The Billion-Dollar Business of Renting Nvidia GPUs

The AI boom has created a strange kind of rental business.

A company can spend billions on Nvidia-powered servers, place them in data centers, and sell other companies access to the computing capacity. Its customers do not need to own the hardware. They need enough processing power, delivered when and where their models require it.

That sounds like a straightforward cloud service. Underneath it is a race between long-term customer commitments and the enormous upfront cost of buying, powering, and connecting GPUs.

CoreWeave offers a useful case study. In 2025 it reported $5.1 billion in revenue and a $1.2 billion net loss. The same annual filing reported $60.7 billion of remaining performance obligations: contracted future work that had not yet been recognized as revenue (see Sources, [1]).

How can a business have a large order book and still lose money? The answer lies in the economics of renting AI computing capacity.

What a GPU cloud actually rents

The phrase “renting GPUs” is convenient, but the customer generally purchases access to a working computing system.

A high-performance AI cluster includes accelerators, CPUs, memory, networking, storage, software, a data-center location, electricity, and cooling. The customer cares about whether the whole system can train a model or run inference reliably. An unpowered GPU in a warehouse has no comparable value.

Specialized AI cloud providers are often called neoclouds. They focus on providing computing systems for AI workloads, sometimes with much larger commitments than a developer buying capacity by the hour.

CoreWeave says it sells access to its platform through multi-year committed contracts as well as on-demand, pay-as-you-go usage. Under its committed contracts, customers reserve a specified amount of capacity on a take-or-pay basis: payment is due under the contract even if the customer uses less than the reserved amount. Committed contracts accounted for more than 98% of its 2025 revenue (see Sources, [1]).

That structure changes the business. Instead of hoping someone will rent every available GPU tomorrow, the provider can point to contracted demand while it finances the equipment required to serve it.

The contract comes before the capacity

Imagine a customer needs a large cluster for several years. It signs a capacity agreement, perhaps with an advance payment. The cloud provider then has to procure the servers, secure space and power, connect the machines, and bring the service online.

CoreWeave says it primarily finances infrastructure development with asset-level debt supported by take-or-pay customer contracts, supplemented by corporate debt and equity. Its annual report describes a weighted-average duration of approximately five years for committed contracts at the end of 2025 (see Sources, [1]).

The basic sequence is:

1. A customer commits to future AI capacity.
2. The provider uses that commitment to help fund infrastructure.
3. It installs and operates the computing system.
4. The customer pays for access over the contract term.
5. The provider must cover equipment, facilities, electricity, operations, and financing costs.

This is an explanatory model, not a promise that every contract or loan works the same way. Actual payment schedules, deployment deadlines, contract protections, and financing terms matter.

It also explains why remaining performance obligations are not cash in the bank. The $60.7 billion figure at the end of 2025 represented future contracted revenue to be recognized as services were delivered, subject to the terms of those arrangements. Fulfilling the contracts requires spending real money along the way (see Sources, [1]).

Growth can consume cash

Buying GPUs is only the beginning of the spending.

They must be installed in facilities with adequate power, cooling, and high-speed connections. CoreWeave reported approximately $10.3 billion of net cash used in investing activities in 2025, driven by infrastructure investment that included its GPU fleet, networking, and servers. It reported $5.1 billion of revenue and a $1.2 billion net loss for that year (see Sources, [1]).

Those figures describe different things. Revenue measures services delivered under accounting rules. Net loss includes expenses and financing effects. Investing cash flow reflects large purchases needed to expand capacity. A company can increase sales quickly while still needing outside capital to build the next wave of systems.

The balance-sheet exposure grew as well. In its filing for the quarter ended June 30, 2026, CoreWeave reported $13.6 billion outstanding under delayed-draw term loan facilities and $16.6 billion of outstanding principal on notes. These are separate categories in the filing, and neither is a substitute for a full assessment of the company's liabilities or liquidity (see Sources, [2]).

The business therefore depends on timing. If a customer's service starts later than expected, a facility takes longer to energize, or financing becomes more expensive, the provider may carry costs before the associated computing capacity is earning its expected revenue.

The physical constraints behind this timing appear in Why AI Data Centers Have a Power Problem Nvidia Can't Solve. Securing a customer and securing electricity are separate achievements.

The concentration question

A long contract can make revenue more predictable. It can also make a provider dependent on a small number of buyers.

For the three months ended June 30, 2026, CoreWeave disclosed that its three largest individually reported customers accounted for 36%, 26%, and 10% of quarterly revenue, respectively (see Sources, [2]). The filing identifies them as Customer A, B, and C in that table; it does not name them there.

Those percentages add to 72%. That does not mean 72% of all future contracts belong to the same three customers. It is a measure of revenue concentration for one quarter.

If one large buyer delays deployment, negotiates different terms at renewal, develops its own capacity, or reduces future demand, the effect on a specialized provider could be substantial. Conversely, a long-term commitment from a strong buyer can make a large construction and equipment program financeable.

Both sides of that relationship are visible in the same business model. The customer commitment is an asset when funding capacity and a source of risk when too much of the business relies on a handful of contracts.

Hardware does not stand still

GPU systems are expensive physical assets in a fast-moving market.

A provider has to decide how long each generation of equipment will earn attractive returns. Newer chips may offer better performance or efficiency, while customers may still want older machines for certain workloads if the price is right. Networking, software, power, and cooling requirements can also change between generations.

CoreWeave's annual report identifies rapid technological change, the need to deploy new infrastructure, and potential declines in demand for existing hardware among its business risks (see Sources, [1]).

That makes the useful life of a GPU fleet more than an accounting assumption. If hardware earns enough over its service life to cover its purchase, operation, financing, and eventual replacement, the model can work. If pricing falls or equipment loses economic value faster than planned, a large contracted backlog alone may not protect the return on every asset.

Where the margin is really made

A neocloud's economics depend on much more than the hourly price of a GPU.

A useful mental model is contracted revenue minus the full cost of delivering the cluster over its working life. That cost includes the servers and network, facility and electrical capacity, cooling, service operations, financing, and the risk that the equipment becomes less valuable before it has paid for itself.

The key questions are practical:

• Can the provider deliver capacity on the schedule promised?
• Does a contract last long enough, and pay enough, to support the hardware investment?
• How much equipment remains useful after the initial customer commitment ends?
• Can the provider refinance or fund the next generation without weakening its economics?
• How much revenue depends on a few customers?

Those questions differ from the ones facing a conventional colocation operator. How Data Centers Actually Make Money explains the business of leasing space, power, and related services. A GPU cloud operator adds another capital-intensive layer: it owns or finances much of the computing equipment and sells the resulting service.

The Data Center Revenue & Capacity Calculator models an illustrative facility's power-capacity revenue. It does not estimate GPU-cloud contract revenue, GPU depreciation, debt service, or a neocloud's profitability.

The next AI infrastructure business

CoreWeave's filings show why GPU clouds attract attention: very large customer commitments can support extraordinary growth. They also show the price of that growth: substantial upfront investment, borrowing, execution risk, and customer concentration.

The business is best understood as a combination of cloud software, industrial deployment, and infrastructure finance. The provider must make the computing system available before it can fully realize the value of its contracts.

That is the billion-dollar business behind “renting Nvidia GPUs.” The valuable product is usable compute on a reliable schedule. The margin depends on how efficiently a company builds, funds, operates, and renews the physical system that delivers it.