Analysis · AI Infrastructure · August 17, 2026
CoreWeave’s Operating Loss Was $49 Million. Its Net Loss Was $626 Million.
CoreWeave grew second-quarter revenue 112% to $2.58 billion and finished the quarter roughly at operating breakeven — a loss from operations of $49 million on that revenue base. It then reported a net loss of $626 million. Almost the entire gap is one line: $640 million of net interest expense in three months, the cost of the debt that bought the graphics processors. The operating business is close to paying for itself. The balance sheet that assembled it is not.
This is the second thing I have found this week by reading the neocloud filings instead of the headlines. The first was an accounting estimate — Nebius quietly extending how long it assumes its servers last, which added $75.7 million to first-half net income. This one is more structural, and it runs through both companies.
Between them, CoreWeave and Nebius disclosed $141.2 billion of contracted, unsatisfied revenue as of June 30. That number is real, it is audited, and it is the strongest argument the bulls have. It is also long-dated, heavily concentrated in a handful of counterparties, partly composed of management estimates, and financed at a cost that currently exceeds what the operating businesses earn. All four of those facts come from the filings themselves.
1. The financing bill
Start with the income statement, because it is unusually legible this quarter. CoreWeave’s revenue nearly doubled year over year. Total operating expenses came in just above revenue, producing a $49 million operating loss — a rounding error against a $2.58 billion quarter, and a genuine improvement in the underlying unit economics. Then interest takes $640 million, other income returns $125 million, tax takes $62 million, and the quarter closes $626 million in the red.
CoreWeave, three months ended June 30, 2026, in millions of U.S. dollars. Every bar is a line item as reported; the bridge is arithmetic, not adjustment.
For the first half the pattern is the same and larger: a $193 million operating loss, $1,176 million of net interest, and a $1,366 million net loss. Interest expense is not a side effect of this business model — at current scale it is the loss.
2. What $141 billion of backlog actually says
Both companies report remaining performance obligations, or RPO — the contracted transaction price allocated to work not yet delivered. It is the disciplined version of “backlog,” and it is the number the bull case rests on.
CoreWeave disclosed $103.7 billion of unsatisfied RPO at June 30, 2026. Nebius disclosed $37.49 billion. Set against each company’s own second-quarter revenue run-rate, that is roughly 10 years of work at CoreWeave’s current pace and roughly 16 years at Nebius’s. Those are extraordinary coverage ratios for any industry.
The schedules are where it gets more textured. Both companies disclose how the backlog is expected to convert, and in both cases the majority lands beyond two years out. CoreWeave’s tail runs all the way to month 78 — six and a half years.
Unsatisfied remaining performance obligations at June 30, 2026, split by each company’s own disclosed recognition schedule. Bars are drawn to a common dollar scale, so the size difference between the two is real.
Turn the near-dated slice into a delivery requirement and the ambition becomes concrete. CoreWeave expects 41% of $103.7 billion — about $42.5 billion — to be recognised over the 24 months to June 2028. That is roughly $21.3 billion a year, against a current run-rate near $10.3 billion. Nebius’s near-dated 36% works out to about $6.7 billion a year against a $2.33 billion run-rate, close to a threefold increase. Neither company has to win new business to hit those numbers. Both have to build, power and staff an enormous amount of capacity on schedule.
3. The concentration did not fall. It redistributed.
The most-repeated good-news story about CoreWeave over the past year is that it has diversified away from a single dominant customer. The filing supports the first half of that claim and complicates the second.
Customer A was 71% of revenue in the second quarter of 2025. In the second quarter of 2026 it was 36% — a genuine and large reduction in single-name dependence. But Customer B, which did not reach the 10% disclosure threshold a year ago, is now 26%, and Customer C is 10%. Add the disclosed names together and they were 71% of revenue then and 72% now.
Second quarter, both years. Customers below the 10% disclosure threshold are not broken out and are excluded here, so the true totals are at least this concentrated.
Receivables tell the same story with less ambiguity: Customers A and B were 32% and 32% of net accounts receivable at June 30 — 64% of the money owed to the company sitting with two counterparties. Three customers failing is a different risk profile from one customer failing, and it is a real improvement. It is not the same as diversification.
One caveat the company itself flags and I will repeat: the letters A through D may refer to different customers between periods. So the year-over-year comparison for any single letter is not necessarily the same buyer.
4. What RPO is not
RPO is a stricter measure than a press-release backlog number, but it is not a receivable and it is not cash. CoreWeave’s own definition is worth reading closely: the figure is stated net of estimated variable consideration, which the company says “primarily consists of potential reductions to the transaction price in the future, such as estimates of future potential credits to customers under availability of service agreements, amounts that may not be recognized as revenue due to delivery delays, and estimates of committed cloud computing capacity that the Company has the right to resell.”
In other words, embedded in the $103.7 billion are management’s estimates of service credits it may owe, revenue it may lose to its own delivery delays, and capacity it may resell. Those estimates are made, per the filing, “based on both historical experience and the specific facts and circumstances of the committed contracts.” That is a normal and appropriate accounting construction. It also means the headline is not a fixed quantity of guaranteed future cash, and the delivery-delay clause is a direct acknowledgement that the schedule can slip.
Nebius adds a different qualifier: its $37.49 billion excludes performance obligations with an original duration of one year or less, so short contracts do not inflate it — a conservative choice that makes the number cleaner but not comparable to CoreWeave’s on a like-for-like basis.
| Measure | CoreWeave | Nebius |
|---|---|---|
| Q2 2026 revenue | $2,575m | $582.3m |
| Year-over-year growth | +112% | +454% |
| Operating loss, Q2 | $(49)m | $(175.9)m |
| Net interest expense, Q2 | $(640)m | $(119.1)m |
| Net loss, Q2 | $(626)m | $(190.4)m |
| Unsatisfied RPO | $103.7bn | $37.49bn |
| RPO ÷ annualised Q2 revenue | 10.1× | 16.1× |
| RPO due within 24 months | 41% | 36% |
| Implied run-rate increase to deliver it | 2.1× | 2.9× |
| Deferred revenue on balance sheet | $9.7bn | — |
5. The case for the other side
Interest against an asset, not a hole
The debt bought GPUs that are generating $2.58 billion a quarter and are contracted years forward. This is closer to how a utility or a shipping company finances hard assets than to a cash-burning software startup. Judging it by net loss alone misreads the structure.
Operating breakeven at scale is the milestone
A $49 million operating loss on $2.58 billion of revenue, from a $19 million operating profit a year earlier at a fifth the size, means the core service is roughly covering its own costs while doubling. If that holds as capex moderates, the interest burden becomes a refinancing question rather than a solvency one.
Prepayment is real money
CoreWeave carries $9.7 billion of deferred revenue and Nebius took in $4.4 billion of additional customer prepayments in the first half. Customers are funding capacity in advance. That is a materially stronger commitment signal than a signed contract alone.
6. What this means for the book
I own both of these companies, in small size, and one of them twice.
| Position | Weight | Conviction | Note |
|---|---|---|---|
| GraniteShares 2× NBIS (NBIL) | 2.06% | Low | Leveraged; roughly 2× daily exposure to Nebius |
| Nebius (NBIS) | 1.24% | Medium | Direct position |
| CoreWeave (CRWV) | 1.17% | Low | Currently below cost |
Roughly 4.5% of the book sits in these three lines, and the economic exposure to Nebius is closer to 5.4% once the 2× fund is counted properly — the double-exposure pattern my own Risk X-Ray exists to flag. What this work changes for me is not the size of those positions but the question I track them on. I have been watching revenue growth and backlog announcements. The filings say the more informative series is interest expense against operating income, and the pace at which the near-dated backlog actually converts.
What I am not claiming
Nothing here suggests either company has misstated anything. RPO is a required disclosure computed under a defined standard, the customer-concentration tables are exactly where they belong, and both companies disclosed the qualifiers I have quoted. I am not forecasting a default, a covenant breach, or a backlog cancellation, and I have not modelled the debt maturity schedule or the terms of the facilities — which matter enormously and which I have not done the work on. The narrow claim is this: the number the market quotes for these companies is the backlog, and the number that currently determines whether they make money is the interest bill.
What would change my read
WatchFour things, in order of how much they would move me
1. Interest expense against operating income. If operating income crosses into consistent profit while interest flattens, the structure works. If interest keeps compounding faster than operating income improves, it does not. 2. RPO conversion versus schedule. CoreWeave has told us it expects $42.5 billion within 24 months. The next four quarters are a measurable test of that, not a narrative. 3. Whether the disclosed-customer total falls below roughly 70%. Redistribution among three large buyers is progress; genuine diversification would show up as that total declining. 4. Any downward revision to RPO. Because the figure is stated net of estimated variable consideration, a reduction could come from credits, delays or resale assumptions rather than from a cancellation — and would be worth reading carefully rather than reacting to.
The demand side of the AI trade has been examined here before, in The AI Capex Reality Check and Data Center Alley. What this quarter’s filings add is the other half of the sentence. The contracts exist. The capacity is being built. The question the income statement is actually asking is what it costs to hold the assets in between — and for now, at CoreWeave, that cost is thirteen times the operating loss it is set against.