What Is a Neocloud? How AI GPU Clouds Make and Lose Money

A neocloud is a cloud company built almost entirely around one product: access to large clusters of AI chips (GPUs), rented to the companies that train and run AI models. If the company’s revenue comes mostly from GPU capacity sold under contracts to AI labs and big tech, it is a neocloud; if GPUs are one service among hundreds, as at Amazon Web Services, Microsoft Azure or Google Cloud, it is a hyperscaler. The word spread in 2025 as these companies grew from crypto-mining or niche hosting roots into multi-billion-dollar businesses. CoreWeave, for example, says its committed contracts are take-or-pay, run one to six years, and made up 98% of its revenue in Q2 2026.

Scope: this page explains the business model, the listed companies and how to read their numbers, for a US investor. It does not cover chipmakers (see the AI compute supply chain), data-center real estate trusts, or private-company valuations. It is education, not a recommendation to buy or sell any stock.

The difference is what else the company earns and who owns the chips. A hyperscaler can fund GPUs from other profitable businesses; a neocloud has to fund them from contracts, debt and new shares; a landlord owns the building and power but not the GPUs. The categories blur at the edges: IREN still mined bitcoin in 2026, Nebius has other businesses, and Oracle rents out GPUs at huge scale.

Is Oracle a neocloud? Not by this definition: it reported $19.3 billion of revenue in its quarter to August 31, 2026, much of it from software, and $664 billion of remaining performance obligations. But its AI build-out has the same funding question: free cash flow was negative $5 billion in that quarter (Oracle release, Sep 10, 2026). See Oracle’s stock page for its price history.

The listed GPU clouds are CoreWeave and Nebius, with IREN converting from bitcoin mining; the listed landlords include Applied Digital and TeraWulf. Each company reports its contracts with its own measure, so the last column is not comparable across rows; the backlog, RPO and ARR guide explains each one. Revenue is GAAP revenue for the latest reported quarter.

Private neoclouds include Lambda, Crusoe, Fluidstack and Nscale; their numbers come from funding reports rather than audited filings, so they are left out of the table. The full side-by-side comparison, with debt, cash and customer share in one definition per column, is in neocloud stocks compared.

They earn a fixed price per GPU hour under contract and lose money after paying for the chips wearing out and the debt that bought them. Nebius said its Q2 2026 deals yielded $20–25 million of annual contract value per megawatt, with prepayments covering 50–60% of the related capex and an estimated payback of 1 year 10 months (company estimate). The unaudited quarterly income statement shows where the money goes. Worked example, CoreWeave Q2 2026 (three months to Jun 30, in millions):

Both lines are true at once: a 59% adjusted EBITDA margin and a 24% net loss margin. The gap is mostly depreciation (the GPUs losing value; CoreWeave depreciates computing equipment over six years) and interest. The cash statement shows how the build is paid for: in the first half of 2026 CoreWeave spent $14.1 billion on property and equipment, generated $3.7 billion from operations and raised $16.7 billion of new debt while repaying $5.2 billion (10-Q). Whether six years is the right life for an AI chip is its own debate, covered in how long AI GPUs last.

A handful of very large buyers on one side and a stack of secured and unsecured lenders on the other. CoreWeave’s 10-K says Microsoft was about 67% of 2025 revenue and the top two customers about 77%; in Q2 2026 the largest three unnamed customers were 36%, 26% and 10% (the letters may refer to different customers in different periods). Nebius has disclosed long-term contracts with Microsoft and Meta; its March 16, 2026 Meta agreement is worth up to about $27 billion. The full contract ledger is in who are CoreWeave’s customers.

On the lending side, CoreWeave owed $35.6 billion of debt principal at June 30, 2026: loans secured by GPUs and customer contracts, unsecured bonds with 8.50%–9.75% coupons, convertible notes and equipment-vendor financing. How those pieces differ, and what the 15% bond yields of October 2026 mean, is in CoreWeave’s debt explained. The suppliers (Nvidia for chips, memory makers, network, power and cooling) are mapped in who gets paid when AI data centers get built.

Much more than the market: the listed neoclouds and landlords have had roughly eight to nine times the S&P 500’s volatility over the past year. In Exent market data to Sep 30, 2026, CoreWeave fell 57.5% from peak to trough within the year while the S&P 500’s largest drop was 9.1%. The full table for every listed name is in neocloud stocks compared.

Those figures end on September 30. October was worse for the group: CNBC reported on October 21, 2026 that CoreWeave was down 36% for the month, Nebius 33% and IREN 38%, as rising bond yields and concerns about AI debt hit the most indebted names first. These moves are history, not a forecast. For how large a slot such stocks get in a plan, see how big a speculative sleeve should be.

The risks are concentrated and checkable in the filings. Five numbers cover most of them; each links to the page that explains it.

Counterexample, when the bearish reading is wrong: much of the secured debt is repaid from contracts with investment-grade customers, and lenders price those loans far below the bonds (IREN financed its Microsoft contract at a weighted 6.0%; Nebius’s first asset-backed loan priced at a floating benchmark rate plus 2.50%). If contracts are delivered and renewed at similar prices, the model can work even with heavy debt. When the bullish reading is wrong: a large backlog is not cash, delivery delays push revenue out while interest keeps running, and a customer that does not renew leaves GPUs to re-rent at whatever the market pays. Gartner’s June 23, 2026 forecast gives neoclouds 20% of a $267 billion AI cloud market by 2030, and its October 14 release expects most of them to be absorbed or to close; both are forecasts, not facts.

Who this is not for: an investor who needs the money within a few years, or whose portfolio already leans on the same AI theme through Nvidia or big tech (see your portfolio if the AI bubble bursts and big tech AI spending in 2026). A neocloud ETF exists: Roundhill launched NCLD on August 6, 2026 with Nebius (30.83%) and CoreWeave (27.30%) at 58% of the fund at launch, so it is a concentrated bet on the same names, not diversification (see thematic ETFs).