A fast-growing AI landlord whose debt and thin returns leave little room for error.

Return on invested capital has been below the 9% hurdle in both years it can be computed. It was low but positive in fiscal 2024 and slightly negative in fiscal 2025. Trailing twelve-month earnings to June 2026 remain loss-making. With only two usable years and no year above the hurdle, the company has not yet shown it earns more than its cost of capital.

A formal five-year incremental return cannot be computed because the history is too short. Capital employed grew very sharply: total debt rose from about $7.9 billion at the end of 2024 to $21.4 billion at the end of 2025 and $35.6 billion by June 2026. Operating income went from a small profit in 2024 to a loss in 2025. New capital has so far not produced added operating profit.

Reported earnings do not turn into owner cash. Free cash flow was negative in each of 2023, 2024 and 2025 because capital spending far exceeded operating cash flow. Net income was also negative. Operating cash flow has been supported by customer prepayments and deferred revenue, which are working-capital inflows. Capex is large and tied to contracts, but it has not yet produced free cash.

Gross margin, which excludes most depreciation, stayed in a narrow band over fiscal 2023 to 2025 and was slightly higher at the end than the start. Management says pricing on newest-generation systems is at record levels and older generations still price at or above prior levels. Independent coverage shows industry GPU-hour prices fell sharply from earlier highs, and no specific price increase is disclosed. Evidence on pricing is therefore mainly management's assertion.

CoreWeave's operating margin has not strengthened in a sustained way. Fiscal 2023 was a small operating loss on $229 million of revenue, fiscal 2024 reached an operating margin of roughly 17% ($324 million on $1.9 billion), and fiscal 2025 fell back to slightly below break-even (a $46 million operating loss on $5.1 billion). The positive three-year trend line comes from the loss-making starting point and the 2024 peak, not from steady improvement. For the latest quarter, the second-quarter 2026 earnings release (August 11, 2026) and call coverage put adjusted operating income at $128 million, a 5% adjusted margin. That is up from $21 million in the first quarter of 2026 but below $200 million a year earlier. Management calls the quarter an inflection point toward operating leverage, and the rapid capacity build-out is a disclosed investment phase, which tempers the negative read. Only three annual periods are available, not five, and the latest-quarter figures are adjusted rather than GAAP, so the direction is an inference. Margins are low and volatile, with an early sequential recovery that is not yet proven. Return on invested capital is not assessed here.

The most plausible moat is scale cost: early access to scarce GPUs, power capacity and a specialised software stack. Observed outcomes are a contracted backlog of about $104 billion, near-term capacity described as sold out, and 98% of revenue from committed contracts. Returns well above peers are not shown, since return on capital is about zero. The moat is demand-side evidence, not yet proven in profits.

CoreWeave is described by Synergy Research as the leader among neoclouds, a group growing very fast and gaining share against traditional hyperscalers. Share direction for CoreWeave itself is not published. Nebius reported faster quarterly growth in mid-2026, and several smaller entrants are growing quickly, so CoreWeave's share of the neocloud pool is not shown to be rising. Hyperscalers remain far larger in broader cloud.

Switching costs rest on contract structure: 98% of revenue comes from committed contracts, and the largest customers signed multi-year orders running to 2031 and 2032. Backlog is about $104 billion and deferred revenue is $9.7 billion. Churn and net revenue retention are not disclosed, and large customers can self-build or multi-source, so lock-in is contractual rather than proven by renewals.

Cost structure compares mixed against named peers. Operating margin averaged slightly positive over fiscal 2023 to 2025, far better than the heavily loss-making crypto-converted peers Applied Digital, IREN and Cipher, but far below Oracle's profitable diversified base. Gross margin excludes most depreciation and cannot be compared with peers. The most recent trailing results are loss-making, so no scale cost advantage shows up in margins yet.

Not applicable. CoreWeave's advantage, to the extent it has one, rests on scale, access to GPUs and power, and contracted capacity, not on patents, licences or exclusive legal rights. The filings say intellectual property protection is limited and may be challenged, so low legal protection is not a weakness here. The moat question is addressed under the moat mechanism question.

A well-funded entrant faces high capital, supply and time barriers, but recent entrants have scaled quickly rather than failed or stayed small. CoreWeave reported on its second-quarter 2026 call (August 11, 2026) that it had secured more than $32 billion of debt and equity capital to date, spent $9.4 billion of capital expenditure in that quarter alone, and grown contracted power to about 3.7 GW. An entrant would need comparable financing, scarce GPU allocation and power-ready data-centre capacity. Entrants nonetheless show momentum. Nebius reported second-quarter 2026 revenue up 454% year on year and about $5.7 billion of quarterly capital expenditure. Earlier, in late 2025, Lambda raised $1.5 billion and announced a multibillion-dollar Microsoft deployment, and Nscale announced an IPO and a $14 billion Microsoft partnership. TechInsights also notes a long tail of smaller neoclouds, including crypto miners converting to AI hosting. The barriers are real and measured in billions of dollars, but credible, well-funded entrants have been able to win large contracts. The Lambda and Nscale evidence is about nine months older than the CoreWeave and Nebius figures, and whether those entrants share customers with CoreWeave is an inference.

The named alternatives are hyperscale clouds (Amazon, Microsoft, Google) that also sell GPU and generative-AI services, and customers building their own infrastructure; the 10-K flags the self-build risk explicitly, and the June 2026 10-Q warns that general-purpose cloud offerings perceived as equivalent could cause price cuts and lost share. None of this is shown taking volume from CoreWeave. Backlog stood at about $104 billion at June 30, 2026, up 246% year on year, excluding more than $25 billion of new commitments added in the early part of the third quarter, and management said near-term capacity is effectively sold out. Price pressure is a caution: a June 2025 industry article reported neocloud GPU rental prices falling from around $8 to under $2 per hour, though that describes competition among neoclouds rather than substitution by hyperscalers or self-build. The weak point is the customer base: the largest customers are best placed to self-build or shift to a hyperscaler, and the filings do not size how much demand could move. The threat is credible but not yet shown to be gaining, and its scale is uncertain, so the view rests on inference.

AI is the demand source for the business. Management reports a backlog of about $104 billion, capacity effectively sold out, and growing managed inference and software products. The main caution is price: independent coverage shows GPU-hour prices falling sharply over time, and efficiency gains could reduce compute needed per task. Current evidence shows AI expanding the paid pool.

CoreWeave holds its customer contracts, but depends on scarce GPUs, leased data-centre space and power, and on large customers who could build their own capacity. A June 2025 industry article says neoclouds compete with each other, hyperscalers and enterprises for Nvidia's latest chips, which makes GPU supply the most contested input; the CoreWeave filings reviewed do not name a supplier agreement. NVIDIA is also a holder: the April 2026 proxy statement lists it with 47,213,353 shares, 10.66% in the ownership table and 3.27% of total voting power as of April 15, 2026. Data-centre space is leased, with rent and power costs in cost of revenue, and the 10-K says suitable additional or alternative space is expected to be available. In the second quarter of 2026 the company said it was further diversifying its portfolio of power and data-centre providers while contracted power grew to about 3.7 GW. Management said on the August 11, 2026 call that pricing and margins on Blackwell and Vera Rubin systems were reaching new highs. No supplier or customer is shown setting terms, but fiscal 2025 operating income was a $46 million loss on $5.1 billion of revenue, and the filings reviewed do not attribute that to any single supplier or customer. The position is a dependency with alternatives, and the likeliest bypass is a large customer building its own infrastructure.

Profit does not rest on a subsidy, tariff protection or reimbursement decision. The filings flag exposure to AI regulation, export controls, energy policy and trade policy, but none is described as dated or underpinning a large share of profit. The exposure is uncertainty on inputs and customers' regulatory environments, not dependence on a favourable policy.

Revenue is not eroding: the latest four quarters to June 2026 show revenue more than doubling. Margins are weaker, though. Both gross margin and operating margin fell over the latest four quarters against the prior four, with operating income turning negative as depreciation and new-capacity costs ramp. Management cites record pricing and is guiding to operating leverage, so the margin decline looks capacity-driven rather than demand-driven.

The market is not consolidated. Besides the hyperscalers there are CoreWeave, Nebius, Lambda, Crusoe, Nscale and a long tail of crypto-miners converting to AI hosting, with a race to add capacity. Network World describes neocloud competition on price as a race to the bottom, with GPU-hour prices falling sharply. Current demand exceeds supply and pricing on new systems is strong, so price conduct is not yet destructive across the board.

Key inputs come from a limited number of suppliers, chiefly NVIDIA for GPUs, whose own supply chain depends on TSMC. The filings say shortages could occur and that supplier disruption would affect availability and results. Pass-through of cost increases is not demonstrated in the filings, and funding relies on large debt facilities with restrictive covenants. Gross margin has held so far, but the concentration is high.

A few customers dominate revenue. One customer was 67% of 2025 revenue and 62% in 2024. In the quarter to June 2026 the top customer was 36% and the second 26%, with a third at 10%. The filings expect concentration to persist given the long-term contracts with OpenAI and Meta. A top customer well above 25% means buyers can set terms.

The company has no ten-year record, and the three years on file contain no demand-driven stress year. Revenue rose every year, from about $0.23 billion in 2023 to $1.9 billion in 2024 and $5.1 billion in 2025. Free cash flow did not fall from a positive peak because it was never positive; it grew more negative each year, from about -$1.1 billion to -$6.0 billion to -$7.3 billion, as capital spending outran operating cash flow. Operating income is the one measure that broke: a profit of about $324 million in 2024 became a loss of about $46 million in 2025, a fall of more than 100%, and it has not recovered. The 2026 quarterly reports show the loss continuing, with a second-quarter 2026 operating loss of $49 million against income of $19 million a year earlier, and a six-month 2026 loss of $193 million against $8 million a year earlier. The filings do not itemize why operating income swung to a loss while revenue more than doubled, so the cause is not established. The co-movement alone does not prove fixed-cost leverage, and a sustained revenue decline has not yet tested the business. Operating leverage can only be judged with low confidence.

The company could not get through a bad two years without outside capital. Net debt was about $30 billion at June 30, 2026, on top of $16.3 billion of operating lease liabilities and $35.5 billion of leases not yet commenced. Net debt to EBITDA was far above four times at the 2025 year end. Liquidity was $15.6 billion including undrawn facilities, but free cash flow is deeply negative and growth depends on raising debt and equity.

Revenue is concentrated by customer and, in effect, by product: all revenue comes from AI cloud services. In 2025 one customer was 67% of revenue and 68% of receivables. In the latest quarter to June 2026 the top customer fell to 36% and the second reached 26%, so concentration is easing but remains well above 30% for one customer. Profit, geography and facility splits are not disclosed in the filings.

Answer withheld: its claims were not cleared by review.

Founders control the company through a dual-class structure that gives them far more votes than their economic stake. Class B shares carry ten votes each, Class A one vote and Class C none. The April 2026 proxy statement shows the 12 executive officers and directors as a group hold 23.9 million Class A shares (5.38% of that class) and all 107.5 million Class B shares, which together give 72.32% of total voting power. By inference, that group owns roughly a quarter of all common shares outstanding, so executives and directors have a substantial economic stake and are not controlling the company without ownership. Voting power is still about three times the economic ownership, which fails the one-share-one-vote test and limits what outside owners can do. On pay, the passages quoted from the proxy show a five-times-salary stock ownership guideline for the chief executive (two times for other executives) and a clawback policy covering incentive pay after a restatement. The metrics that determine incentive pay are not shown in the proxy text reviewed, and the company's adjusted operating income, a metric it reports, adds back stock-based compensation, which is a real cost to owners. Board independence is not detailed in the passages reviewed.