Home โ€บ Blog โ€บ The neoclouds
AI Factory Case Studies ยท Part 7 ยท CoreWeave & the neoclouds

The Neoclouds: How CoreWeave Turned GPUs Into a Debt-Financed Rental Business

The first six case studies were built by labs and big tech companies for themselves or one main customer. The neoclouds do something different: they borrow against GPUs, put them in buildings owned by others, and rent the compute to whoever will sign a multi-year contract, including Microsoft, OpenAI and Meta. CoreWeave, the largest, has about $104 billion of contracted work and about $35 billion of debt. This case study explains how the model works and where it could break.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) CoreWeave figures come from its Q2 2026 results announcement and its 10-Q filing for the quarter ended 30 June 2026. The Core Scientific vote and IPO come from TechCrunch and others; credit market moves from 24/7 Wall St.; sector debt from analyst write-ups and Gartner. Customer contract values are rarely published in full, so those we found only in the press are marked โš ๏ธ. This is not investment advice. See Reading the numbers.

1. Case file

โ˜๏ธ
CoreWeave51 data centers, mostly in buildings it leases ยท listed on Nasdaq (CRWV)
Q2 2026 revenue$2.58B ยท +112% YoY
Backlog (30 Jun)~$104B
Active power1.5 GW
Contracted power3.7 GW โ†’ 4.2 GW
Total debt (30 Jun)~$35.1B
2026 revenue guidance$12.4โ€“13.2B
2026 capex guidance$35โ€“39B
Bond yields (Oct 2026)~15% (unsecured)

The backlog does not include more than $25 billion of new customer commitments CoreWeave said it signed early in the third quarter. The debt figure is net of discounts and issuance costs: about $31.4B recourse and $3.7B non-recourse, per the 10-Q.

2. Timeline

CoreWeave, 2017 โ†’ 2026

Solid dots are confirmed by a filing or company results, hollow dots come from the press, โš ๏ธ marks claims we could not confirm and dashed dots are planned. Click a dot for its source.

Click a milestone to see where it comes from.

CoreWeave was founded in 2017 and, by its founders' account, started out mining cryptocurrency with GPUs โš ๏ธ before pivoting to renting GPUs for AI. It listed at $40 a share in March 2025. In July 2025 it agreed an all-stock takeover of Core Scientific, a former bitcoin miner that hosts much of its capacity; Core Scientific's shareholders voted the deal down on 30 October 2025, with the proxy adviser ISS and its largest shareholder recommending against. In March 2026 CoreWeave closed an $8.5 billion facility backed by GPUs. By October 2026, its unsecured bonds yielded about 15% and the cost of insuring them against default had risen 30% in a month.

3. The bet: GPUs as collateral

A neocloud's business is simple to state. It signs a multi-year contract with a customer that needs compute, borrows money against that contract and the GPUs it will buy, buys the GPUs, puts them in a leased data center, and repays the loan from the customer's payments. If the customer pays and the GPUs keep earning until the loan is repaid, the neocloud keeps the margin and owns hardware it can rent again. If demand falls, prices drop or a customer walks away, the loan is still there and the collateral is a depreciating chip.

๐Ÿ’ก Abilene, one layer upStargate Abilene (Part 2) financed the building against a lease from Oracle. Neoclouds finance the computers against contracts from labs and hyperscalers. Both turn AI demand into something lenders understand, and both depend on a small number of very large customers continuing to pay.

4. How a GPU loan works

The GPU loan model

An illustration with made-up but plausible inputs, not CoreWeave figures. Each GPU is financed fully with debt. Watch whether the loan is repaid before the GPU's value runs out.

Includes its share of servers and networking.
Per GPU per hour.
Years before the GPU is worth little.

The model shows why the useful life of a GPU is the most argued-over number in the business. A longer life means lower annual depreciation and more years to repay; a shorter life, as each new NVIDIA generation makes the last one less valuable, can leave a lender holding chips worth less than the debt. Rental prices matter just as much: if the price per GPU-hour falls as supply grows, the same loan takes longer to repay.

5. Contracted vs active

The power funnel

CoreWeave's power at different stages, as of mid-2026. The gap between contracted and active is capacity promised to customers but not yet running.

At the end of June 2026 CoreWeave had 1.5 GW running across 51 data centers and 3.7 GW contracted; by the August earnings call contracted power had reached 4.2 GW. Turning the other 2.7 GW into revenue means building or leasing, powering and filling that capacity on time, which is where the capex guidance of $35โ€“39 billion for 2026 goes.

Backlog, revenue and debt

Billions of dollars. Backlog is future contracted revenue; revenue is the 2026 guidance range (top); debt is total at 30 June 2026.

6. Customers

Neoclouds depend on a handful of very large customers. In 2025 one report said Microsoft accounted for more than 70% of CoreWeave's revenue in a single quarter โš ๏ธ. CoreWeave has since signed multi-billion-dollar contracts with OpenAI and Meta โš ๏ธ, and a report on its Q2 2026 results said new contracts with Meta and Anthropic drove the quarter โš ๏ธ. A broader customer base reduces the risk of any one leaving, but the customers are all exposed to the same thing: whether AI revenue keeps up with AI spending.

โš–๏ธ ConcentrationA neocloud whose top customers are also building their own AI factories (Microsoft's Fairwater, Meta's Hyperion) is renting to companies that may one day need it less. Long contracts protect against that, but only for as long as they run.

7. Buildings it does not own

Unlike every other site in this series, a neocloud usually owns neither the land nor the building. CoreWeave leases space from data center developers and from former bitcoin miners such as Core Scientific, which already had power connections and buildings that could be converted for AI. The failed takeover was an attempt to own more of that layer. The approach lets a neocloud grow quickly across many sites, but adds another party, and another lease, between the GPUs and the power.

Who owns what: neocloud vs hyperscaler

8. The wider field

The neocloud landscape

Selected companies and how they are financed, as reported. Tap a card.

Analysts estimate the sector holds more than $20 billion of GPU-backed debt. Gartner projects that more than 80% of independent neocloud companies will no longer operate independently by 2030, through acquisitions, mergers or failure. Related reading: who pays for the AI build-out.

9. Seven bets compared

ColossusAbileneFairwaterRainierMetaGoogleCoreWeave
The betSpeedOutside capitalNetworkOwn siliconEvery building typeWhole stack earlyRent GPUs, financed with debt
Owns the building?YesJVYesYesPartlyYesMostly no
Main userItselfOpenAIOpenAI, itselfAnthropicItselfItself, AnthropicMany: Microsoft, OpenAI, Metaโ€ฆ
Main riskPermitsOne customerUtility gasWellsStranded plantsEmissionsDebt and GPU value
Compute for rentThe AI factory as a service business
GPUs as collateralLoans repaid by customer contracts
Asset-light buildingsLeased from developers and ex-miners
Credit is the signal~15% bond yields in Oct 2026
VerdictThe neoclouds are the most financially exposed AI factories in this series. They let labs and hyperscalers add capacity without building it, and they grew faster than almost anyone expected. The same leverage that made that possible makes them the first place stress shows up if GPU prices or AI demand disappoint, which is why their bond yields are worth watching. Next, China: AI factories built under export controls.

10. Reading the numbers

From CoreWeave
Q2 2026 revenue $2.58B (+112%); backlog ~$104B; 51 data centers; 1.5 GW active; 3.7 GW โ†’ 4.2 GW contracted; 2026 guidance $12.4โ€“13.2B revenue and $35โ€“39B capex; total debt ~$35.1B at 30 Jun 2026 (10-Q).
Press
$40 IPO (Mar 2025); Core Scientific vote (30 Oct 2025, ~$9B deal); $8.5B GPU-backed facility (Mar 2026); ~15% bond yields and a 30% rise in default insurance cost (Oct 2026).
Unverified โš ๏ธ
Crypto-mining origins; Microsoft's share of revenue; OpenAI, Meta and Anthropic contract values.
Sector
>$20B of GPU-backed debt is an analyst estimate; the 80% figure is a Gartner projection.
Illustrations
The GPU loan model uses made-up inputs and simplified economics (a fixed share of revenue for operating costs, straight-line depreciation, no taxes or refinancing).