CoreWeave's Backlog Hits $104 Billion; Interest Is Four Times Profit

Two Numbers, Released Half an Hour Apart

On the evening of August 11, minutes after the New York close, CoreWeave reported its second quarter. Revenue had doubled in a year, from $1.21 billion to $2.58 billion — a 112% increase. The stock rose 13% in after-hours trading, touching $102.

The same release carried a second number. Net interest expense for the quarter was $640 million, against $267 million a year earlier. In the guidance issued that night, the company said the bill would rise to between $860 million and $940 million in the third quarter. In the same guidance, adjusted operating income was projected at $200 million to $260 million.

CoreWeave had, in its own words, told the market it expects to pay roughly four times its operating profit in interest next quarter. The market heard that sentence and bought the stock. This piece is about how both reactions can be right at once.

The first two numbers are a growth story. The last two are a financing story. What the company is worth depends on which one is running faster.

The Business of Renting Out Compute

CoreWeave is the largest of the so-called neoclouds. The model is plain: buy GPUs from Nvidia in volume, install them in data centers, and rent them to companies training AI models under contracts that run for years. Unlike Amazon or Microsoft, it does not sell a software layer on top. It sells compute.

The appeal is that revenue can be locked in up front. Customers commit to pay whether or not they use the capacity. CoreWeave's IPO filings disclosed that 62% of its 2024 revenue came from a single customer, Microsoft. The base has since widened: contracts with Meta covering 2027 through 2032 now total $35 billion, OpenAI and Google were added, and a multi-year agreement with Anthropic was signed. The company now says no single customer will account for more than 35% of contracted revenue.

The cost of the model is this: you buy the chip today and collect the money over years. You bridge the gap with debt.

That equation holds for the whole sector. Financing AI data centers became the busiest corner of the bond market, private credit and securitization in 2026, with lending structures that pledge the GPUs themselves as collateral. CoreWeave is that market's most visible borrower. On the call, management pointed to its ability to securitize shorter-duration contracts — financing enterprise customers who commit for two or three years rather than five — as evidence that the credit window is open. That it is open is true. How long it stays open is a separate question.

Total debt stood at $35.07 billion on June 30, against $5.52 billion of cash. Annualize the quarterly interest and the blended cost of that debt works out to roughly 7.3% — well above Treasury yields, because the collateral is not a building but a stack of processors whose worth three years out is genuinely unknown.

A Backlog Is a Promise; a Megawatt Is a Calendar

The reason the stock rose was the backlog: $104.2 billion as of June 30, up 246% from a year earlier, with more than $25 billion added in the opening weeks of the third quarter. The mix also lengthened — contracts running beyond four years grew from 10% to 21% of the book.

But a backlog is not revenue. A contract becomes revenue only when the capacity is actually running, and the narrowest link in that chain is electricity.

The distance between those figures shows up in the company's own targets. Active power was 1.5 GW at quarter-end against 4.2 GW contracted — barely a third of the committed capacity is live. The year-end target is 1.85 GW: 0.35 GW added in six months. Management says it sees a clear path to at least 8 GW by 2030. That path is four and a half years long.

The backlog tells you how strong demand is. Megawatts tell you when that demand turns into cash. They are not the same thing.

Where a Hundred Dollars Goes

Here is the arithmetic at the center of it. CoreWeave's adjusted EBITDA margin is 59% — a software-like figure, and the number its defenders cite most. In the same quarter, the GAAP net loss margin was negative 24%. What sits between the two?

The point of that arithmetic is not that the 59% EBITDA margin is fake. It is that the same $58.60 has to both replace the aging machine and service the loan. Today it does not stretch far enough to do both.

The guidance sharpens the picture. Third-quarter revenue guidance has a midpoint of $3.53 billion; the interest midpoint is $900 million. Interest as a share of revenue rises from 24.9% to 25.5%. Revenue grows 24% sequentially while interest grows 40%.

The Sentence Punished in May and Rewarded in August

The company also updated its 2026 outlook that night, lifting the revenue midpoint 2.4% to $12.8 billion and the capital expenditure midpoint 12.1% to $37 billion.

That means raising this year's spending by $4 billion to raise this year's revenue by $300 million. The company's defense is reasonable: the spending belongs to today, the revenue to 2027 and beyond. But three months ago the market read the identical sentence the other way. On May 7, reporting the first quarter, CoreWeave again raised spending while leaving revenue guidance soft, and the stock fell 10%.

What changed was not the structure of the numbers. It was the size of the backlog. In May, the extra spending looked like an unbacked bet. In August, it looked like the investment a $104 billion book requires.

The Other Side of the Table

Who is on the other side of that book? A backlog is worth exactly as much as the people who signed it can pay.

Contracted commitments across the four largest AI labs have been reported at $66.8 billion. All are strong balance sheets, or companies backed by strong balance sheets. But all are making the same wager: that AI models will eventually earn enough to justify today's spending. The backlog is diversified by customer. It is not diversified by thesis.

What Is Left Over

Free cash flow for the quarter was negative $5.74 billion. Sources differ on capital expenditure: the company's cash flow statement shows $6.42 billion for the three months, while outlets working from the presentation slides cite $9.35 billion on a committed basis. Both describe the same thing — somewhere between two and three times quarterly revenue leaving not the till but the credit market.

The full-year outlook makes the ratio plainer still: $12.4-13.2 billion of revenue against $35-39 billion of capital expenditure. Roughly $3 spent for every $1 earned.

That is not an accusation in itself. Railways, telecom networks and power plants were all built this way: a large debt first, a long collection after. But in this model the outcome turns on how long the asset lasts. A railway runs for forty years, and the technology of the rail does not go obsolete. How many years a GPU produces revenue is the number the industry argues about most and knows least. Lengthening the depreciation schedule means booking profit from the same machine for longer; shortening it does the reverse. With $1.393 billion of depreciation in a single quarter, a one-year change in that assumption reshapes reported profit directly.

The maturity of the debt sits on the same calendar. If part of the $35 billion comes due before the contracts it financed run out, the company has to refinance before the revenue arrives. Nobody knows what mood the credit market will be in at that moment.

The chart above shows the ground this argument is being fought on over the past three months; CoreWeave joined the Nasdaq 100 in early August and is now on the receiving end of passive fund flows as well.

The Lesson: Growth Has to Outrun Interest

A company that grows on debt has two curves. One is revenue, the other is interest. The story is about which curve is steeper — the fact that both point up tells you nothing.

At CoreWeave, second-quarter revenue rose 112% year over year while interest rose 140%. Third-quarter guidance has revenue growing 24% sequentially and interest 40%. The company says the gap will close, because money spent today earns its return in 2027 and beyond. That claim may well be right. For it to be right, the length of the contracts, the working life of the processors and the maturity of the debt all have to meet on the same calendar.

Figures in this piece come from CoreWeave's second-quarter press release of August 11, 2026, the guidance given on that evening's earnings call, and the financial press reporting both. Where sources diverge on capital expenditure, the discrepancy is preserved in the text as noted. Several customer-level contract values have not been disclosed by the company and rest on press reporting. Share prices and charts are live data and may differ at the time of reading.