Neocloud is an industry label for a newer group of cloud providers whose main business is renting GPU-backed servers and virtual machines for artificial intelligence workloads. Uptime Institute, which analysed the segment in February 2025, described these firms as specialists in AI infrastructure as a service and listed CoreWeave, Nebius, Lambda Labs, Vast AI, OVH, Vultr, DigitalOcean and Paperspace among them. [1]

Equinix, whose facilities host several of these providers, defines a neocloud simply as a vendor offering AI-specific infrastructure and services, and dates the wider use of the term to late 2024, when GPU scarcity pushed buyers to look beyond the largest clouds. Its examples include CoreWeave, Crusoe, Denvr Dataworks, GroqCloud, Lambda and Nebius. [2]

The boundaries of the label are loose. It is applied both to companies founded around GPUs and to older hosting providers that have added large GPU fleets, and some firms that build their own data centers describe themselves primarily as infrastructure developers rather than clouds. What unites them is that accelerated compute is the core product, not one service among hundreds.

Hyperscalers such as AWS, Microsoft Azure and Google Cloud operate broad platforms spanning compute, storage, databases, analytics and many managed services across dozens of regions. Uptime Institute noted that the three largest controlled roughly two-thirds of global cloud spending by the end of 2023. Neoclouds, by contrast, offer a limited product line centred on GPU servers, and Uptime characterised them as closer to commodity GPU suppliers that position themselves as complements to hyperscalers in multi-cloud strategies rather than full replacements. [1]

Price is the most visible difference. In its February 2025 comparison Uptime found that list prices for a comparable Nvidia H100-based instance were about two-thirds lower at neoclouds than at hyperscalers, and argued that most of the gap reflected differences in gross margin rather than in underlying cost. Equinix cites the same research and adds that neoclouds tend to emphasise GPU availability and fast scaling for training and experimentation. [1][2]

Hyperscalers have not stood still. AWS, for example, places its H100 and H200 instances in large, tightly networked clusters it calls UltraClusters and offers several purchasing options for them, from on-demand use to reservations of GPU clusters for fixed future dates. The practical difference for buyers is less about raw hardware and more about pricing structure, contract flexibility, surrounding services and how quickly capacity can actually be obtained. [3][4]

CoreWeave is the largest publicly listed example. Its 2025 results, published in February 2026, reported revenue of about 5.1 billion dollars, more than 850 megawatts of active power and roughly 3.1 gigawatts of contracted power at year end. Customers named in that release range from AI labs to software companies such as Cursor, Cognition and Midjourney. [5]

Nebius runs its own AI cloud and in September 2025 signed a multi-year agreement to provide dedicated GPU capacity to Microsoft from a new data center in Vineland, New Jersey. Lambda, which calls itself the Superintelligence Cloud, announced a multi-year agreement in November 2025 to deploy AI infrastructure for Microsoft using tens of thousands of Nvidia GPUs, including GB300 NVL72 systems. [7][8]

Other providers illustrate the range of models. Crusoe presents itself as a vertically integrated AI infrastructure company and built the Abilene, Texas campus that serves Oracle Cloud Infrastructure, with Nvidia GB200 racks arriving from June 2025. Together AI combines model inference services with self-service GPU clusters. Vultr, a general-purpose cloud with 32 regions, rents both AMD and Nvidia accelerators as on-demand virtual machines, bare metal and dedicated clusters. [9][10][11]

At the small end, neoclouds sell GPU-as-a-service by the hour. Equinix notes that this lets organisations treat AI infrastructure as an operating expense rather than a capital purchase, which suits projects whose return is still uncertain. Together AI, for instance, lets users create GPU clusters from a web console or command line in minutes and run them with Kubernetes or Slurm. [2][10]

At the large end, the business looks more like infrastructure leasing. CoreWeave reports a revenue backlog, which it defines as remaining performance obligations plus other amounts expected under committed customer contracts, of 66.8 billion dollars at the end of 2025. The Nebius and Lambda agreements with Microsoft follow a similar pattern: a single buyer takes dedicated capacity for several years. [5][7][8]

This mix creates concentration. CoreWeave's IPO filing showed Microsoft accounted for 62 percent of its 2024 revenue and a second customer for another 15 percent, and the company listed reliance on a small customer base as a key risk. [6]

Most neoclouds offer three broad tiers. On-demand capacity is billed hourly with no commitment. Reserved capacity trades an upfront or term commitment for a lower rate; Together AI, for example, sells reservations of 1 to 90 days paid in advance and allows reserved and on-demand nodes to be mixed in one cluster. Vultr similarly offers pay-as-you-go instances alongside reserved GPU servers. [10][11]

Beyond those published tiers sit privately negotiated contracts for whole clusters or data halls, typically running for years. These are the agreements that appear in company backlogs and financing announcements, and their terms are rarely published in detail. [5][7]

Buying GPUs at scale requires capital well ahead of revenue, and neoclouds have leaned heavily on debt. In August 2023 CoreWeave raised a 2.3 billion dollar facility that used Nvidia H100 hardware as collateral, which Quartz reports was the first time H100-based equipment had backed such a loan. By the end of 2024 CoreWeave carried 12.9 billion dollars of asset-backed debt, and its total debt reached 21.4 billion dollars at the end of 2025. [13][6][5]

Lenders increasingly look at the customer contract as much as the chips. In March 2026 CoreWeave closed an 8.5 billion dollar delayed-draw term loan secured by GPU infrastructure and an associated contract with an unnamed AI enterprise, rated A3 by Moody's, which the company described as the first investment-grade financing of that kind. Nebius said it would fund its Microsoft build-out partly with debt secured against that contract. [12][7]

Smaller players use similar tools. TechCrunch reported in August 2026 that Lambda secured about 1 billion dollars of short-dated private debt, arranged by JPMorgan, to buy Nvidia chips to be leased to Microsoft, after a separate secured credit facility in May 2026. [14]

The main concern raised by credit analysts is depreciation: GPUs lose value quickly as new generations arrive, so loans secured by them depend on the hardware earning enough before it becomes obsolete. Quartz describes these structures as unlike traditional corporate credit for that reason. [13]

For buyers, the practical questions are different. A neocloud that depends on one or two anchor tenants may have less spare capacity for smaller customers, or may reprioritise when a large contract lands. Conversely, the same financing model is what has allowed these firms to add capacity quickly. We think buyers should look past headline prices to the contract terms, cluster networking, support model and the provider's track record of actually delivering the capacity it lists.

Kovara tracks published GPU prices and availability signals across roughly 100 providers, including neoclouds, hyperscalers and marketplaces. Readers can compare list prices for the same GPU model across provider types on the GPU prices page, browse provider profiles and regions, ask Kova to explain how a specific provider sells capacity, or request a quote when a workload needs reserved or dedicated capacity that is not published. Because list prices do not guarantee capacity, we recommend confirming availability before planning around any single offer.