CoreWeave has not funded a dollar of its GPUs from operations
A case study of the largest GPU rental business: what it earns, what it owes, and what the credit market is really pricing.
Everything here comes from filed documents. CoreWeave's FY2025 Form 10-K and its Q2 2026 Form 10-Q, both filed with the SEC. Market prices are from Bloomberg and I give the date on each one. Where I have not established something, I say so rather than estimating it.
Since the start of 2023, CoreWeave has generated $11.3bn of cash from operations, after paying $1.9bn of cash interest. Over the same period it repaid $9.2bn of debt principal, which leaves $2.1bn. And it spent $36.1bn on property and equipment.
So 94% of the hardware was paid for with money the business did not earn. Gross borrowing over the same period was $37.0bn, which is within 2.6% of the capex, and there was $4.5bn of equity on top.
Source: FY2025 Form 10-K and Q2 2026 Form 10-Q, consolidated statements of cash flows. These are gross cash movements over the period, not a sources and uses statement, so they do not sum. Cash interest paid sits inside cash from operations and is not a separate bar.
That is not a scandal and it is not a trick. It is what an infrastructure build looks like while it is being built. But it means the business has never once paid for its own growth, and the question that follows is how long it has to keep not doing so, and what happens the day it cannot.
What CoreWeave sells
Almost all of it is contracted. Committed contracts were 98% of revenue in the second quarter of 2026 and 98% in the second quarter of 2025. Consensus has committed revenue at $12.8bn of the $12.9bn expected this year. The on-demand business is a rounding error.
So this is not a cloud company in the utility sense. It is closer to a leasing business. Customers sign for dedicated GPU capacity over several years, prepay part of it, and CoreWeave buys the hardware and the power to deliver it.
The scale moved quickly. At the end of 2023 it ran 10 data centres and about 70MW of active power. At the end of 2024, 32 sites and 360MW. At the end of 2025, 43 sites, over 850MW, 250,000 GPUs online and $5.1bn of revenue.
Work the revenue backwards to a rate per GPU-hour and you get something you can test against the market:
| Year | Revenue | Average GPUs online | At 100% utilisation | At 85% |
|---|---|---|---|---|
| 2024 | $1,915m | 65,500 | $3.34 | $3.93 |
| 2025 | $5,131m | 178,000 | $3.29 | $3.87 |
| 2026 consensus | $12,895m | 580,896 | $2.53 | $2.98 |
Average GPUs is the mean of opening and closing counts. The closing figure for 2025 is the 250,000 disclosed in the 10-K. The 2026 figures are consensus, not actuals. The 2024 opening count is scaled from active power and is the softest input on this page.
Utilisation is never disclosed, which is why the rate is given at two assumptions rather than one.
Two things fall out of that table. The realised rate sits well above the hourly market, because the book was sold as multi-year committed capacity rather than at spot. And the forecast already has it converging on spot, from $3.87 to $2.98 at the same utilisation. That is a 23% fall in revenue per unit of hardware, and it is in the consensus, not in anyone's bear case.
For reference, the Silicon Data H100 neocloud index was $2.85 per GPU-hour on 9 October 2026. B200 was $5.93. A100 was $1.53.
Where the cash goes
| $m | 2023 | 2024 | 2025 | H1 2026 | Cumulative |
|---|---|---|---|---|---|
| Operating cash flow | 1,833 | 2,749 | 3,058 | 3,663 | 11,303 |
| Capex, cash | (2,943) | (8,702) | (10,309) | (14,117) | (36,071) |
| Debt issued, net of fees | 1,404 | 7,018 | 11,829 | 16,747 | 36,998 |
| Debt repaid | (2) | (589) | (3,399) | (5,219) | (9,209) |
| Equity issued | 0 | 0 | 1,491 | 2,982 | 4,473 |
| Cash interest paid | (0) | (184) | (869) | (806) | (1,859) |
Cash interest is shown for reference and is not an additional outflow. It is a supplemental disclosure at the foot of the cash flow statement and is already deducted inside the operating line above it, because the statement starts from net loss.
Read the cumulative column as a single sentence and you have the business. Eleven billion earned, nine billion of it straight back out to repay principal, thirty-six billion spent on hardware, and thirty-seven billion borrowed to do it.
- Operating cash flow
- Capex
- Debt raised
- Debt repaid
| Period | Operating cash flow, $m | Capex, $m | Debt raised, $m | Debt repaid, $m |
|---|---|---|---|---|
| 2023 | 1,833 | 2,943 | 1,404 | 2 |
| 2024 | 2,749 | 8,702 | 7,018 | 589 |
| 2025 | 3,058 | 10,309 | 11,829 | 3,399 |
| H1 2026 (six months) | 3,663 | 14,117 | 16,747 | 5,219 |
Source: FY2025 10-K and Q2 2026 10-Q statements of cash flows. H1 2026 is six months and is not annualised, so the final group is understated against the full years beside it.
In the first six months of 2026 alone, CoreWeave raised $16.7bn of gross debt. That is more than four times the operating cash flow it generated in the same period, and more than the entire operating cash flow of the company's history.
Most of that operating cash is not earnings
| $m | 2023 | 2024 | 2025 | H1 2026 | Cumulative |
|---|---|---|---|---|---|
| Operating cash flow | 1,833 | 2,749 | 3,058 | 3,663 | 11,303 |
| of which, deferred revenue inflow | 1,986 | 2,049 | 4,174 | 1,365 | 9,574 |
| Operating cash flow excluding it | (153) | 700 | (1,116) | 2,298 | 1,729 |
Eighty-five percent of all the operating cash this business has ever produced is customers paying in advance. Strip it out and the figure is $1.7bn over three and a half years, against $36.1bn of hardware.
Deferred revenue stood at $9.7bn on 30 June. This is a real source of cash and a perfectly normal feature of a contracted business. It has two properties that matter here. It is a liability that has to be worked off by delivering service. And it only keeps contributing cash while the contracted book is growing.
In the first half of 2026, only $554m of the $8.2bn opening balance was recognised as revenue. So this is overwhelmingly long-dated prepayment, not near-term unearned income. The day bookings stop growing, this line goes to zero and then turns negative, and it takes most of the operating cash flow with it.
That is the first half of why the business cannot fund itself. The second half is that the obligations it has signed run on four different clocks.
Four clocks
Source: Q2 2026 10-Q Note 5, Note 10 and the lease note; useful lives from the FY2025 10-K. Bars show contractual length, not value. The lease bars are undiscounted payment periods, not present values.
| Clock | Length |
|---|---|
| GPU fleet in service | 6-year accounting life, 1.1 years weighted average age |
| Customer contracts | $103.7bn of RPO, 41% in 24 months, 39% in months 25 to 48, the rest to month 78 |
| Debt | $35.6bn of principal, final maturities 2030 to 2033 |
| Leases, commenced | $29.1bn undiscounted, 12-year weighted average remaining term |
| Leases, not yet commenced | $35.5bn undiscounted, terms of 7 to 16 years, commencing 2026 to 2029 |
The fleet age is mine, estimated from year-end gross technology equipment cohorts. Everything else is disclosed.
The order matters. The building commitment is the longest and the hardest to cancel. The hardware inside it is the shortest-lived. The customer contract that pays for both sits in between and ends years before the lease does. And there is a further $35.5bn of lease obligations already signed but not yet started, most of it without customer contracts attached, plus a single site contractually capped at $14.7bn over sixteen years that is excluded from that figure.
This mismatch is not an accident. It is what the business is. Nobody will sell you a twelve-year power and cooling envelope on a three-year commitment, and nobody will buy a twelve-year contract on hardware that is obsolete in six. Every operator in this sector carries a version of it. The only question is whether the contracted book keeps refilling the gap, and at what price.
The schedule almost nobody reads
The common view is that CoreWeave has no near-term maturities, because the bonds mature between 2030 and 2033 and the term loans run to 2032. I held that view myself until I read Note 10.
| Years ending 31 December | Debt principal | Lease payments |
|---|---|---|
| Remaining 2026 | 4,413 | 1,028 |
| 2027 | 6,184 | 2,116 |
| 2028 | 4,416 | 2,306 |
| 2029 | 2,421 | 2,373 |
| 2030 | 3,221 | 2,289 |
| Thereafter | 14,896 | 19,023 |
| Total | 35,551 | 29,135 |
$10.6bn of principal, 30% of all of it, falls due inside eighteen months. The delayed-draw term loans amortise monthly once drawn, which is the right structure for lending against hardware that loses value monthly, and it means the repayment profile is front-loaded even though the final maturity dates are years away.
Against that, annualised operating cash flow is about $7.3bn.
To be clear about what this does and does not mean: CoreWeave can pay it. There was $5.5bn of unrestricted cash on the balance sheet at 30 June, and $3.7bn of operating cash came in during the first half. The amortisation is not a solvency problem. It is a claim on the operating cash flow that would otherwise have gone towards capex, which means more borrowing, which is where the pressure actually sits.
What happens if the capex stops
The bull case for any capital-intensive business under stress is that the spending is discretionary. Stop building and you are left with a cash-generative installed base.
Here is why that does not work cleanly.
Capex buys power, buildings and GPUs. Capacity is what is actually sold, so new bookings require new capacity. New bookings are what grow deferred revenue, and deferred revenue is 85% of operating cash flow. So cutting capex removes new bookings, which removes the deferred revenue inflow, which removes most of operating cash flow, in the same year.
And the number that makes it concrete: consensus has $26.5bn of revenue in 2027. The contracted book at 30 June delivers $21.3bn of that. The remaining $5.3bn, a fifth of next year's revenue, has to come from contracts that did not exist at the balance sheet date. Those contracts need capacity, and that capacity needs the $47.8bn of capex consensus has pencilled in for 2027.
So capex is not growth spending here. It is the revenue line.
Two things qualify this. The lease note discloses $500m to $1.2bn of contractual obligations to procure and install equipment at leased premises through 2028, on top of the $35.5bn of signed but uncommenced leases, so a meaningful part of the spend is already committed and not discretionary at all. And the mix shift is real: B200 rents at $5.93 against H100 at $2.85, so a fleet moving to newer silicon can hold its blended rate. But buying newer silicon is capex, which brings you back to the start.
What the credit market did in July
The credit market worked all of this out in July 2026, and the way it sorted itself is the most useful piece of evidence in this entire piece.
| Instrument | Rank | 30 Jun | 9 Oct | Change |
|---|---|---|---|---|
| 9.25% Jun-2030 | Senior unsecured | 100.69 | 91.92 | -8.77 |
| 9.75% Oct-2031 | Senior unsecured | 99.83 | 85.35 | -14.49 |
| 8.50% Jul-2032 EUR | Senior unsecured | 98.70 | 83.04 | -15.66 |
| CB 1.75% Dec-2031 | Senior unsecured | 119.45 | 95.43 | -24.02 |
| DDTL 3.0 Aug-2030 | 1L, guaranteed | 101.62 | 101.62 | 0.00 |
| DDTL 5.0 Nov-2031 | 1L, guaranteed | 102.13 | 100.51 | -1.62 |
| DDTL 4.0 Mar-2032 | 1L, non-recourse | 99.88 | 100.62 | +0.75 |
Clean price per 100 of face, Bloomberg, mid.
$1,312m of market value lost on $9,930m of senior unsecured face. $40m lost on $7,690m of secured face. The non-recourse facility went up.
yield, %
| Series | From | To | First, yield, % | Last |
|---|---|---|---|---|
| Senior unsecured 9.75% 2031 (B1 / BB-) | 1 June 2026 | 9 October 2026 | 8.95 | 13.93 |
| Secured, parent guaranteed, DDTL 5.0 (Ba2 / BB+) | 1 June 2026 | 9 October 2026 | 7.64 | 8.30 |
| Secured, non-recourse, DDTL 4.0 (A3 / A-) | 1 June 2026 | 9 October 2026 | 4.81 | 4.93 |
Source: Bloomberg, daily, to 9 October 2026. These are quoted yields. The two loans are floating rate, so their yields track SOFR and should not be compared directly to a fixed-rate spread.
Over the same period, US high yield OAS moved from 270bp to 308bp and the Nasdaq 100 rose. This was CoreWeave, not the market. Regressing daily CoreWeave spread changes on daily high yield index changes through 2026 gives a beta of 2.51 and a correlation of 0.44, so the index explains about a fifth of the variance.
Why the split is so clean is in Note 10. The delayed-draw facilities sit in bankruptcy-remote special-purpose subsidiaries formed to hold the financed infrastructure and, in the filing's own words, the related customer contracts. DDTL 4.0 is non-recourse to the parent except for customary carve-outs. The lender there owns the hardware and the cash flow that services it. Moody's rates it A3. The senior unsecured notes at the parent are B1, seven notches lower.
The equity, incidentally, peaked at $137.98 on 6 May and was at $72.91 by 16 July, before any of the July news that has been blamed for the fall. It closed at $82.08 on 9 October, down 41% from the peak.
The 60% that is not a forecast
Several published pieces have taken CoreWeave's five-year CDS and converted it into a default probability of 50 to 64%. The arithmetic is right. The number is not a forecast.
A credit default swap spread converts to an implied default rate by dividing by the loss given default. Loss given default is one minus the recovery rate, and every published figure I have seen uses 40%, because 40% is the ISDA convention baked into the contract, not because anyone estimated it.
At the 9 October CMAN mid of 903bp:
| Recovery assumption | Implied annual | Five-year cumulative |
|---|---|---|
| 40%, the market convention | 15.1% | 55.8% |
| 32%, my base case | 13.2% | 50.7% |
| 16%, my liquidation case | 10.8% | 43.4% |
My recovery figures come from a waterfall built off the 30 June balance sheet, with the ring-fenced assets and the claims that attach to them separated out. The headline moves by twelve points depending on an assumption nobody in those pieces examined.
The cash bonds agree with the CDS, which matters. The 9.75% 2031 had an option-adjusted spread of 894bp on 9 October against the 903bp CDS, so this is not a derivatives artefact. The credit market really is charging something in the region of an 11 to 15% annual default rate, against 3 to 5% for B-rated US corporates through a cycle.
What that is not is a Lehman comparison. Lehman was a broker-dealer funded overnight on repo and it failed in days. CoreWeave has term debt on a published amortisation schedule, no repo book and no depositors. The failure mode is a different shape, and a spread level alone does not tell you otherwise.
What this tells you
The business is solvent and it is not self-funding, and those are different sentences.
Interest is covered 4.6 times on annualised EBITDA and there is enough cash on the balance sheet to meet every obligation due this year. What there is not is any history of the operations paying for the hardware. $2.1bn of operating cash left after repaying principal, against $36.1bn spent on property and equipment. And of the $11.3bn earned, $9.6bn was customers paying in advance, so stripping that out the business did not cover even its own debt repayments.
The risk is access, not a maturity.
Consensus needs $27.2bn of external funding this year and $30.3bn next. CoreWeave raised $16.7bn of gross debt in six months, so the plan is consistent with what it has been doing. It requires that to continue, at a price that has moved from 9.8% to 13.9% on the unsecured paper since June.
The hardware and the contracts are already pledged.
If the funding does close, the unsecured creditor at the parent is behind a ring fence that holds both the GPUs and the revenue that pays for them. That is why the secured loans did not move in July while the notes fell fifteen points, and it is why a 40% recovery assumption deserves more scrutiny than it has had.
The next thing I want to see is the third-quarter RPO figure. Not revenue, not EBITDA. If the contracted book is still growing faster than capacity is coming online, the loop holds. If it is not, the deferred revenue inflow goes first and everything else follows from there.
Figures from CoreWeave's FY2025 Form 10-K and Q2 2026 Form 10-Q. Market data from Bloomberg as at 9 October 2026 unless stated. Consensus from the Bloomberg company model. Published default probability estimates referenced above appear in TechTimes, 30 July 2026, TokenPost and Real Investment Advice, November 2025.
Nothing here is investment advice.