoreWeave Form 10-K for 2025
• U.S. SEC EDGAR: Oracle Form 8-K on $10 billion common stock and $10 billion notes offerings (Nov. 4, 2026)
• U.S. SEC EDGAR: Meta Form 8-K on $15 billion notes offering (Nov. 5, 2026)
• Learn to Love Money: AI compute as a new asset class