In this note
- What is a neocloud?
- Neocloud vs hyperscaler vs GPU marketplace
- Neocloud business model: how neoclouds make money
- Neocloud unit economics: price, utilization and cost
- What it takes to build a neocloud
- What a neocloud owns and what it can outsource
- The main risks for a neocloud
- Questions to answer before building a neocloud
- How Lightyard works with neoclouds
A neocloud is a cloud provider whose main business is renting out GPU compute for AI, mostly NVIDIA GPUs, by the hour or on multi-year contracts. A hyperscaler sells hundreds of services and GPUs are one of them. A neocloud's product is the GPU cluster itself: the servers, the network between them, storage and the software tenants use to run jobs. CoreWeave, Nebius, Lambda and Crusoe are the examples analysts name most often.
Revenue is GPU-hours billed times the price per hour. Most of the cost is fixed the day the hardware is bought, and most of that hardware is financed with debt. Whether a neocloud works depends on how many hours it bills, at what price, for how long and what its capital costs.
Numbers come from filings, research and press, linked where used. Where we compute something, we say so and show the assumptions.
What is a neocloud?
SemiAnalysis (opens in a new tab) defines an AI neocloud as "a new breed of cloud compute provider focused on offering GPU compute rental." Synergy Research Group (opens in a new tab) describes neoclouds as "purpose-built to deliver high-performance, GPU-centric computing for artificial intelligence workloads", selling GPU as a service (GPUaaS), generative AI platforms and high-density data center capacity.
SemiAnalysis sorts the companies that sell GPU time into four groups:
- Traditional hyperscalers, such as AWS, Microsoft Azure, Google Cloud and Oracle
- Neocloud giants: it names Crusoe, Nebius, Lambda and CoreWeave
- Emerging neoclouds, including regional sovereign AI providers
- Brokers, platforms and aggregators, which don't own GPUs
The segment is young and growing fast. Synergy puts neocloud revenue above US$25 billion in 2025, with the fourth quarter at US$9 billion, up 223% on the year. It forecasts close to US$400 billion by 2031. McKinsey counts over 100 neoclouds, of which only 10 or 15 operate at "meaningful scale" in the US, as reported by InfoWorld (opens in a new tab).
Neocloud vs hyperscaler vs GPU marketplace
Three kinds of company sell GPU-hours, and all three get called GPU cloud providers. They differ in what they own and what else they sell.
| Hyperscaler | Neocloud | GPU marketplace | |
|---|---|---|---|
| What it sells | GPUs as one product among compute, storage, databases and AI platforms | GPU clusters: bare metal or virtual machines, the fabric, storage, schedulers, sometimes inference endpoints | Other operators' GPUs, listed and matched on price |
| Owns the GPUs | Yes | Yes, mostly financed with debt | No |
| Main way of selling | On demand, with one- to three-year commitments | Multi-year take-or-pay contracts for most revenue, plus on demand | On demand and short term |
| Median on-demand H100 list price, October 2026 | US$10.53 per GPU-hour | US$3.95 per GPU-hour | Not in the index |
Uvik's index (opens in a new tab) of on-demand list prices from 19 providers, captured on 7 October 2026, found hyperscalers charging 1.7 to 2.7 times the GPU-cloud price for the same GPU, median against median. Commitments narrow it: three-year terms at AWS, Google Cloud and Microsoft Azure cut H100 prices by 56% to 57%.
The Uptime Institute (opens in a new tab) calls neoclouds "primarily commodity players" that "predominantly match supply with demand, pricing their offerings inexpensively to attract volume", while hyperscalers add infrastructure, platform and software services on top. McKinsey's view, reported by The Register (opens in a new tab), is that renting out GPUs is "inherently commoditized".
Hyperscalers are also neocloud customers. Microsoft was about 67% of CoreWeave's 2025 revenue, according to its 10-K (opens in a new tab). In September 2025 Nebius signed a five-year agreement (opens in a new tab) to give Microsoft dedicated GPU capacity, worth about US$17.4 billion through 2031.
GPU marketplaces don't own GPUs; SemiAnalysis calls them capital-light intermediaries. They suit short jobs and price shopping, not a large cluster on one fabric for years.
Neocloud business model: how neoclouds make money
A neocloud sells GPU-hours in two ways, and the mix sets most of its business model.
On-demand GPU pricing
On demand, a customer pays by the hour and can leave at any time. It's the highest price per hour, earned only for hours someone uses. CoreWeave's price page (opens in a new tab) lists an 8-GPU HGX B200 instance in North America at US$68.80 an hour, or US$8.60 per GPU-hour. Across seven providers, Uvik found a median B300 list price of US$9.50 per GPU-hour, with a range of US$7.40 to US$17.80. Uvik notes that committed contracts, spot capacity and marketplaces all come in below those list prices.
Reserved and take-or-pay contracts
On a committed contract the customer reserves a block of capacity for a term and pays whether it uses it or not. CoreWeave's 10-K describes customers buying "a specified amount of capacity on a take-or-pay basis over the contract term." Committed contracts were over 98% of its 2025 revenue, with a weighted-average duration of about five years. Its remaining performance obligations, revenue contracted but not yet recognized, were US$60.7 billion at the end of 2025, against US$15.1 billion a year earlier.
Customers get a lower rate for committing. CoreWeave offers "up to 60% discounts" off on-demand prices for committed usage. In April 2026 SemiAnalysis (opens in a new tab) wrote that neoclouds prefer terms of at least a year, with two or three years preferred, that five years is the most popular tenor for long-term deals and that some labs were agreeing to prepayments above 20% on deals of four years or more.
The neocloud gives up price for three things. Every contracted hour is billed. Revenue is known years ahead. And lenders will lend against it, which is what pays for the next cluster.
What moves the price
GPU-hour prices follow supply. SemiAnalysis's one-year H100 contract index rose from US$1.70 per GPU-hour in October 2025 to US$2.35 in March 2026, and it reported H100s being renewed at the rates they were signed at two to three years earlier. A neocloud's on-demand book reprices every day; its contract book reprices at renewal.
Neocloud unit economics: price, utilization and cost
Most of a neocloud's cost is fixed once the cluster is bought and the lease is signed. So the unit economics come down to revenue per productive GPU-hour (the price times the share of hours billed) against the all-in cost of that hour.
The tables below are our estimate. Costs come from our five-year model for a 32-node HGX B300 cluster, laid out line by line in buy vs rent a GPU cluster; prices come from the sources above.
| Assumption | Value | Basis |
|---|---|---|
| All-in cost per productive GPU-hour, 32 HGX B300 nodes in colocation, five years | US$3.53 low, US$5.10 middle, US$7.23 high | Hardware, network, storage, colocation, power, support, team, insurance, residual value and cost of capital (8% a year in the middle case); 9 months to first revenue and 2% of hours lost to failures in the middle case |
| On-demand list price, B300 | US$9.50 median, US$7.40 lowest | Uvik, 7 October 2026 |
| Discount for a multi-year commitment | 40%, 50% or 60% off the median list price | Scenarios; CoreWeave advertises up to 60%, and three-year H100 commitments at three hyperscalers are 56% to 57% off |
| Hours billed | 50% to 70% on demand; every contracted hour on take-or-pay | Scenarios, not observed rates |
| How the hours are sold | Price per GPU-hour | Hours billed | Revenue per productive GPU-hour | Margin at low cost | At middle cost | At high cost |
|---|---|---|---|---|---|---|
| On demand, median list price | 9.50 | 70% | 6.65 | 3.52 | 1.81 | -0.58 |
| On demand, median list price | 9.50 | 50% | 4.75 | 1.62 | -0.09 | -2.48 |
| On demand, lowest list price | 7.40 | 60% | 4.44 | 1.31 | -0.40 | -2.79 |
| Multi-year, 40% off the median | 5.70 | 100% | 5.70 | 2.57 | 0.86 | -1.53 |
| Multi-year, 50% off the median | 4.75 | 100% | 4.75 | 1.62 | -0.09 | -2.48 |
| Multi-year, 60% off the median | 3.80 | 100% | 3.80 | 0.67 | -1.04 | -3.43 |
All figures are US$ per GPU-hour, before sales, corporate overhead and platform development. Three things follow.
Utilization decides the on-demand case. At the median B300 list price, a cluster covers its middle-case cost once 51% of its hours are billed. At our low cost that falls to 33%. At our high cost it's 76%.
A long contract fixes the margin for the term. There's no upside from a busy month, so the contract price has to clear the all-in cost from the first day. At a 50% discount, a cluster with our middle-case costs breaks even for five years.
Cost decides who can sign long contracts. Our low case assumes cheaper hardware and capital, a faster start and a higher resale value. Scale helps as well: in our model the same build at 128 nodes comes to about US$4.58 per productive GPU-hour, mostly because the team is spread over more GPUs. An operator near our low case can sign at a 60% discount and keep a margin; one near our middle case can't. SemiAnalysis wrote in April 2026 that long-term deals can lock in "a teens project IRR in most cases".
At scale, financing is one of the largest lines. CoreWeave's 2025 revenue was US$5.1 billion and its net loss US$1.2 billion. Net interest expense was about US$1.2 billion, close to a quarter of revenue by our arithmetic, on total indebtedness of US$21.6 billion at the end of the year.
What it takes to build a neocloud
Building a neocloud means putting seven things in place and keeping them in step: capital, power, hardware, the network, software, the people who run it and the customers who pay for it. Together they're the neocloud infrastructure stack.
Capital
In our cost model, a 32-node HGX B300 cluster needs about US$19 million to US$24 million of capex, or about US$73,000 to US$94,000 per GPU once the network, storage, fit-out and spares are counted. The servers are only about a third of the five-year cost. The rest is space, power, support, people and capital.
The first clusters are usually paid for with equity. Growth after that runs on debt secured on contracts. Wing VC (opens in a new tab) counts more than US$20 billion of GPU-backed facilities announced and describes the usual structure: the GPUs and the assigned customer contracts sit in a bankruptcy-remote vehicle, and the loan is sized to be repaid before the hardware ever has to be sold. Pricing follows the customer's credit. Wing lists SOFR plus 2.25% for a facility backed by investment-grade offtakers, SOFR plus 4.5% for one backed by non-investment-grade customers and reported quotes of SOFR plus 8% to 9% for small neoclouds without investment-grade contracts.
Financing and sales depend on each other. The obligations under Microsoft's agreement with Nebius began only once Nebius confirmed it had secured any additional financing it needed for the capital spending.
Power and sites
Power is the scarce input. CBRE (opens in a new tab) put North American primary-market vacancy at a record low of 1.4% in the first half of 2026, with 7,481 MW under construction and 80.4% of it preleased. It names power availability and delivery timelines as the most decisive factors in site selection, and reports energization delays from permitting, planning and zoning.
Many neoclouds lease colocation space rather than build. CoreWeave's 10-K describes its cost of revenue as mainly data center costs, including rent and power. At the end of 2025 it operated 43 data centers with over 850 MW of active power, against about 3.1 GW contracted. The gap between power contracted and power energized is where much of a neocloud's schedule risk sits. Our deployment timeline covers how long each kind of site takes.
Hardware supply
GPUs come through NVIDIA's allocations and the server makers, and in a tight market the delivery date is part of the price. In April 2026 SemiAnalysis reported Blackwell lead times running into June and July and capacity coming online through August and September already booked. It also reported LPDDR5 and DDR5 contract prices heading toward roughly four and five times their year-earlier level, which pushed server costs up. A neocloud also needs spare parts on site, so a failed node doesn't wait on a return.
Network and storage
SemiAnalysis's 2024 playbook counts eight 400 Gb/s NICs per H100 server, 3,200 Gb/s per node, and shows how much the design choices weigh: in its examples, oversubscribing or simplifying the backend saved 24.9% to 34.8% of its cost, and the cheapest design performed worse under congestion. Next to the backend fabric sit a frontend network, a storage network and an out-of-band management network.
Multi-tenancy adds work a single-user cluster doesn't need: each tenant has to be isolated on the backend fabric too, and SemiAnalysis notes that InfiniBand partition keys alone aren't enough.
Software for tenants and billing
Tenants expect to get nodes, images, a scheduler (Slurm, Kubernetes or both), storage and monitoring without opening a ticket. The business needs tenants, quotas, contracts, metering and invoices that match what was sold. SemiAnalysis's ClusterMAX 2.0 ratings (opens in a new tab), which tracked 209 providers in November 2025, look for passive health checks that drain a node on events such as a lost GPU or ECC errors, and active tests that run around each job and weekly on idle nodes. A neocloud can build this stack or license it, but it can't sell capacity without it.
Operations and on-call
GPUs fail often enough that operations is a standing cost. Meta's Llama 3 paper (opens in a new tab) reports 419 unexpected interruptions in a 54-day snapshot of a 16K-GPU H100 training run, about 78% of them attributed to confirmed hardware issues. Someone has to detect each failure, drain the node, swap the part or the node and prove it healthy before a tenant gets it back, at any hour. ClusterMAX gives 99.9% for compute nodes and 99% for racks as example uptime targets for an SLA. Our SLA and goodput note covers failure rates and spare sizing.
Sales
The large buyers want large blocks. SemiAnalysis describes offtake deals of 50 or 100 MW, about 24,000 to 48,000 GB300 NVL72 GPUs. A few such customers can fill a neocloud, which is why concentration runs high: Microsoft was about 67% of CoreWeave's 2025 revenue, and OpenAI has committed to pay CoreWeave up to about US$11.9 billion through October 2030. For smaller players, McKinsey suggests growing with startups as a route to larger customers, per The Register.
What a neocloud owns and what it can outsource
A neocloud doesn't have to do all of this itself. What it should keep is what its customers buy and its lenders underwrite: the contracts, the hardware and the customer relationship.
| Layer | What a neocloud usually owns | What it can outsource | What to keep control of |
|---|---|---|---|
| Customers and contracts | Sales, pricing, the contract and the SLA | Some sales through partners | Terms and the customer relationship |
| GPUs, servers and network gear | Owned and financed | Wholesale capacity from another cloud, resold | Specification, acceptance tests and title |
| Building, power and cooling | Sometimes; many lease | Colocation: the building, power, cooling, security and remote hands | Power per rack, expansion rights, lease term against contract term |
| Cluster design and build | The design standard | Site search, design, sourcing, racking, cabling and burn-in, project by project | Acceptance criteria and dates |
| Day-to-day operations | Usually, once the fleet is large | Monitoring, failed-part replacement and on-call, to an SLA | Response and restore times, access to health data |
| Software | The layer tenants see | Licensed fleet and cloud software, open-source schedulers | Billing data and tenant isolation |
In our view the split that works for most is to own the customer, the contract and the hardware and to lease the building. While the team is small, each build can be bought as a project; operations can move in-house once the fleet keeps an on-call rotation busy. Reselling another cloud's capacity is the fastest way in, but then the hardware and the terms belong to someone else.
The main risks for a neocloud
Utilization and price
Uncommitted capacity earns only when it's billed, and the on-demand price can move by tens of percent within months, in either direction. A long contract removes that risk for its term and hands it back at renewal.
Financing
Debt is cheap only when the customer's credit is good, so a neocloud's cost of capital depends on who it signs. Wing VC argues that lenders' real collateral is contracted cash flow, and that the loans are built to amortize before anyone has to find out what a three-year-old GPU is worth. If a customer defaults or a contract ends early, the hardware has to earn elsewhere while the loan is still being repaid.
Depreciation
How long GPU servers should be depreciated over, and how long they actually keep earning, are both disputed. A shorter economic life than the plan assumes leaves less time to recover the capex. GPU depreciation and useful life compares the schedules in the filings with the evidence on how long GPUs earn.
Generation turnover
NVIDIA works on a one-year rhythm (opens in a new tab) for data center products, and each generation tends to pull down what customers will pay for an hour on the one before. Capacity under a five-year contract has its price fixed; capacity sold on demand competes with each new generation as it ships. So far older GPUs have held their rental value better than that cadence suggests, while capacity has been short. GB300 or wait for Vera Rubin works through the timing.
Customer concentration
A handful of large customers can be most of the revenue. That makes the contract book easier to finance and harder to replace. When a large contract ends without a renewal, the neocloud needs a new customer for the whole block at once.
Questions to answer before building a neocloud
- Who are the first customers, for how much capacity and on what term? Will they prepay?
- Where is the power and when will it be energized? Does the lease run as long as the customer contracts?
- When can the hardware be delivered, and at what price?
- What will the debt cost, with those contracts and without them?
- Who builds and accepts each cluster, and against which tests?
- Who responds to a failed node at night, and how long until a healthy one replaces it?
- What do tenants use to get nodes, run jobs and see their bill?
- What happens to the cluster when the next generation ships and the first contracts end?
How Lightyard works with neoclouds
Lightyard designs, sources, builds and operates NVIDIA GPU clusters on its own software. For a neocloud that covers the build side of the list above, and operations if you want them. You own the cluster and keep the customers; we don't sell its hours to anyone else.
With cluster deployment, we choose the site with you or find one, design the cluster, source GPUs and servers through our relationships with NVIDIA and server manufacturers, then build and burn it in, handing capacity over in monthly waves of 32 nodes. Most clusters are 32 to 128 nodes; larger ones are scoped case by case. With a managed cluster, our team also operates it on our software, under an SLA written into the contract. The neoclouds page has the details, or talk to us about your next cluster.
Questions this note answers
A cloud provider whose main business is renting out GPU compute for AI, mostly NVIDIA GPUs, by the hour or on multi-year contracts. GPUs are its core product rather than one service among hundreds. CoreWeave, Nebius, Lambda and Crusoe are the examples analysts name most often.
A hyperscaler sells hundreds of services and GPUs are one of them. A neocloud sells GPU clusters and little else, usually for less: an October 2026 index of on-demand list prices found hyperscalers charging 1.7 to 2.7 times the GPU-cloud price for the same GPU. Hyperscalers are also large neocloud customers.
By billing GPU-hours. At the large neoclouds most revenue comes from multi-year take-or-pay contracts priced below on-demand rates; CoreWeave's committed contracts were over 98% of its 2025 revenue. The rest is on-demand capacity, which pays more per hour but only for the hours someone uses.
Renting GPU compute instead of buying it: bare-metal servers, virtual machines or whole clusters, billed per GPU-hour or reserved for a term. It's the core product of a neocloud and one of many products at a hyperscaler.
Capital (mostly debt secured on customer contracts), powered data center space, a hardware allocation, a backend network with storage, software for tenants and billing, an operations team on call and customers who will commit. On our estimates a single 32-node HGX B300 cluster needs about US$19 million to US$24 million of capex.
Unbilled hours and falling prices on uncommitted capacity, the cost of debt, GPUs earning for less time than planned, each new NVIDIA generation and dependence on a few large customers. Long contracts reduce the first risk and make debt cheaper, at a lower price per hour.
Sources
- SemiAnalysis: AI Neocloud Playbook and Anatomy (October 2024) (opens in a new tab)
- SemiAnalysis: ClusterMAX 2.0 (November 2025) (opens in a new tab)
- SemiAnalysis: The great GPU shortage, rental capacity (April 2026) (opens in a new tab)
- Synergy Research Group: Neocloud market forecast to approach $400B by 2031 (April 2026) (opens in a new tab)
- Uptime Institute: Neoclouds, AI's shock absorbers (October 2025) (opens in a new tab)
- InfoWorld: Neoclouds vs. hyperscalers, citing McKinsey (August 2026) (opens in a new tab)
- The Register: Rent-a-GPU neoclouds need to adapt, citing McKinsey (November 2025) (opens in a new tab)
- Uvik: GPU price index, on-demand list prices captured 7 October 2026 (opens in a new tab)
- CoreWeave: pricing page (on-demand and committed discounts) (opens in a new tab)
- CoreWeave Form 10-K, 2025 (opens in a new tab)
- Nebius Form 6-K: agreement with Microsoft (September 2025) (opens in a new tab)
- Wing VC: How the GPU became collateral (September 2026) (opens in a new tab)
- CBRE: North America Data Center Trends H1 2026 (August 2026) (opens in a new tab)
- Meta: The Llama 3 Herd of Models, reliability section (opens in a new tab)
- NVIDIA: Computex 2024 keynote, one-year rhythm (opens in a new tab)