A year ago, one customer paid about two of every three dollars CoreWeave earned. In the quarter that ended September 30, that customer's share was 24%.

That shift sits in a small table in CoreWeave's quarterly report, filed on November 12. It's also the clearest way to see what a neocloud is. A neocloud's business depends on a short list of promises. When one of them changes, the whole model moves.

What is a neocloud?

A neocloud is a company whose main business is renting out AI computing power. Customers pay by the hour or sign contracts that run for years.

The computing power comes from graphics processing units, or GPUs. These are the chips that train and run AI models, and most of them are made by Nvidia. A neocloud buys or leases thousands of them, puts them in data centers and rents the capacity out.

The big general clouds are called hyperscalers. That means Amazon Web Services, Microsoft Azure and Google Cloud. They rent out GPUs too, but GPUs are one product among hundreds. For a neocloud, GPUs are the business.

There is no official definition. Runpod, which rents GPUs itself, says in its guide that there is no formal definition or register of neoclouds. CoreWeave and Nebius appear on almost every list. Some former bitcoin miners now count as well. IREN, for example, reported $200 million of AI cloud revenue last quarter.

How does a neocloud make money?

A neocloud makes money by renting chips for more than they cost to own, power and finance. That works only while four partners keep their promises.

Think of it as a landlord who rents out computers instead of apartments. The landlord needs four things to line up:

The chips

CoreWeave's 10-Q says all of the GPUs in its infrastructure today are Nvidia GPUs, and that customers have contractually specified Nvidia chips. Nvidia is also an owner. In January 2026 it bought about 23 million CoreWeave shares at $87.20 each, for $2.0 billion, according to the same filing.

The money

GPUs are bought before the rent comes in, so neoclouds borrow. Nebius said it would pay for its Microsoft contract with cash from the deal and debt secured against the contract. That contract is worth about $17.4 billion through 2031. When lenders get nervous, the whole group feels it. In October, worries about CoreWeave's bonds pulled down Oracle and other neocloud shares.

The buildings

Chips need a building with power and cooling. CoreWeave says its data center leases generally run three to 16 years. Smaller neoclouds lean on landlords and backers. AMD, for one, signed credit support behind a $6.5 billion data center lease for TensorWave, a neocloud that runs AMD chips.

The customers

The tenants are mostly very large companies. CoreWeave's filing names several multi-year commitments:

CoreWeave customer commitments named in its November 12, 2026 filing
Customer Commitment Runs through
Meta Up to about $21.0 billion (March 2026 order) December 2032
OpenAI Up to about $6.5 billion (September 2025 order) May 31, 2031
Jane Street About $6.0 billion (April 2026) Not stated in the filing
Microsoft Named as a significant customer; amount not given Not stated

Source: CoreWeave Form 10-Q for the quarter ended September 30, 2026. "Up to" means the most the customer could pay, not money already received.

Neocloud vs hyperscaler: rivals or customers?

Both. The big clouds compete with neoclouds for AI customers. Several of them also pay neoclouds for capacity, and Microsoft is the biggest example.

CoreWeave's 10-Q lists Amazon, Google, Microsoft and Oracle as competitors, "a number of which are also our current customers." Microsoft is the clearest case. CoreWeave's annual report says it got about 67% of its 2025 revenue from Microsoft. Microsoft also signed the $17.4 billion deal with Nebius.

Why would a giant cloud rent from a smaller one? The filings don't say. Our reading is that renting gets capacity online without the giant building it first. The giant pays rent, and the neocloud carries the debt.