AI-generated
Moatsight's own AI workflow wrote this edition, and Moatsight corrected it before it was published.
- Written Nov 11, 2026 from 20 sources, including the earnings release (8-K) filed Nov 10, 2026.
- Argued by a bear and a bull critic.
Business explainer
Loud claims, checked: 2 partly supported · 3 open
| Verdict | Claim | Who and when |
|---|---|---|
| OpenThe 10-K confirms the six-year life and its sensitivity; no filing shows whether six years is right | Long useful lives for AI servers overstate earnings, and CoreWeave's six-year GPU life is too longIn What changed | Michael Burry, post of November 10, 2025 reported by CNBC on November 11, 2025, since applied widely to CoreWeave |
| Partly supportedSupported for the cost of parent-level unsecured debt; the company still raised convertible notes and a term loan in Q3 2026 | CoreWeave's funding is breaking down as credit markets tightenIn What changed | Market commentary reported by CNBC on November 6 and November 9, 2026, including Paul Meeks of Freedom Capital Markets |
| OpenMicrosoft was about 67% of 2025 revenue; existing contracts are take-or-pay with fixed terms | Microsoft's move to fewer neocloud partners threatens CoreWeave's revenueIn What changed | CNBC, November 6, 2026, citing people familiar with the matter |
| Open | Most neocloud providers will not survive as independent companiesIn What changed | Gartner forecast of October 2026, reported by CNBC on November 4, 2026 |
| Partly supportedPlausible if capex per dollar of revenue is near 2.5 or below; the backlog still needs large capex to deliver | Contracted cash flows from the backlog will service and repay all of CoreWeave's debtIn What changed | CoreWeave's CFO, third-quarter 2026 earnings call, November 10, 2026 |
A verdict is the Moatsight view of what the filings show, not a finding of fact.
The core#
CoreWeave rents out large clusters of NVIDIA graphics processors (GPUs) to companies that train and run artificial-intelligence models, mostly under multi-year take-or-pay contracts signed before the capacity is built. It is a capital-heavy business that makes money on a spread: the contracted price per year must cover the data-centre rent and power, pay for servers that are depreciated over six years, pay interest on the debt that bought them, and still leave something for shareholders. Demand is not the open question: the revenue backlog stood at about $132.5 billion on September 30, 2026, 13 times trailing revenue. The open questions are what each dollar of capacity costs, how long a GPU keeps earning after its first contract, and whether lenders keep funding the build-out at rates the spread can bear. Our arithmetic on a single contract shows how thin that spread is when capacity costs more than about 2.5 times its annual revenue and the GPU has no second life.
How the money comes in#
CoreWeave sells computing capacity, not chips. Customers pay for access to GPU clusters, with the storage, networking and software needed to run them, inside data centres that CoreWeave mostly leases from third-party developers. The company reports a single operating segment and no split of revenue by product.
- Committed contracts are the business. The FY2025 10-K describes them as typically take-or-pay, "requiring payment regardless of the level of utilization", with an initial term that "generally ranges from one to six years", recognised evenly over the term. Revenue from customer commitments was 98% of revenue in 2025, 96% in 2024 and 88% in 2023. The weighted-average duration of committed contracts was about five years at the end of 2025.
- Prepayments are a second source of funding. Customers often pay part of the contract in advance, booked as deferred revenue: $10.0 billion at September 30, 2026 ($2,269 million current, $7,733 million non-current). The Q3 2026 release says about $9 billion of further prepayments and deposits are expected in the three months after the quarter.
- On-demand capacity is billed monthly on hourly use. It is small in revenue but it sets the market price that renewals are signed against.
- Software and managed services sit on top: Kubernetes and Slurm orchestration (SUNK), Mission Control, Weights & Biases (bought in May 2025 for about $1.03 billion), and the Forge development environment launched in the third quarter of 2026. The company does not report their revenue separately.
How to read this: "backlog" is the company's own measure. It is remaining performance obligations ($132.3 billion at September 30, 2026, per the Q3 2026 earnings presentation) plus about $0.2 billion of other amounts the company expects to recognise under committed contracts, "subject to the satisfaction of delivery and availability of service requirements". It is revenue that still needs capacity to be built, powered and delivered. At September 30, 2026, the company expected 36% of the backlog to be recognised within 24 months, 38% in months 25 to 48 and 26% after 48 months, against 40%, 39% and 21% a year earlier (Q3 2026 earnings presentation). The backlog is getting longer, so more of it rests on customers that are still paying years from now.
Customers: still few, and large#
| Period | Largest customer, share of revenue | Other customers at 10% or more |
|---|---|---|
| FY 2023 | 35% | two others; top three 73% |
| FY 2024 | 62% | one other at 15% |
| FY 2025 | 67% (Microsoft, named in the 10-K) | none |
| Q2 2026 | 36% | 26% and 10% |
| H1 2026 | 40% | 23% |
The 10-K names Microsoft as the 67% customer for 2025; the 10-Q labels customers by letter only, and the letters "may represent different customers than in a previous period". The concentration is falling as other contracts start to bill, but three customers still made up about 72% of second-quarter revenue on our sum of the 10-Q figures. The large contracts named in the filings and releases are:
- OpenAI: an order form of September 2025 for up to about $6.5 billion through May 2031 (10-K), on top of earlier contracts announced in 2025.
- Meta: an order form of September 2025 for up to about $14.2 billion through December 2031 (10-K), and a further order form for up to about $21.0 billion announced in April 2026 (Q2 2026 10-Q).
- Others: Anthropic (April 2026), Jane Street, Cohere, Perplexity and others named in the Q3 2026 release. On the first-quarter 2026 call the company said ten customers had committed at least $1 billion each.
On the third-quarter 2025 call (November 10, 2025) the CFO said no single customer was more than about 35% of backlog, down from about 85% at the start of 2025, and that more than 60% of backlog was tied to investment-grade customers. On the first-quarter 2026 call he said non-investment-grade AI companies and model labs were less than 30% of backlog. The Q3 2026 release gives no update on either figure.
NVIDIA is on every side of the business: the sole GPU supplier, an investor (it bought about 23 million shares for $2.0 billion at $87.20 a share in January 2026, per the Q2 2026 10-Q) and a customer under a master services agreement filed with the 10-K.
The market and CoreWeave's share of it#
Synergy Research Group estimated on April 2, 2026 that revenue of "neocloud" providers (specialist GPU clouds such as CoreWeave) exceeded $25 billion in 2025, with $9 billion in the fourth quarter alone, up 223% year on year, and forecast the market to approach $400 billion by 2031, a 58% compound annual growth rate. CoreWeave's 2025 revenue of $5.1 billion was at most about 20% of that market, on our calculation.
The company's FY2026 revenue guidance of $13.0 to $13.3 billion implies growth of about 156% at the midpoint, which we calculate against 2025's $5,131 million. That is fast, but in a market that grew more than threefold in the fourth quarter of 2025, it does not show share gains on its own. It shows CoreWeave keeping pace in a market whose growth is set by how fast power, buildings and chips can be delivered. Management's long-term aims are stated in power, not market share: on the August 11, 2026 call the CEO spoke of "at least 8 gigawatts" of active power by 2030, against more than 1.7 GW active and more than 4.6 GW contracted at September 30, 2026.
Unit economics: one contract, three outcomes#
The filings do not give the economics of a single contract, so we model one from what they and the calls do give. Revenue is 1 a year for five years. Operating costs other than depreciation (rent, power, staff, overhead) run at 36% of revenue, consistent with the 62% adjusted EBITDA margin in Q3 2026. Capex is paid up front, at a multiple of annual revenue. On the second-quarter 2026 call the company described contracts that repay the asset-level debt and leave free cash flow, but it gave no capex multiple. Our Q3 2026 figures imply about 2.7 times: FY2026 capex guidance of $37 to $39 billion against the rise in annualised revenue from the fourth quarter of 2025 to the $20 to $21 billion exit-2026 guidance.
| Capex per $1 of annual revenue | Extra years after the contract | Asset IRR, no prepayment | Asset IRR, 15% prepaid | Equity cash per $1 of annual revenue, 80% debt at 7.5% (NPV at 16%) |
|---|---|---|---|---|
| 2.2 | 0 | 13.9% | 20.5% | 0.48 |
| 2.2 | 2 at half price | 17.0% | 24.0% | 0.64 |
| 2.5 | 0 | 8.8% | 12.4% | 0.22 |
| 2.5 | 2 at half price | 12.2% | 16.4% | 0.38 |
| 2.8 | 0 | 4.6% | 6.3% | -0.03 |
| 2.8 | 2 at half price | 8.2% | 10.6% | 0.13 |
Three things decide the outcome, and the headline numbers show none of them:
- Capex per dollar of revenue. At 2.8 times, a five-year contract returns less than the company's 7.9% weighted cost of debt (CEO, November 10, 2026 call). At 2.2 times it is a good business. The third-quarter price rises the CEO described on the November 10, 2026 call (about 25% across products in July and another 10% later in the quarter) move new contracts towards the better rows.
- Second life. The extra years after the first contract are where most of the equity's return sits. Management says older GPUs re-contract well. On the November 10, 2026 call the CEO said prices for older Ampere and Hopper GPUs rose 30% to 50% quarter on quarter, to levels above their original contracts. On the second-quarter 2026 call the CFO described an A100 contract running into 2029, for a chip introduced in 2020.
- Prepayments and debt. Customer prepayments lift the asset return by three to seven points in our table. Debt at 80% of capex turns a thin asset return into a positive equity cash flow, but only while lenders keep lending at around 7.5%.
What moves the results#
| Driver | Where it shows | Direction of effect |
|---|---|---|
| Capacity delivered (active MW) | Revenue, with a lag of one to two months while a site ramps | More MW, more revenue; delays push revenue and margin out |
| Price per GPU-hour on new and renewed contracts | Revenue per MW, adjusted operating margin | Higher prices lift margin on the same capex |
| Capex per MW (GPUs, networking, memory) | Depreciation, debt needed | Component inflation raises it; the CEO said contracts are priced "with purchase orders in hand" |
| Useful life of GPUs (six years in the books) | Depreciation, 51.6% of revenue in Q3 2026 | A shorter life raises depreciation |
| Cost and volume of debt | Interest expense, 23.1% of revenue in Q3 2026 | Each 100 basis points on floating debt moved interest by about $30 million a quarter at June 30, 2026 (Q2 2026 10-Q) |
| Rent and power of leased sites | Cost of revenue, 32.3% of revenue in Q3 2026 | Leases are long and fixed; idle capacity costs money |
The ramp effect matters for reading any single quarter. On the first-quarter 2026 call the CFO said new deployments run at a negative contribution margin for "one to two months" while they are fitted out, before normalising "in the mid-20s". When a lot of capacity arrives at once, as in early 2026, margins fall first and recover later.
The moat, and why customers choose CoreWeave#
The 10-K says CoreWeave competes mainly with hyperscalers, "several of which are also customers of, and partners to, CoreWeave", and with smaller GPU clouds. It claims speed to market with each new NVIDIA generation, its own orchestration software, security standards and total cost of ownership. Outside evidence supports the speed claim. The Q3 2026 release reports multi-rack deployments of NVIDIA's Vera Rubin systems in production. It also says CoreWeave is "the only provider to earn SemiAnalysis' Platinum ClusterMAX rating three consecutive times".
The real advantages, as we read them, are three:
- Execution at scale. CoreWeave can turn a signed contract into a powered, networked cluster of tens of thousands of GPUs faster than most buyers can themselves.
- A financing machine. Its delayed-draw term loans (DDTLs) are secured on contracted cash flows and lend against the GPUs. DDTL 4.0, closed in early 2026, carries an investment-grade rating, which the CFO described on the first-quarter 2026 call as A-minus. Its fixed tranche cost about 5.9% at the first draw (Q1 2026 release).
- A pipeline of contracted power. It had more than 4.6 GW contracted at September 30, 2026, at a time when power is the binding constraint.
None of these is a patent or a network effect. Each can be copied by a customer with a larger balance sheet, which is what the market fears.
Honest weaknesses#
- No product of its own in the core input. All GPUs are NVIDIA's, and the 10-K says customers often specify them. Gross margin depends on NVIDIA's pricing as much as on CoreWeave's.
- Customers that can leave. The largest customers are also the largest builders of their own data centres. CNBC reported on November 6, 2026, citing people familiar with the matter, that Microsoft is "shrinking its roster of neocloud partners" as it builds out its own data centres. The report did not say what this means for CoreWeave's existing contracts, which run to fixed end dates.
- A thin equity cushion. Total debt was $46.9 billion at September 30, 2026, against stockholders' equity of $4.0 billion, plus $18.7 billion of operating lease liabilities.
- Controls still being built. The 10-K and the Q2 2026 10-Q report material weaknesses in internal control over financial reporting, first identified in 2024, and disclosure controls that were not effective.
- Founders' control. Class B shares carry ten votes, and the co-founders control the outcome of shareholder votes.
Risks the filings name, in plain words#
- One or two customers can move the results (10-K): a cut in spending by a top customer "would adversely affect" the business.
- The company needs ever more capital (10-K): it says its ability to lower its cost of capital "depends upon a number of factors, many of which are beyond our control".
- Construction and power delays at partners' sites. The 10-K cites the delays the company announced in November 2025, caused by one third-party developer.
- GPU useful life (10-K): changes to useful-life estimates, or "any inability to redeploy components of our existing infrastructure to extend past their contracted life", could "significantly affect" results.
- AI demand itself. Efficiency gains in AI models, or a plateau in the benefit of more computing, could reduce demand for centralised capacity.
- Litigation. The 10-K and the Q2 2026 10-Q describe a putative securities class action filed on January 12, 2026 in the US District Court for the District of New Jersey, and related stockholder derivative actions, which are stayed. The company says it believes the claims are without merit; no loss is accrued.
What we do not know#
- The capex per megawatt and per dollar of revenue of the contracts signed in 2026. The company has not disclosed it, and it decides the equity's return.
- How much of the $132.5 billion backlog each of the largest customers accounts for. The company gave shares on calls in 2025 but not for September 30, 2026.
- The price at which GPUs re-contract after their first term, beyond the examples management has given on calls.
- The third-quarter 10-Q, which had not been filed when this edition was written. Customer concentration and the debt schedule for September 30, 2026 are therefore taken from June 30, 2026, and the backlog's timing from the Q3 2026 presentation.