CoreWeave's EBITDA was $1.5B. Depreciation and interest were $2.0B.

CoreWeave rents access to graphics processors. It buys them with borrowed money, signs multi-year contracts with customers who need compute, and reports the result. In the June quarter revenue grew 112.5% to $2,575 million.

Operating income went from positive $19 million to negative $49 million over the same period.

Both numbers are correct, and the distance between them is the whole business. This article works through where it comes from, using the company’s own tables.

The quarter

The last row is the one that does not fit the story of scale. Revenue doubled, adjusted EBITDA doubled, and adjusted operating income — the same measure after depreciation — fell by more than a third.

What adjusted EBITDA leaves out

Adjusted EBITDA is a disclosed non-GAAP measure with a published reconciliation, and CoreWeave provides one. Here are its two largest components for the quarter.

Net loss                                −$626M
add back: depreciation & amortisation +$1,393M
add back: interest expense, net         +$640M
add back: other reconciling items       +$103M
─────────
Adjusted EBITDA                         $1,510M

Depreciation and interest together are $2,033 million — $523 million more than the adjusted EBITDA they are excluded from.

For many companies, adding back depreciation and interest is a reasonable way to see operating performance before capital structure and accounting choices. Here it is worth pausing, because:

• Depreciation is the cost of the product. CoreWeave’s product is access to GPUs. The GPUs wear out, or become obsolete, on a schedule. That schedule is the cost of goods.
• Interest is how the GPUs were bought. This is not legacy leverage from an old acquisition. The company drew $13,457 million of new debt in this quarter.

Neither is a one-off. Both are growing faster than revenue.

The engine

The scale of the build is easier to see in the cash flow and balance sheet than in the income statement.

Capital expenditure in a single quarter was $6.4 billion, against $14.1 billion across the six months. The funding came from debt at roughly twice the rate the capex consumed it — capex was 47.7% of gross new borrowing in the quarter, with the balance going to repayments, cash and other uses.

The balance sheet records the result:

Debt rose $13.7 billion in six months. Total assets rose $27.8 billion. This is what a deliberate, financed infrastructure build looks like in accounting, and CoreWeave has never described it as anything else.

The backlog, and the thing under it

Against that, the company reports a revenue backlog of approximately $104 billion as of 30 June, excluding more than $25 billion of commitments it says were added in early Q3.

At the June quarter’s revenue rate, $104 billion is roughly ten years of revenue already contracted. That is the number that makes the debt legible: you borrow against contracted cash flows, not hopes.

Which is why one sentence in a separate filing is worth reading closely. On 10 August, CoreWeave announced a $2.6 billion delayed draw term loan and described its structure this way:

Facility’s approximate five-year maturity extends beyond the average three-year length of underlying customer contracts, reflecting lender confidence in long-term GPU demand

The company presents this as a positive, and from a financing-flexibility standpoint it plainly is — it lets shorter contracts support longer debt, which widens the customer base it can serve.

It also states the arithmetic plainly: for this facility, the debt outlives the contracts behind it by roughly two years. Whether that gap closes depends on renewals at rates that have not been agreed yet, on hardware that will be two years older, in a market where the price of compute is not fixed.

That is not a prediction and this article does not make one. It is the risk the structure contains, disclosed by the company, in its own sentence.

What would make this wrong

• Adjusted EBITDA is disclosed and reconciled. Nothing here suggests the measure is improper. The point is narrower: the two largest exclusions are, for this specific business model, recurring operating costs rather than accounting artefacts.
• Depreciation schedules are an estimate. How quickly GPUs are written down materially changes reported income, and the useful-life assumptions are in the annual report rather than this release. A longer life would raise reported earnings without changing any cash flow.
• Debt secured against contracted revenue is not the same as unsecured leverage. Part of CoreWeave’s borrowing is non-recourse and tied to specific customer contracts. We have shown recourse and non-recourse separately for that reason.
• Backlog is not revenue. The company states it includes remaining performance obligations plus estimated future amounts, subject to delivery and availability. The ten-year figure is arithmetic on a current run rate, not a schedule.
• One quarter of adjusted operating income decline is one quarter. It may reflect the timing of deployments rather than a trend.
• These are unaudited results in an 8-K. The 10-Q and eventually the 10-K will carry the audited and more detailed versions, including customer concentration.
• This article makes no claim about whether the strategy works. Financing long-lived assets with debt against contracted demand is an ordinary way to build infrastructure.

Check it yourself

Both filings are linked below and were made a day apart. The income statement, cash flow statement, balance sheet and EBITDA reconciliation are all in the 11 August earnings release; the maturity sentence is in the 10 August facility announcement.

CoreWeave also appears at the top of a screen we ran of companies whose capital spending exceeded their operating cash flow, at 337% for 2025 — the highest of any listed non-utility company above $1 billion of operating cash flow.

If a figure here does not match a filing, tell us and it will be corrected on the article and on the corrections log, with the date.