AIAIExplainer NO. ---

What is a neocloud? How GPU clouds like CoreWeave make money

Neoclouds rent Nvidia GPUs to AI labs and big tech on multi-year contracts, funded mostly with debt. Here is how CoreWeave-style providers work, how they differ from AWS, Azure and Google Cloud, and why their bonds are under pressure.

By Rowan Ashby Published 23.10.26 14:16 UTC 9 min read

What is a neocloud? How GPU clouds like CoreWeave make money — Unknown /

Key takeaways

A neocloud is a cloud provider built almost entirely around renting Nvidia GPUs for AI training and inference. CoreWeave, Nebius, Lambda, Crusoe and Nscale are the best-known names.
Neoclouds mostly earn money from multi-year, take-or-pay capacity contracts with a few very large customers, such as Microsoft, Meta and OpenAI.
The model needs a lot of borrowing. CoreWeave reported about $104 billion of revenue backlog on June 30, 2026, and about $35 billion of debt on its balance sheet, before lease liabilities.
Nvidia sells the chips, owns equity in some neoclouds and, in CoreWeave's case, has agreed to buy unused capacity until 2032.
In October 2026, CoreWeave's bond yields and credit default swap spreads climbed to distressed levels. On October 22, Nvidia's Jensen Huang said a "scarcity of capital" is holding the sector back.

What is a neocloud?

What is a neocloud? It is a specialist cloud company that rents GPU computing power, mostly Nvidia chips, for AI work, and sells few other cloud services. CoreWeave is the largest. Where Amazon Web Services sells hundreds of services, a neocloud mainly sells GPU time.

The term spread through industry analysis in 2024 and 2025 as start-ups raced to fill GPU clusters faster than the big clouds could. It reached mainstream financial coverage in October 2026 for three reasons. On October 14, Gartner forecast that most neoclouds will not survive as independent companies. Over the following week, CoreWeave's debt sold off. On October 22, Nvidia chief executive Jensen Huang said in a CNBC interview at the company's GTC event in Berlin that "the demand for the neo clouds is off the charts" but that financing is the bottleneck.

This explainer covers how neoclouds work, how they make money, why they borrow so much, what Nvidia's role is, and which risks are now showing up in credit markets.

Neocloud vs AWS, Azure and Google Cloud

The big three clouds, often called hyperscalers, are run by companies with large, diversified profits that can fund data centres from their own cash flow. They design their own AI chips (Google's TPU, Amazon's Trainium) as well as buying Nvidia's, and they sell AI compute as one part of a much wider product range. Our explainer on Google TorchTPU, Amazon TorchNeuron and Nvidia CUDA covers how that chip competition works.

Neoclouds are narrower, and in some ways faster. They put most of their capital into the latest Nvidia systems, tune their networks and storage for large training jobs, and sell bare-metal or lightly managed GPU clusters. Their customers include the hyperscalers themselves, which rent neocloud capacity when their own data centres cannot be built quickly enough.

Feature	Neocloud (e.g. CoreWeave, Nebius)	Hyperscaler (AWS, Azure, Google Cloud)
Main product	GPU clusters for AI training and inference	Hundreds of services: compute, storage, databases, software, AI
Chips	Almost entirely Nvidia	Nvidia plus in-house chips (TPU, Trainium, Maia)
Typical customer	A small number of AI labs and big tech firms on multi-year contracts	Millions of businesses, mostly on pay-as-you-go
Funding	Mostly debt, equity raises and customer prepayments	Mostly internal cash flow plus investment-grade bonds
Credit rating	Speculative grade (CoreWeave's secured GPU facility: Ba2 Moody's, BB+ Fitch)	Investment grade

Sources: CoreWeave DDTL 5.5 press release, August 10, 2026; company filings.

Who are the main neocloud providers?

CoreWeave (Nasdaq: CRWV). The largest neocloud, which listed in March 2025. Microsoft accounted for 62% of its 2024 revenue, according to its SEC filings. It also has large contracts with OpenAI and Meta.
Nebius (Nasdaq: NBIS). Based in Amsterdam. In September 2025, according to its SEC filing, it signed a five-year deal to supply Microsoft with dedicated GPU capacity from Vineland, New Jersey, worth about $17.4 billion, or up to about $19.4 billion if Microsoft buys more. In November 2025 it added a Meta contract worth about $2.9 billion.
Lambda, Crusoe and Nscale. Privately held providers that range from self-serve GPU rentals for developers to large, energy-linked campuses.

How do neoclouds make money?

The business is simple to describe. A neocloud buys GPUs, servers and networking gear, leases or builds data centre space with enough power, and rents the finished clusters out by the GPU-hour.

The key feature is the contract. Large customers usually sign take-or-pay agreements: they commit to pay for a fixed block of capacity for several years, whether or not they use it. Those commitments appear as backlog, which is revenue the company expects to recognise in the future. In its August 11, 2026 results, CoreWeave defined backlog as remaining performance obligations plus other amounts it expects to earn under committed customer contracts.

CoreWeave metric (quarter to June 30, 2026)	Value	Year earlier
Revenue	$2.575 billion	$1.212 billion
Adjusted EBITDA (margin)	$1.510 billion (59%)	$753 million (62%)
Net interest expense	$640 million	$267 million
Net loss	$626 million	$290 million
Capital spending (property and equipment)	$6.42 billion	n/a
Revenue backlog	About $104 billion	n/a
Active / contracted power	About 1.5 GW / 3.7 GW	n/a

Source: CoreWeave Q2 2026 results release, August 11, 2026. CoreWeave said backlog excluded more than $25 billion of net new commitments signed in early Q3.

The table shows how the model works. Operating cash generation, measured by adjusted EBITDA, is large. Yet the company still lost money in the quarter, mainly because interest costs ($640 million) were higher than its adjusted operating profit after depreciation. The faster a neocloud grows, the more it borrows ahead of the revenue.

Why do neoclouds borrow so heavily?

GPUs are expensive and have to be paid for long before customers' monthly payments arrive. CoreWeave spent about $14.1 billion on property and equipment in the first half of 2026 alone, according to its results. That is far more than it could fund from operations.

The usual tool is a delayed draw term loan (DDTL) placed in a separate special purpose vehicle. Lenders are secured on the GPUs and on the customer contracts those GPUs serve. The company draws money as each tranche of hardware goes live. Because repayment depends mainly on the end customer paying under a take-or-pay contract, these loans can be priced more tightly than the neocloud's own credit would allow.

A recent example shows both the strengths and the weak points. On August 10, 2026, CoreWeave closed a $2.6 billion facility, which it called DDTL 5.5, priced at SOFR plus 5.50% and rated Ba2 by Moody's and BB+ by Fitch. The loan runs about five years, while the underlying customer contracts average about three. CoreWeave said lenders were signalling "confidence in the long-term value of NVIDIA GPUs" and were willing to take renewal risk. In other words, the lenders are betting the chips can be re-rented at useful prices once the first contracts end.

At June 30, 2026, CoreWeave's balance sheet showed about $31.4 billion of recourse debt and $3.7 billion of non-recourse debt. About $7.5 billion of that was due within 12 months. Lease obligations come on top.

Nvidia's role: supplier, shareholder and backstop

Nvidia is linked to the neocloud model in three ways:

Supplier. Neocloud capital spending goes mostly into Nvidia systems. That is why neocloud stress hits Nvidia shares, as our coverage of the Nvidia stock correction and the 5.6% chip-stock slump showed.
Shareholder. On January 26, 2026, Nvidia invested a further $2 billion in CoreWeave at $87.20 a share, according to the companies' announcement as reported by Bloomberg and Yahoo Finance.
Backstop. In a September 2025 filing, CoreWeave disclosed an order worth an initial $6.3 billion. Under it, Nvidia buys any cloud capacity that CoreWeave's customers do not use, through April 13, 2032.

Critics, cited in Bloomberg's coverage of the January deal, call this "circular" financing: the chipmaker puts money into a customer that then buys its chips. Nvidia says the agreements secure computing capacity that it and the wider ecosystem need.

Why neocloud credit risk is in the news

In mid-October 2026, investors started asking whether the debt could be refinanced at reasonable cost. According to CNBC, CoreWeave's unsecured bonds were yielding about 15% around October 19. On October 21, CNBC reported that its five-year credit default swap spread stood near 1,145 basis points, a level usually associated with distressed borrowers. For how to read that number, see our guide to credit default swaps explained and our report on Meta's record CDS levels.

CoreWeave shares fell about 10% on October 22, according to CNBC's market coverage. That was the day Huang spoke. He said "the working capital to fuel that growth" is what is missing and that the gap "is going to get solved." Wells Fargo traders, quoted by CNBC, said his comments added to worries about appetite for risk in credit markets.

Not every investor is selling. John Lloyd of Janus Henderson, whose fund holds CoreWeave's 2031 bonds, told CNBC on October 20 that the sell-off looked like "a little bit of overreacting." No default or missed payment had been reported as of October 23, 2026.

The longer-term warning came from Gartner. On October 14 it predicted that more than 80% of neocloud providers will stop operating independently by 2030, pointing to cost pressure and the lack of a differentiating business model. It also said more than $3.5 trillion of AI infrastructure spending could face impairment by the end of 2030. Gartner did not name which companies it counts as neoclouds.

The main risks in the neocloud model

Customer concentration. A few buyers make up most revenue. If one cuts back or renegotiates, backlog shrinks quickly.
Refinancing. Debt that comes due has to be rolled over. At yields near 15%, new borrowing can cost more than the GPUs earn.
Chip obsolescence. Nvidia releases new architectures about every year. If older GPUs rent for much less when contracts renew, the collateral behind the loans is worth less.
Power and construction. Revenue starts only when a site is energised. Delays push back cash flow but not interest bills. Huang also pointed to power as a constraint in Europe.
Delivery penalties. Nebius's filings say Microsoft can take service credits or end tranches if delivery schedules or uptime targets are missed.

What it means

A neocloud is easiest to understand as a leveraged landlord for AI chips. When demand is strong and credit is cheap, the model grows very fast: CoreWeave's revenue more than doubled in a year. When credit tightens, the same structure becomes fragile, because growth depends on borrowing ahead of revenue that is promised but not yet received.

For anyone following AI, neocloud bond yields and CDS spreads are now one of the clearest public signals of how much funding is still available for the AI build-out. They also matter beyond neoclouds. Stress there feeds back to Nvidia's order book, to the hyperscalers that rent extra capacity, and to the wider debate over whether AI revenue will grow fast enough to pay for the data centres. The things to watch are CoreWeave's next quarterly results, its next refinancing deals, and whether any neocloud is taken over, as Gartner expects many to be.

Frequently asked questions

What is a neocloud?

A neocloud is a specialist cloud provider that mainly rents out GPU computing power, mostly Nvidia chips, for AI training and inference, rather than offering the broad range of services sold by AWS, Microsoft Azure or Google Cloud. CoreWeave, Nebius, Lambda, Crusoe and Nscale are common examples.

How is a neocloud different from AWS or Azure?

Hyperscalers sell hundreds of services to millions of customers and fund data centres largely from their own profits. Neoclouds focus on GPU clusters, depend on a few large customers on multi-year take-or-pay contracts, and fund growth mostly with debt and equity raises. Hyperscalers such as Microsoft also rent capacity from neoclouds.

Why is CoreWeave's debt under pressure?

In October 2026, investors questioned how costly it would be for CoreWeave to refinance roughly $35 billion of debt reported at June 30, 2026. CNBC reported unsecured bond yields near 15% and a five-year credit default swap spread near 1,145 basis points on October 21. No default had been reported.

What did Jensen Huang say about neoclouds?

In a CNBC interview at Nvidia's GTC event in Berlin on October 22, 2026, Huang said demand for neoclouds is "off the charts" but that the working capital to fund their growth is scarce, adding that the gap "is going to get solved."