CoreWeave's Business Model Explained: How a Neocloud Makes Money

CoreWeave makes money by renting clusters of Nvidia GPUs to AI labs and hyperscalers, mostly on multi-year take-or-pay contracts. It signs a contract with a creditworthy buyer, borrows against that contract and the chips it will buy, leases a powered building, installs the cluster, and repays the debt from contract cash flows. That model produces an EBITDA margin of about 60% but, after depreciation and interest, losses, which is why CoreWeave is better understood as a leasing company with a technology story than as a software firm.

The market's first verdict was mixed. On 27 March 2025 CoreWeave priced its IPO at $40 a share after hoping for $47–55, raising about $1.5 billion with Nvidia reportedly anchoring the deal with an order of about $250 million (reported). Its prospectus showed revenue up from about $229 million in 2023 to about $1.92 billion in 2024, a loss of about $863 million, and about 62% of revenue from one customer, Microsoft. The shares reached an intraday high of about $187 in June 2025, then fell back to roughly $60–80 by October 2026 after a stock-and-bond sell-off (market data, reported).

What is a neocloud?

A neocloud rents large clusters of the newest accelerators, usually Nvidia's, on contracts of two to five years, to model labs and hyperscalers. CoreWeave, Nebius, Lambda and Crusoe are the best-known. Unlike AWS or Azure, they offer few other services; unlike a colocation landlord, they own the chips.

Neoclouds exist because of a specific moment. In 2023 Nvidia was short of H100s and chose to allocate them partly to specialists that would never design chips of their own, rather than only to hyperscalers that were building rivals. At the same time Microsoft needed capacity for OpenAI faster than Azure could build it. CoreWeave had the chips, the operating skills and a way to finance them.

Where did CoreWeave come from?

CoreWeave began in 2017 as Atlantic Crypto, an Ethereum-mining operation run by three former commodity traders, Michael Intrator, Brian Venturo and Brannin McBee. When Ethereum's price collapsed in 2018, they had to decide whether to sell their GPUs or find something else for them to do. They rented them out, to visual-effects studios, rendering and scientific computing, and renamed the company around 2019. Mining had taught them to buy GPUs at scale, find cheap power and keep large fleets running. In April 2023 Nvidia joined a $221 million round led by Magnetar (reported), and the company was positioned when H100s became the scarcest product in technology.

How does CoreWeave make money?

The research summarizes the model in one formula: contract revenue × utilization − power and leases − operations − depreciation − interest. Its business lines are simple.

CoreWeave mostly does not own its buildings. It leases powered shells from developers and former crypto miners on 10–15-year leases, then fills them with servers it buys. Operating lease liabilities were about $16.3 billion at 30 June 2026, on top of its borrowings.

Why does a 60% EBITDA margin still lose money?

At the operating level the model looks strong. In 2025 CoreWeave reported about $5.13 billion of revenue and adjusted EBITDA of about $3.09 billion, about 60% (company). Below EBITDA it is thin or negative. GPUs are written off over six years under CoreWeave's policy, and its high-yield debt is expensive: net interest was about $536 million in the first quarter of 2026 and about $640 million in the second. The result was GAAP net losses of about $1.17 billion in 2025 and about $626 million in the second quarter of 2026 alone.

Guidance for 2026 was about $12.4–13.2 billion of revenue and about $0.96–1.18 billion of adjusted operating income, under 10% (company, August 2026). Capex guidance for 2026 rose to $35–39 billion. The company is growing fast, and almost all of the growth is financed.

What are the three clocks in a GPU cloud?

The book frames neocloud risk as a mismatch between three clocks: the contract term, the GPU's economic life and the debt's maturity. A stylized example in Chapter 6, not based on CoreWeave's figures, isolates the contract clock. A cluster of 20,000 GPUs costs $45,000 each, $900 million in all, financed 75% with debt at 10% over five years. A customer pays $2.10 per GPU-hour, and operations cost $0.30.

With a five-year contract matched to five-year debt, the $225 million of equity receives about $699 million of pre-tax cash and still owns the GPUs. Shorten the contract to three years and re-rent the chips at a price that falls when the next generation arrives, and the equity's five-year cash drops to about $209 million, less than it put in. Nothing about the chips changed. The only difference is who bears the price of a GPU in its fourth year.

CoreWeave's early debt facilities amortized over the life of each contract, which matched the clocks. Its later unsecured notes, at coupons of about 8.5–9.6% due 2030–32, and convertible notes do not, and must be refinanced or repaid from equity value. The depreciation question matters too: the investor Michael Burry argued in November 2025 that the useful lives the big clouds assign to servers are too long (reported), and Nvidia now releases a new architecture every year.

What are CoreWeave's biggest risks?

Customer concentration

Microsoft accounted for about 35% of revenue in 2023, about 62% in 2024 and about 67% in 2025 (company filings). Diversification went toward other very large, capital-hungry buyers rather than many small ones. OpenAI signed contracts worth about $11.9 billion, $4 billion and $6.5 billion in 2025; Meta up to $14.2 billion in September 2025 and about $21 billion more in April 2026; Anthropic a multi-year deal in April 2026. Several of these customers are themselves loss-making and raising money to pay. The research's first question for any cloud is whose credit stands behind the backlog.

Delivery

Backlog becomes revenue only if capacity is delivered on time. On 10 November 2025 CoreWeave cut guidance because one third-party developer delivered a powered shell late. Contracts typically let customers terminate or claim credits for missed deliveries (S-1 risk factors). An attempt to buy its landlord, Core Scientific, for about $9 billion in stock was voted down by Core Scientific's shareholders on 30 October 2025.

Supplier dependence

Nearly all of CoreWeave's compute is Nvidia's. The relationship gives it early access to each generation, and Nvidia has become supplier, shareholder and backstop at once: the $6.3 billion residual-capacity order in September 2025 and a $2 billion share purchase at $87.20 on 26 January 2026. Such backstops make the neocloud's credit look like Nvidia's, which is the point, but they also show how much of the model depends on one company's support. Our article on why Nvidia dominates AI chips explains why Nvidia values buyers like this.

How does CoreWeave compare with hyperscalers?

Hyperscalers rent the same chips but fund them from enormous existing cash flows and sell many other services. AWS earned an operating margin of about 38.7% in the third quarter of 2026. Microsoft itself leases capacity from neoclouds, including Nebius (about $17.4 billion over five years, September 2025), because leasing shifts construction and financing risk to suppliers. A neocloud's edge is speed and Nvidia's favor; its weakness is that it rents the same chips as everyone else, financed with debt. The cloud, data centers and compute infrastructure study compares the two models.

Which factor of the intelligence equation decides a neocloud?

The Intelligence Equation argues that durable value needs four factors at once: capability, unit economics, a defensible position and the capital and permission to last the cycle. The cloud layer turns on the fourth. CoreWeave had capability and, for a time, a position Nvidia granted it. Its returns depend on matching borrowed money to contracts with a handful of customers who are themselves raising money to pay. Where the capital clock and the contract clock line up, the business works like a lease. Where they do not, the value goes to whoever is left holding the chips. For how this layer fits the wider money flow, see our guide to how AI companies make money.

The takeaway

CoreWeave's business model is GPU leasing financed with debt: it rents Nvidia clusters to a few very large buyers on multi-year contracts, earns about 60% EBITDA margins, and loses money after depreciation and interest. Its value depends on three clocks lining up, the contract term, the GPU's useful life and the debt's maturity, and on the credit of customers who are themselves burning cash. Backlog of about $104 billion is real only if data centers arrive on time and customers keep paying. Chapter 6 of the book, Clouds and Data Centres, sets out the full analysis, and the CoreWeave company profile follows each move.