Neocloud

A neocloud is a cloud computing company whose business is renting out GPU capacity for artificial intelligence workloads, rather than selling the broad catalogue of storage, database, networking and application services that defines a general purpose cloud. The research firm SemiAnalysis set out the widely cited definition in an October 2024 report, describing an AI Neocloud as "a new breed of cloud compute provider focused on offering GPU compute rental" ^[1].

The category sits between two familiar things. Hyperscalers such as Amazon Web Services, Microsoft Azure and Google Cloud have the lowest cost of capital and the widest product surface, but they allocate GPU capacity across many internal and external claims. Traditional colocation providers sell space and power but not the machines. Neoclouds buy the machines, wire them into large InfiniBand or Ethernet fabrics, and sell the resulting clusters on multi-year committed contracts, usually as bare metal or thin managed layers. The commercial consequence is that a neocloud is closer to a leveraged infrastructure operator than to a software company: it raises debt against hardware and customer contracts, spends it on data centers and accelerators, and earns a spread over the life of the assets. The result is one of the most capital-intensive corners of the AI infrastructure market.

By 2026 the largest firms in the category were operating at meaningful scale. CoreWeave reported $5.13 billion of 2025 revenue and $60.7 billion of remaining performance obligations at year end ^[3]. Nebius reported $529.8 million of 2025 revenue and $399.0 million in the first quarter of 2026 alone ^[9][10]. Private firms including Lambda Labs, Crusoe, Nscale, Fluidstack and Together AI had each raised hundreds of millions to billions of dollars alongside contracts with frontier labs and hyperscalers.

Definition and scope

SemiAnalysis divided the GPU rental market into four groups: traditional hyperscalers; "Neocloud Giants" such as Crusoe, Nebius, Lambda Labs and CoreWeave, whose current or planned capacity runs well in excess of 100,000 H100 equivalents across all their sites; a long tail of emerging neoclouds that are relatively inexperienced at running data center infrastructure and carry a higher cost of capital, a group that includes many regional players under the sovereign AI umbrella; and capital-light brokers, platforms and aggregators. It treated a fifth model, clusters set up by venture capital firms for the exclusive use of portfolio companies, as sitting outside those four categories ^[1].

The distinction matters because "neocloud" is used loosely for very different businesses. The brokers and aggregators match supply and demand without owning hardware; the asset owners, CoreWeave and Crusoe among them, carry the machines and the debt ^[1]. Only the second group faces the depreciation and refinancing problems described below.

Buyers began to treat the category as a rated market. SemiAnalysis runs ClusterMAX, a rating system that scores GPU clouds on ten criteria including security, orchestration, storage, networking, reliability, monitoring, pricing, partnerships and availability, and assigns four medallion tiers, Platinum, Gold, Silver and Bronze, plus a Not Recommended band split into Underperforming and Unavailable. Version 1.0 was published in March 2025, version 2.0 in November 2025 and version 2.1 in April 2026 ^[2].

Origins

Two supply constraints created the opening. The first was the scarcity of high-end accelerators relative to demand, the condition SemiAnalysis was already mapping when it described the neocloud build-out in 2024 ^[1] and which returned as an acute shortage in 2026, when on-demand rental capacity sold out across GPU types ^[22]. The second was power and shell availability: sites with grid interconnects already energized were worth more than sites that merely had land.

That second constraint explains why so many neoclouds emerged from cryptocurrency mining. CoreWeave began in 2017, when Michael Intrator, Brian Venturo and Brannin McBee established Atlantic Crypto Corp. to mine ethereum with Nvidia GPUs; the company built a GPU cloud rental business from 2020 and shut the mining unit down in September 2022 ^[27]. Bitcoin miners including IREN and TeraWulf converted or leased mining sites into AI capacity, in some cases becoming landlords to neoclouds rather than operators themselves ^[18][19].

Nebius took a different route. Its predecessor was Yandex N.V., the Dutch parent of the Russian search company. Nebius completed the divestment of its Russian businesses in July 2024; those businesses had constituted more than 95% of the group's consolidated revenues, assets and employees at the time. The remaining Amsterdam-based entity changed its name from Yandex N.V. to Nebius Group N.V. in August 2024 and refocused on AI cloud infrastructure ^[9].

Business model

The defining commercial feature is the take-or-pay committed contract. Customers buy a fixed amount of capacity for a fixed term and pay whether or not they use it. In CoreWeave's case, committed contracts accounted for over 98% of 2025 revenue, up from 96% in 2024 and 88% in 2023; initial contract periods generally run one to six years, the weighted-average duration of committed contracts was approximately five years at the end of 2025, and the weighted-average prepayment across active contracts was 15% to 25% of total contract value ^[3].

Those contracts are the collateral. CoreWeave states that it primarily finances its infrastructure development "through asset-level debt supported by take-or-pay customer contracts, supplemented by corporate-level equity and debt financing" ^[3]. Nebius described the same logic when it announced its Microsoft agreement, saying it expected to finance the associated capital expenditure through a combination of cash flow from the deal and debt secured against the contract, at terms enhanced by the credit quality of the counterparty ^[11].

The economics are therefore driven by four variables: the price and delivery date of the accelerators, the cost of debt, the contracted rental rate, and the number of years the hardware keeps earning. CoreWeave's disclosed debt costs show how expensive the early years were and how quickly they improved.

Scroll sideways for more →

Source: CoreWeave FY2025 Form 10-K ^[3]. Total principal was $8,033 million a year earlier; total interest expense net of capitalized interest was $1,148 million in 2025.

In March 2026 CoreWeave closed an $8.5 billion delayed draw term loan maturing in March 2032, priced at SOFR plus 2.25% floating or approximately 5.9% fixed, and rated A3 by Moody's and A (low) by DBRS. The company described it as the first investment grade rated financing secured by high performance computing infrastructure and an associated customer contract ^[6].

Major providers

Scroll sideways for more →

Smaller operators including Voltage Park, RunPod and Vast.ai occupy the on-demand end of the market, selling capacity without a long-term commitment rather than by the year. Voltage Park advertises a single H100 for one hour with no contract ^[32], RunPod prices dedicated GPU pods by the hour ^[33], and Vast.ai runs a marketplace whose prices are set by supply and demand across more than 40 data centers and billed by the second ^[34].

Several of the largest contracts run to a single counterparty. Microsoft signed the Nebius agreement on 8 September 2025, the Nscale agreement on 15 October, and the Lambda and IREN agreements on 3 November, all within two months ^{[11][28][13][19]}. That buying pattern is one reason the category's revenue stays concentrated even as the list of providers grows.

Economics and risk

Depreciation assumptions

Because accelerators are the dominant asset, the assumed useful life directly determines reported margins. CoreWeave depreciates technology equipment over six years, having changed its estimate from five to six years effective 1 January 2023 ^[3]. Nebius stated in its 2025 annual report that "our main assets, GPUs, are currently depreciated over four years," then disclosed that in January 2026 it concluded the useful lives of servers and network equipment should be extended from four to five years, a change expected to reduce fiscal 2026 depreciation expense by approximately $167.6 million ^[9].

Those disclosures sit at the center of a public argument. In a post on X in November 2025 the investor Michael Burry accused hyperscalers of understating depreciation by assuming longer useful lives than the hardware cycle warrants, calling it "one of the more common frauds of the modern era" and singling out Oracle and Meta Platforms ^[24]. A counterargument is that the relevant question is economic rather than accounting: SemiAnalysis noted in April 2026 that higher rental rates extend the economic useful life of existing GPUs, because invested capital keeps generating cash flow for longer before it must be replaced ^[22]. CoreWeave's own risk disclosure concedes the uncertainty, warning that changes to the significant assumptions underlying its estimates of component useful lives, or any inability to redeploy infrastructure past its contracted life, could significantly affect results ^[3].

Customer and supplier concentration

Neocloud revenue is unusually concentrated. CoreWeave recognized approximately 67% of its 2025 revenue from a single customer, Microsoft, and its largest customer also accounted for 68% of net accounts receivable at year end. The equivalent top-customer share in its concentration table was 62% in 2024 and 35% in 2023, though the company cautions that the anonymized customer labels in that table may refer to different customers in different periods ^[3]. Nebius reported that its largest customer accounted for 25% of 2025 revenue and a second for 15%, with one customer representing 83% of the gross fair value of accounts receivable at year end ^[9].

Supply is concentrated too. CoreWeave states that all of the GPUs used in its infrastructure are Nvidia GPUs as a result of obligations in its customer contracts, and that three suppliers accounted for 23%, 20% and 17% of total purchases in 2025 ^[3].

Circular financing

The most contested feature of the category is the degree to which chip vendors and cloud buyers finance their own customers and suppliers. In September 2025 CoreWeave disclosed a $6.3 billion order under which Nvidia will acquire residual cloud capacity not taken up by other customers through 13 April 2032 ^[8]. In January 2026 Nvidia invested $2 billion in CoreWeave Class A common stock at $87.20 per share ^[3], and it was an investor in Nscale's Series B, Crusoe's Series E and Together AI's Series C ^[15][14][20].

Google used a different instrument. Rather than lending to Fluidstack, it backstopped Fluidstack's lease obligations to data center landlords: in August 2025 TeraWulf signed ten-year hosting agreements worth about $3.7 billion in contracted revenue for more than 200 MW of critical IT load at its Lake Mariner campus, with Google backstopping $1.8 billion of Fluidstack's lease obligations in exchange for warrants over approximately 41 million TeraWulf shares, about 8% of the company on a pro forma basis ^[18].

Supporters argue these structures are ordinary credit enhancement for a young asset class. Critics place them in the AI bubble debate, on the grounds that a chip vendor guaranteeing demand for the capacity built with its own chips obscures how much genuine third-party demand exists.

Price and refinancing risk

Rental prices have moved sharply in both directions. SemiAnalysis observed meaningful declines in on-demand GPU pricing as early as its October 2024 report ^[1], and rates for H100 capacity kept falling into 2025 before reversing: SemiAnalysis reported that one-year H100 rental contract pricing rose almost 40% from a low of $1.70 per GPU-hour in October 2025 to $2.35 by March 2026, with on-demand capacity sold out across GPU types and all capacity coming online until August to September 2026 already booked ^[22]. For an operator whose debt is fixed and whose contracts are signed years ahead, that volatility cuts both ways.

Consolidation has not been smooth either. CoreWeave agreed in 2025 to acquire the data center operator Core Scientific, but Core Scientific did not receive the votes necessary to approve the merger agreement and the deal was terminated on 30 October 2025 ^[25].

Relationship to sovereign AI

Neoclouds became a default delivery vehicle for national compute programs, because governments wanting capacity inside their own jurisdiction generally lack the operating capability to build it. Nscale is the clearest example. In July 2025 it announced Stargate Norway, a 50/50 joint venture with Aker at Kvandal outside Narvik, with an initial 230 MW and ambitions to expand by a further 290 MW, closed-loop direct-to-chip liquid cooling, and a target of 100,000 Nvidia GPUs by the end of 2026 ^[16]. A companion project, Stargate UK, was announced with OpenAI and Nvidia on 16 September 2025; Nscale said OpenAI would explore offtake of 8,000 Nvidia GPUs in the first quarter of 2026 with potential to scale to 31,000, alongside a separate 50 MW Loughton campus for Microsoft holding 23,040 GB300 GPUs ^[26]. Nscale's Series B materials describe the company as deploying infrastructure across Europe, North America and the Middle East, with data centers in Norway, the United Kingdom, Portugal, Iceland and the United States ^[15].

The Stargate program in the United States follows the same pattern with a different owner mix: Crusoe built the flagship Abilene, Texas campus, whose first phase went live on Oracle Cloud Infrastructure on 30 September 2025 ^[29]. The full eight-building site is designed to support hundreds of thousands of GPUs on a single integrated network fabric ^[29].

Developments in 2025 and 2026

CoreWeave's initial public offering was the category's first public market test. The company priced 37.5 million Class A shares at $40.00 on 27 March 2025, below the $47 to $55 range it had marketed and on a reduced deal size, raising $1.5 billion; the shares began trading on the Nasdaq Global Select Market under the ticker CRWV the following day ^[4][7]. For fiscal 2025 it reported $5,131 million of revenue and a $1,167 million net loss, capital expenditures of $10,309 million, a revenue backlog of $66.8 billion, and 43 data centers with over 850 MW of active power and approximately 3.1 GW of contracted power ^[3][5]. On its 26 February 2026 earnings call it guided to 2026 revenue of $12 billion to $13 billion and capital expenditure of $30 billion to $35 billion, more than double the 2025 figure ^[23]. The shares fell 18% the next day, and chief executive Michael Intrator publicly defended the spending plan ^[30].

Nebius reported first quarter 2026 revenue of $399.0 million, up 684% year over year, with adjusted EBITDA of $129.5 million against a $53.7 million loss a year earlier, $2,472.9 million of purchases of property, equipment and intangible assets in the quarter, and $9,298.2 million of cash at 31 March 2026. It also announced up to 1.2 GW of power and land for a new owned AI factory in Pennsylvania ^[10]. In March 2026 it signed an agreement with Meta worth up to approximately $27 billion over five years, comprising $12 billion of dedicated capacity plus up to $15 billion of additional available capacity, based on the Nvidia Vera Rubin NVL72 platform, with deliveries starting in early 2027 ^[12].

On the private side, Anthropic announced on 12 November 2025 that it would invest $50 billion in American AI infrastructure built with Fluidstack, with initial sites in Texas and New York, about 800 permanent and 2,400 construction jobs, and facilities coming online through 2026 ^[17]. Fluidstack disclosed an $830 million Series A at a $7.5 billion valuation in July 2026, led by Situational Awareness ^[21]. Together AI announced $800 million of Series C funding on 1 July 2026 from a group of investors including Aramco Ventures and Nvidia, alongside commitments for more than 500 MW of compute capacity ^[20]. Lambda announced more than $1.5 billion in a Series E led by TWG Global, with the US Innovative Technology Fund and existing investors, on 18 November 2025, two weeks after its Microsoft agreement ^[31].

What the category has not yet done is operate through a period of surplus. Every model in it assumes that hardware can be kept earning past its initial contract or refinanced on better terms, and the 2026 shortage postponed that test rather than settling it. Meanwhile the largest operators kept committing capital well ahead of revenue: CoreWeave's 2026 capital expenditure guidance of $30 billion to $35 billion is between roughly two and three times the revenue it guided for the same year ^[23].

See also

• CoreWeave
• Hyperscaler
• AI infrastructure
• Sovereign AI
• GPU cluster
• AI bubble

References

1. ^SemiAnalysis, "AI Neocloud Playbook and Anatomy," 3 October 2024. newsletter.semianalysis.com/...laybook-and-anatomy
2. ^SemiAnalysis, "ClusterMAX Overview." clustermax.ai/overview
3. ^CoreWeave, Inc., Annual Report on Form 10-K for the fiscal year ended 31 December 2025. s205.q4cdn.com/...CoreWeave-Inc-FY25-10-K-7.pdf
4. ^CoreWeave, "CoreWeave Announces Pricing of Initial Public Offering," 27 March 2025. investors.coreweave.com/...default
5. ^CoreWeave, "CoreWeave Reports Strong Fourth Quarter and Fiscal Year 2025 Results," 26 February 2026. investors.coreweave.com/...default
6. ^CoreWeave, "CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment Grade Rated GPU-backed Financing," 31 March 2026. investors.coreweave.com/...default
7. ^CNBC, "CoreWeave prices IPO at $40 a share, below expected range," 27 March 2025. cnbc.com/...ipo-at-40-a-share-below-expected-range
8. ^RCR Wireless News, "Nvidia commits $6.3B for CoreWeave's unused cloud capacity," 16 September 2025. rcrwireless.com/...nvidia-cloud
9. ^Nebius Group N.V., Annual Report on Form 20-F for the fiscal year ended 31 December 2025. sec.gov/...nbis-20251231x20f
10. ^Nebius, "Nebius reports first quarter 2026 financial results," 13 May 2026. nebius.com/...first-quarter-2026-financial-results
11. ^Nebius, "Nebius announces multi-billion dollar agreement with Microsoft for AI infrastructure," 8 September 2025. nebius.com/...with-microsoft-for-ai-infrastructure
12. ^Nebius, "Nebius signs new AI infrastructure agreement with Meta," 16 March 2026. nebius.com/...i-infrastructure-agreement-with-meta
13. ^Lambda, "Lambda announces multibillion-dollar agreement with Microsoft to deploy AI infrastructure powered by tens of thousands of NVIDIA GPUs," 3 November 2025. lambda.ai/...d-by-tens-of-thousands-of-nvidia-gpus
14. ^Crusoe, "Crusoe Announces Series E Funding," 24 October 2025. crusoe.ai/...crusoe-announces-series-e-funding
15. ^Nscale, "Nscale Raises the Largest Series B in European History, at $1.1 billion," 25 September 2025. nscale.com/...nscale-series-b
16. ^Nscale, "Nscale, Aker and OpenAI to establish Stargate Norway," 31 July 2025. nscale.com/...stargate-norway-nscale-aker-openai
17. ^Anthropic, "Anthropic invests $50 billion in American AI infrastructure," 12 November 2025. anthropic.com/...ion-in-american-ai-infrastructure
18. ^TeraWulf, "TeraWulf Signs 200+ MW, 10-Year AI Hosting Agreements with Fluidstack," 14 August 2025. investors.terawulf.com/...reements-with-fluidstack
19. ^IREN, "IREN signs $9.7 billion agreement with Microsoft to deploy AI Cloud infrastructure," 3 November 2025. iren.com/...soft-to-deploy-ai-cloud-infrastructure
20. ^Together AI, "Announcing our Series C," 1 July 2026. together.ai/...announcing-our-series-c
21. ^SiliconANGLE, "AI data center builder Fluidstack raises $830M at $7.5B valuation," 21 July 2026. siliconangle.com/...ack-raises-830m-7-5b-valuation
22. ^SemiAnalysis, "The Great GPU Shortage: Rental Capacity," 2 April 2026. newsletter.semianalysis.com/...age-rental-capacity
23. ^The Motley Fool, "CoreWeave (CRWV) Q4 2025 Earnings Call Transcript" (call held 26 February 2026), published 27 February 2026. fool.com/...-crwv-q4-2025-earnings-call-transcript
24. ^CNBC, "Big Short investor Michael Burry accuses AI hyperscalers of artificially boosting earnings," 11 November 2025. cnbc.com/...lers-of-artificially-boosting-earnings
25. ^Core Scientific, "Core Scientific Announces Termination of Merger Agreement with CoreWeave," 30 October 2025. investors.corescientific.com/...ent-with-coreweave
26. ^Nscale, "Nscale announces UK AI infrastructure commitment," 16 September 2025. nscale.com/...le-uk-ai-infrastructure-announcement
27. ^CNBC, "CoreWeave's 7-year journey to IPO wound through crypto before AI," 30 March 2025. cnbc.com/...-to-ipo-wound-through-crypto-before-ai
28. ^Nscale, "Nscale signs expanded agreement with Microsoft for approximately 200,000 NVIDIA GB300 GPUs," 15 October 2025. nscale.com/...nscale-microsoft-2025
29. ^Crusoe, "Crusoe Announces Flagship Abilene Data Center Is Live," 30 September 2025. crusoe.ai/...-flagship-abilene-data-center-is-live
30. ^CNBC, "CoreWeave CEO defends spending plans as stock plummets," 27 February 2026. cnbc.com/...coreweave-ceo-intrator-stock-ai-spending
31. ^Lambda, "Lambda raises over $1.5B from TWG Global & USIT to build superintelligence cloud infrastructure," 18 November 2025. lambda.ai/...uperintelligence-cloud-infrastructure
32. ^Voltage Park, "Pricing." voltagepark.com/pricing
33. ^RunPod, "Pricing." runpod.io/pricing
34. ^Vast.ai, "Pricing." vast.ai/pricing