What Is a Neocloud? How GPU Clouds Work, Compete and Get Financed
Key Takeaways
- A neocloud is a cloud provider specialized in GPU compute for AI, often built on owned hardware, with bare-metal access and simple per-GPU-hour pricing.
- Neoclouds sell a narrow GPU product below hyperscaler prices, and hyperscalers are often among their biggest customers.
- Most operators outside the headline names deploy into colocation and finance each cluster against the customer contract it serves.
- Neoclouds follow three business models: owning data centers, deploying into colocation, or running asset-light marketplaces.
- Neocloud compute is priced like a commodity, so access to capital and the quality of customer contracts decide which operators scale.
Introduction
A neocloud is a cloud provider built around GPUs for AI. The term, sometimes written neo cloud, covers companies that install accelerated servers in data centers and rent the compute to AI labs and enterprises. AI neoclouds grew out of a shortage, as demand for training and inference capacity outran what the largest clouds could supply.
That shortage has made neoclouds some of the fastest-growing companies in infrastructure. It has also made them some of the most capital-hungry, because every new cluster starts as a large hardware purchase.
This guide is written for the operator side of the market: founders, CFOs and infrastructure leads deciding what to deploy next. It explains what a neocloud is, how it differs from a hyperscaler, who the leading providers are and which business models work. It closes with how neoclouds finance GPU capacity and the risks that shape every deployment decision.
What Is a Neocloud?
A neocloud is a cloud provider that specializes in GPU compute for AI training, inference and high-performance computing, sold as GPU-as-a-Service. Neoclouds offer bare-metal access to current accelerators, high-speed interconnects between servers and simple per-GPU-hour pricing. Many own the hardware they rent out, and their revenue comes from selling that compute to customers.
ABI Research defines a neocloud as "a new generation of cloud infrastructure provider that specializes in High-Performance Computing (HPC) and AI-focused workloads." Uptime Institute describes these providers as focused on GPU-backed servers, often priced below the hyperscalers.
Many neoclouds own their hardware and earn back the cost through customer contracts, though ABI also describes asset-light neoclouds that rent capacity from others. The ownership model describes many smaller GPU-as-a-Service operators that sell capacity to a handful of customers.
Most neoclouds share five traits:
- GPU-first hardware: Clusters are built around current data center GPUs, with CPUs and storage sized to keep them busy.
- High-speed interconnects: InfiniBand or high-bandwidth Ethernet links servers so many GPUs can train one model together.
- Bare-metal access: Customers run workloads directly on the servers, with no virtualization layer taking a share of performance.
- Simple pricing: Capacity is sold per GPU-hour, on demand or under reserved contracts, with few add-on charges.
- Narrow product range: A short list of compute, storage and networking products stands in for a hyperscaler's broad service catalog.
The word came into wide use after the launch of large language models pushed AI companies toward specialist GPU providers. CoreWeave and Crusoe, began as digital-asset mining ventures. Those ventures brought industrial land and access to low-cost power into AI compute.
CoreWeave is a neocloud and the best-known public example and Microsoft is its top customer.
Neocloud Vs. Hyperscaler: Key Differences
Hyperscalers are the three largest cloud providers: Amazon Web Services, Microsoft Azure and Google Cloud. In Q2 2026, Synergy Research Group reported worldwide cloud infrastructure shares of 28% for Amazon, 20% for Microsoft and 15% for Google. Synergy put enterprise spending on these services at $143 billion for the quarter.
The same release lists CoreWeave, Nebius, Crusoe and Nscale among the fastest-growing tier-two providers. AI hyperscalers and neoclouds both sell GPU capacity, yet they are built for different jobs.
| Dimension | Neocloud | Hyperscaler |
|---|---|---|
| Focus | GPU compute for AI training, inference and HPC | General-purpose cloud, with AI as one of many services |
| Hardware | Homogeneous clusters of current-generation GPUs | Mixed estate of CPUs, GPUs and custom accelerators |
| Product range | Compute, storage and networking with a thin software layer | A broad catalog of managed services, databases and developer tools |
| Pricing | Per GPU-hour, with few add-on charges | Per instance-hour, with discounts tied to committed spend |
| Customer base | AI labs, AI-native startups, enterprises and other clouds | Enterprises, governments and developers of every size |
| Contract style | Reserved capacity contracts, often multi-year, plus on-demand | Self-serve on-demand, savings plans and enterprise agreements |
| Regions | Fewer sites, concentrated where power is available | A broad global footprint of regions |
The two groups are partners as well as rivals. CoreWeave's FY2025 annual report names Microsoft, a hyperscaler, as its top customer, at about 67% of revenue.
For a technology buyer, a neocloud works as complementary capacity in a multi-cloud plan. Training runs and inference clusters can sit with a neocloud while data and applications stay with a hyperscaler. For an operator, hyperscaler demand is both a source of large contracts and a concentration risk.
Neocloud Pricing: Why GPU Clouds Cost Less
Price is one of the main reasons customers choose a neocloud. At October 2026 list prices, an on-demand H100 averaged about $10 per GPU-hour across AWS, Google Cloud and Azure. Across CoreWeave, Nebius and Lambda, the average was about $4.90, roughly half the hyperscaler rate.
| Provider | Type | On-demand price per H100-hour (8-GPU instance) |
|---|---|---|
| Azure (ND96isr H100 v5) | Hyperscaler | $12.29 |
| Google Cloud (a3-highgpu-8g) | Hyperscaler | $11.06 |
| AWS (p5.48xlarge) | Hyperscaler | $6.88 |
| CoreWeave (HGX H100) | Neocloud | $6.16 |
| Nebius (HGX H100) | Neocloud | $4.50 |
| Lambda (8× H100 SXM) | Neocloud | $3.99 |
US list prices as of October 2026, compiled by IntuitionLabs. Reserved and committed contracts are priced separately, and on-demand rates change often.
Prices vary widely within each group. AWS has cut its H100 prices and now sits close to CoreWeave, while Google Cloud and Azure remain near the top. Egress fees, storage and minimum commitments also change what a buyer ends up paying.
Most of the gap comes from gross margin. Neoclouds sell a narrow product line on uniform hardware, which keeps their offer simple and their prices low. Hyperscalers can hold higher prices because many of their customers find switching providers too costly and complex.
Lower cloud prices also raise the bar for owning hardware. The cheaper the rental option, the busier an owned cluster has to be to compete on cost per GPU-hour.
For operators, the lowest all-in cost per GPU-hour wins, and the cost of capital is one of the largest inputs. Financing terms become a competitive variable alongside power prices and utilization.
Why Neoclouds Are Growing So Fast
AI demand keeps pulling capacity toward neoclouds. According to ABI Research forecasts, neocloud GPU-as-a-Service revenue will reach nearly $250 billion by 2030, a 43% compound annual growth rate. ABI's forecast starts from $42 billion in 2025 and expects inference to make up about 80% of the neocloud market by 2030.
The capital required across the industry is far larger. McKinsey estimates that data centers equipped for AI processing will require $5.2 trillion in capital expenditures by 2030 in its central scenario. Its constrained and accelerated scenarios put the figure at $3.7 trillion and $7.9 trillion, with about 60% going to chips and computing hardware.
USD.AI's view from its deal flow is that neoclouds win a share of that spend by adding capacity where power already exists. Gigawatt campuses need permitting, power and construction before they host GPUs. A neocloud can install a cluster in existing colocation space and start selling compute once the servers are online.
Deal flow at USD.AI shows who is seeking that capital. As of September 2026, USD.AI had reviewed over 700 GPU financing deals with about $25 billion of loan demand, and neoclouds make up the majority of qualified borrowers.
Those operators belong to a much larger group of companies that own compute, most of them well below the headline names. For most of them, the limit on growth is the capital they can put into new GPUs.
Leading Neocloud Companies
A handful of providers dominate coverage of the neocloud market. The table below summarizes how ABI Research's neocloud profiles describe seven of them, and the order carries no ranking.
| Company | How ABI Research describes it |
|---|---|
| CoreWeave | Offers customers early access to cutting-edge GPUs in dense, liquid-cooled clusters |
| Nebius | Takes a partner-first approach, with a large agreement with Microsoft |
| Lambda | A developer-first brand with transparent pricing |
| Crusoe | Runs energy-focused data centers |
| Nscale | Building large-scale deployments in Europe and the United States |
| Vultr | An independent cloud provider with more than 32 sites |
| Civo | Emphasizes affordability and a developer-first culture |
Below these names sits the Compute Middle Market. In USD.AI's deal flow, these are GPU cloud operators running one cluster or a few sites, selling GPU-hours to the same kinds of customers. Many sign their capacity to one anchor customer under a multi-year contract.
The larger providers raise billions in public and private markets. A middle-market operator with one signed contract has a different problem. In USD.AI's view, that operator needs financing sized and structured around a single cluster and its contract.
The middle-market operators USD.AI sees add capacity in smaller steps, often one cluster per signed contract. When each cluster can be financed against its own contract, these operators compete on price with far larger providers.
Neocloud Business Models
Neoclouds differ most in what they own. ABI Research sorts them into three business models, and each makes a different trade between capital intensity, speed to capacity and control.
| Model | What the operator owns | Capital intensity | Speed to capacity | Control |
|---|---|---|---|---|
| Own data centers | Land, buildings, power infrastructure and GPUs | Highest | Slowest, since permitting, power and construction come first | Full control of power, cooling and site design |
| Colocation | GPUs, servers and networking in leased data center space | Moderate, with capital going into hardware | Fast, since power and space already exist | Control of the hardware, with the facility running power and cooling |
| Asset-light or marketplace | Software and customer relationships, with GPUs owned by others | Lowest | Limited by partners' supply | Little control over hardware or availability |
Operators that own their data centers capture the facility margin and control their power, which matters as racks grow denser. They also carry the most capital and the longest lead times. Some data center operators also own the GPUs they sell compute from, which puts them in both groups at once.
Colocation is the core model of the Compute Middle Market. The operator buys the GPUs, leases space and power in an existing facility and signs customers for the capacity. In USD.AI's deal flow, middle-market clusters of 1–10 MW deploy into existing colocation this way, so the operator's capital goes into revenue-producing hardware.
Asset-light providers and marketplaces aggregate capacity owned by others and resell it. They need the least capital, but their margins depend on partners' pricing and they have no hardware of their own to finance. When a marketplace doesn't own the GPUs, the hardware owner is the party that finances them.
How Neoclouds Finance GPU Capacity
GPUs are the largest cost in a neocloud deployment, and equity is the most expensive way to pay for them. Funding every cluster with equity dilutes owners and raises the operator's cost per GPU-hour against competitors that use debt.
The largest neoclouds rely heavily on debt. CoreWeave's annual report for FY2025 states that its total indebtedness was $21.6 billion at 31 December 2025. CoreWeave raised part of that debt through delayed-draw term loans at wholly owned subsidiaries.
Middle-market operators need the same tool at a smaller scale: asset-backed GPU financing underwritten on contracted cash flow. The GPUs secure the loan, and the customer contract repays it. Contract tenor, counterparty strength and pricing matter as much as the hardware itself.
GPU financing through USD.AI applies one standardized structure to every deal, for new deployments and for GPUs already in service. A bankruptcy-remote special-purpose vehicle holds the GPUs, the customer contracts and the revenue accounts. Capital is held in escrow until the servers are independently verified as installed, and servicing stays in place for the full term of the loan.
| Term | Financing through USD.AI |
|---|---|
| Loan-to-value | Up to 80% of the deployment cost, with the operator contributing at least 20% as equity |
| Rate | Fixed or floating |
| Term and amortization | 3–5 years, with straight-line monthly amortization |
| Reserve | Three-month debt-service reserve funded at closing |
| Size | From about $5 million, with larger facilities available |
| Collateral location | Tier III or Tier IV data centers |
Every deal is scored by the Compute IQ underwriting system against published underwriting standards on a 100-point scale. Compute IQ assesses the operator, the customer and the contract, and the score sets financeability, the pricing tier and the loan's size and structure. Every operator is assessed on the same evidence, so the terms follow from the deployment itself.
Every loan's hardware carries residual-value insurance, designed to reduce loss severity if the GPUs sell below their insured value. The lender group holds a first-priority security interest in the hardware.
Consider an operator adding NVIDIA B300 deployments in colocation under a 36-month contract with an AI company. The contract runs for the full loan term, so contracted revenue is in place for every scheduled payment. The GPUs stay deployed and earning while they secure the loan, and the operator keeps ownership of the hardware.
Challenges Facing Neocloud Providers
Power comes first. The International Energy Agency projects global data center electricity use will double to around 945 TWh by 2030 in its Base Case. That compares with about 415 TWh in 2024, and the IEA notes substantial uncertainty, with its Headwinds case plateauing near 700 TWh.
On density and cooling, the Uptime Institute 2026 survey found the average of operators' most typical rack densities passed 11 kW across all data centers. AI racks run far higher, and Uptime Institute research on AI factories puts current rack-scale GPU systems at 130 kW and above. Operators need facilities with the power delivery and liquid cooling to match.
GPU depreciation is the hardest risk to measure. Uptime Institute's 2026 survey notes that shorter refresh cycles mean older equipment is likely to lose value more quickly.
CoreWeave's annual report shows it extended the useful life of its computing equipment from five to six years in 2023. That is an accounting estimate and says little about resale value.
Customer concentration is common at every scale. CoreWeave's annual report attributes about 67% of its FY2025 revenue to Microsoft. Many operators in USD.AI's deal flow also rely on one anchor customer, a risk that falls when the contract runs as long as the loan.
Commodity pricing squeezes operators that sell on-demand capacity. As new supply arrives, spot prices can fall faster than debt service, while reserved contracts lock in pricing for their full tenor.
Regional concentration adds another layer. The International Energy Agency expects the US and China to account for nearly 80% of data center electricity growth to 2030. An operator tied to one market carries that market's power, permitting and pricing risk.
Each of these risks has a practical response that shows up in how a deployment is planned and financed.
| Challenge | What it puts at risk | What reduces it |
|---|---|---|
| Power availability | The date a cluster can start earning | Colocation space with energized capacity before the hardware ships |
| Density and cooling | Whether the facility can run current GPUs at full load | Facilities built for liquid cooling and high-density racks |
| GPU depreciation | Collateral value if the hardware has to be sold | Amortization that pays down principal as the hardware ages, plus residual-value cover |
| Customer concentration | Revenue if the anchor customer leaves | Contracts with strong counterparties that run as long as the loan |
| Commodity pricing | Margins on on-demand capacity | Reserved contracts that fix pricing for the full tenor |
| Regional concentration | Exposure to one market's power and permitting | Sites in more than one market as the operator grows |
Conclusion
A neocloud is a GPU cloud that sells compute to AI customers, usually below hyperscaler prices, and many own the hardware they sell. The market is growing fast, and neoclouds make up the majority of qualified borrowers in USD.AI's deal flow. Because neocloud compute is priced like a commodity, capital and contract quality decide which operators scale.
If you're planning a deployment, line up three things first. You need a signed customer contract whose tenor matches the loan and colocation space with power in place. You also need the equity for your share of the deployment.
USD.AI's neocloud financing page explains how financing through USD.AI fits a neocloud deployment. Once the contract, space and equity are in place, talk to USD.AI about financing your next cluster.
FAQ
These answers cover the questions operators and buyers ask most often about neoclouds.
What is a neocloud?
A neocloud is a cloud provider specialized in GPU compute for AI training, inference and high-performance computing. Many own their GPU clusters, offer bare-metal access and high-speed interconnects, and sell capacity per GPU-hour on demand or under reserved contracts.
Is CoreWeave a neocloud?
Yes, CoreWeave is a neocloud and the best-known public example.
What is the difference between a neocloud and a hyperscaler?
A neocloud specializes in GPU compute with a narrow product line, while hyperscalers such as AWS, Azure and Google Cloud offer broad service catalogs. Neoclouds usually price GPUs lower, and some hyperscalers are also neocloud customers.
Why are neocloud GPUs cheaper than hyperscaler GPUs?
Uptime Institute's February 2025 benchmark found on-demand DGX H100 capacity at $34 an hour from neoclouds and $98 from hyperscalers. Uptime Intelligence believes most of the gap reflects gross margin, since neoclouds run a narrow product line on homogeneous hardware.
What are the biggest risks for neocloud operators?
The biggest risks are power availability, rack density and cooling, GPU depreciation, customer concentration and commodity pricing. Matching customer contract tenor to loan tenor reduces the revenue risk behind a GPU deployment.
How much can a neocloud finance through USD.AI?
Financing through USD.AI covers up to 80% of the deployment cost, with the operator contributing at least 20% as equity. Financing starts from about $5 million, with larger facilities available.
