Neocloud Business Model and Unit Economics

A neocloud buys GPU servers with debt, houses them in rented data center space, and sells compute by the GPU-hour. It earns the spread between what it charges per hour and what it costs to own and operate each GPU: depreciation, colocation, power, and interest. Contract neoclouds like CoreWeave lock in multi-year deals at fixed rates. On-demand neoclouds like Lambda adjust prices to demand.

Contract pricing vs on-demand pricing

Every neocloud buys GPUs and sells hours. They differ in how they price those hours.

CoreWeave operates on multi-year take-or-pay contracts: the customer commits to a fixed number of GPU-hours at a fixed rate for 2-5 years and pays whether it uses the capacity or not. Nearly all of CoreWeave's revenue comes from these commitments, ^{[3]CoreWeave, S-1 registration statement filed with the SEC (March 2025)https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&company=coreweave&CIK=&type=S-1&dateb=&owner=include&count=40&search_text=&action=getcompany} and its backlog (contracted future revenue not yet recognized) hit $66.8 billion at the end of 2025, roughly 13x annual revenue. ^{[1]CoreWeave, Q4 2025 earnings and 10-K annual report (February 2026)https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Fourth-Quarter-and-Fiscal-Year-2025-Results/}

Lambda prices to market. In January 2024, Lambda raised its H100 SXM price from $2.59 to $3.49 per hour, a 35% increase. By August 2024, the price had dropped back to $2.99. As of February 2026, Lambda's published rate is $3.44 per hour. ^{[4]American Compute internal pricing data: 2,326 pricing observations across 8 providers and 10 GPU models, July 2022 through March 2026}

CoreWeave's pricing barely moves. Its H100 PCIe rate held at exactly $4.25 per hour across three years of tracking. ^{[4]American Compute internal pricing data: 2,326 pricing observations across 8 providers and 10 GPU models, July 2022 through March 2026} Its SXM rate did move once, from $4.76 to $6.16 in December 2025.

CoreWeave charges 1.6x Crusoe's rate and 1.8x Lambda's. The premium buys guaranteed multi-year capacity and service-level agreements backed by a $66.8 billion contract base.

Outside the neocloud market, several other business models compete for the same GPU buyer. RunPod ($2.69/hr for H100 SXM) aggregates capacity from multiple sources into a single cloud platform. Vast.ai ($1.60/hr) runs a peer-to-peer marketplace where anyone with GPUs can list them for rent, pricing set by individual sellers. SF Compute ($1.35/hr for committed volume) brokers bulk deals between buyers and GPU owners. ^{[4]American Compute internal pricing data: 2,326 pricing observations across 8 providers and 10 GPU models, July 2022 through March 2026}

Revenue per server depends on two unknowns: the actual contract rate (which is lower than the published rate card) and GPU utilization. At published rates and 100% utilization, an 8-GPU server generates $241,000 per year at Lambda's $3.44/hr ($3.44 x 8 GPUs x 8,760 hours) or $432,000 at CoreWeave's $6.16/hr. No provider runs at 100%. CoreWeave does not disclose utilization rates, but its $5.1 billion in 2025 revenue across an estimated 250,000-400,000 GPUs implies a blended yield well below the rate card.

What a neocloud sells

CoreWeave is the largest public neocloud as of early 2026: $5.1 billion in 2025 revenue, up from $1.9 billion in 2024. ^{[1]CoreWeave, Q4 2025 earnings and 10-K annual report (February 2026)https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Fourth-Quarter-and-Fiscal-Year-2025-Results/} Nebius (NBIS) and Applied Digital (APLD) also trade publicly but at smaller scale. Lambda, Crusoe, and Voltage Park remain private and do not disclose revenue.

All buy NVIDIA GPU servers from OEMs like Dell, HPE, and Supermicro, and house them in third-party data centers. As of early 2026, a neocloud sells GPU compute hours and not much else: no managed AI services, no model hosting, no databases. Sometimes orchestration tooling (Kubernetes, Slurm) sits on top, but the core product is closer to an infrastructure lease than a software platform. AWS started the same way in 2006: just EC2 and S3, raw compute and storage. The hundreds of managed services came later. Whether neoclouds follow the same path or stay as bare-metal providers depends on whether they can build software margins on top of hardware margins.

What it costs to run a GPU

Depreciation outweighs every other cost.

An 8-GPU H100 SXM server from an OEM like Dell or Supermicro costs roughly $250,000-$400,000 as of early 2026, depending on configuration. CoreWeave depreciates servers over 6 years using straight-line depreciation. ^{[3]CoreWeave, S-1 registration statement filed with the SEC (March 2025)https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&company=coreweave&CIK=&type=S-1&dateb=&owner=include&count=40&search_text=&action=getcompany} Lambda uses 5 years. Nebius uses 4. ^{[5]SiliconAngle, "Resetting GPU Depreciation: Why AI Factories Bend, But Don’t Break, Useful Life Assumptions" (November 2025)https://siliconangle.com/2025/11/resetting-gpu-depreciation/}

At 6-year depreciation, a $300,000 server costs $50,000 per year. At 4 years, the same server costs $75,000, a gap of $25,000 per server per year. A fleet of 1,250 servers (10,000 GPUs) on a 4-year schedule reports $31 million more in annual depreciation expense than the same fleet on 6 years.

Neoclouds lease rack space from data center providers like Equinix, QTS, and CyrusOne, typically on 5-10 year terms. GPU clusters draw 40-80 kW per rack, 3-8x what traditional enterprise IT requires. ^{[10]Eaton, "2025 Data Center Progress Report" (2025)https://www.eaton.com/content/dam/eaton/digital/eaton-data-center-segment-report-2025-whitepaper-wp152032-en-sg.pdf} Purpose-built AI colocation is scarce because of this density. Crusoe deploys GPU clusters at stranded natural gas sites, converting gas that would otherwise be flared into electricity for computing. ^{[11]Crusoe Energy, "Bringing Computing to Where the Energy Happens"https://crusoe.ai/energy/}

Operations costs (technicians, spare GPU inventory, RMA logistics) round out the rest. All of these costs combined, everything except depreciation and interest, stay well below 40% of revenue. CoreWeave's adjusted EBITDA margin of 60% means 60 cents of every dollar earned is cash profit before accounting for depreciation and interest.

Interest on debt is the other major cost. CoreWeave carried $21.4 billion in debt at the end of 2025 ^{[1]CoreWeave, Q4 2025 earnings and 10-K annual report (February 2026)https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Fourth-Quarter-and-Fiscal-Year-2025-Results/} and paid $1.2 billion in interest for the year. ^{[1]CoreWeave, Q4 2025 earnings and 10-K annual report (February 2026)https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Fourth-Quarter-and-Fiscal-Year-2025-Results/} Unlike every other cost on this list, interest payments come due whether the GPUs are running or not.

All figures below come from CoreWeave's Q4 2025 earnings press release: ^{[1]CoreWeave, Q4 2025 earnings and 10-K annual report (February 2026)https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Fourth-Quarter-and-Fiscal-Year-2025-Results/}

The gap between adjusted EBITDA ($3,093M) and operating loss (-$46M) comes to $3,139M. CoreWeave's cash flow statement breaks this out: $2,454M in depreciation and amortization, $630M in stock-based compensation, and roughly $55M in other adjustments. ^{[1]CoreWeave, Q4 2025 earnings and 10-K annual report (February 2026)https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Fourth-Quarter-and-Fiscal-Year-2025-Results/} Depreciation on GPU hardware dominates, consistent with $30.6 billion in net property and equipment on the balance sheet.

The gap between operating loss (-$46M) and net loss (-$1,167M) comes almost entirely from interest expense at $1.2 billion, the interest cost on $21.4 billion in debt.

CoreWeave keeps roughly 60% of revenue as “profit” but still reports a net loss. Each server generates cash from day one, but the depreciation and interest from the debt used to buy it exceed that cash in the early years. The net loss persists because CoreWeave is still deploying billions in new hardware to fill its $66.8 billion backlog. As long as it keeps growing, the company looks like Amazon in the 2000s: cash-flow positive on existing operations, unprofitable on paper because of reinvestment.

How neoclouds finance hardware

A 576-GPU cluster costs $35-36 million. Scaling to tens of thousands of GPUs requires billions.

Neoclouds draw on two sources of capital. Equity: venture capital, private equity, and public markets. Debt: asset-backed loans from private credit firms. (Raising debt for a cluster of your own? American Compute introduces operators to lenders with GPU appetite: see GPU financing.)

Lambda raised $480 million in a Series D at a $2.5 billion valuation in February 2025, followed by an additional $1.5 billion in November 2025 tied to a multibillion-dollar Microsoft deal. ^{[2]Reuters, "AI cloud startup Lambda raises $480 million in new round, Nvidia among investors" (February 2025)https://www.reuters.com/technology/ai-cloud-startup-lambda-raises-480-million-new-round-2025-02-06/ [6]TechCrunch, "Lambda raises $1.5 billion following multibillion-dollar Microsoft deal" (November 2025)https://techcrunch.com/2025/11/lambda-raises-1-5-billion/} OpenAI's Stargate project, which Crusoe is helping build and operate, secured $11.6 billion in funding in May 2025. ^{[7]Reuters, "OpenAI’s biggest data center secures $11.6 billion funding" (May 2025)https://www.reuters.com/technology/openais-biggest-data-center-secures-116-billion-funding-2025-05/} CoreWeave raised $7.5 billion in debt from Blackstone and other private credit firms in 2024 ^{[8]CNBC, "CoreWeave raises $7.5 billion in debt financing" (May 2024)https://www.cnbc.com/2024/05/coreweave-debt-financing/} before its March 2025 IPO.

The debt is asset-backed. Collateral is the GPU hardware itself plus contracted future revenue. Private credit lenders evaluate three things when underwriting a neocloud:

1. Contracted backlog (compute offtake). Signed customer commitments that guarantee future revenue. CoreWeave's $66.8 billion backlog lets it borrow at this scale.
2. Utilization rate. The percentage of installed GPUs generating revenue. Idle GPUs depreciate without generating income to cover loan payments.
3. GPU residual value. What the hardware will be worth when the loan matures. If next-generation GPUs collapse resale prices, the collateral behind the loan loses value.

Long-term contracts support more borrowing. More borrowing funds more GPU procurement. More GPUs support more contracts. CoreWeave's backlog is both future income and the foundation of its ability to raise debt.

The same mechanism runs in reverse. If contracts are not renewed, borrowing capacity shrinks. If GPU resale values drop, collateral loses value. And loan payments are fixed. CoreWeave's $1.2 billion in annual interest expense comes due whether utilization is at 95% or 50%.

What threatens the model

On-demand pricing is unpredictable. Lambda's H100 SXM rate swung from $2.59 to $3.49 to $2.99 to $3.44 between August 2023 and February 2026. ^{[4]American Compute internal pricing data: 2,326 pricing observations across 8 providers and 10 GPU models, July 2022 through March 2026} An on-demand neocloud financing hardware over 5-6 years cannot reliably forecast what rates will look like in year 3.

Customer concentration. Microsoft accounted for 62% of CoreWeave's revenue in 2024 and 67% in 2025. ^{[1]CoreWeave, Q4 2025 earnings and 10-K annual report (February 2026)https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Fourth-Quarter-and-Fiscal-Year-2025-Results/ [3]CoreWeave, S-1 registration statement filed with the SEC (March 2025)https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&company=coreweave&CIK=&type=S-1&dateb=&owner=include&count=40&search_text=&action=getcompany} Microsoft is also building its own AI accelerator, Maia, which could reduce its need for external GPU capacity.

Next-generation hardware. NVIDIA ships a new GPU architecture roughly every two years. Each generation improves performance per watt but also increases power draw per GPU. A B200 draws up to 1,000W per GPU compared to 700W for an H100 SXM. Many existing colocation facilities cannot deliver that power density without infrastructure upgrades, and those upgrades take months to years. A neocloud that wants to refresh its fleet needs both the capital for new servers and a facility that can support them. A neocloud that stays on current-generation hardware keeps its facility but risks losing customers to providers with newer GPUs.

Hyperscaler competition. AWS cut P5 (H100) on-demand instance pricing by up to 45% effective June 2025. ^{[9]AWS News Blog, "Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances" (June 2025)https://aws.amazon.com/blogs/aws/announcing-price-reduction-nvidia-gpu-instances/} Google and Amazon are deploying proprietary accelerators (TPUs, Trainium) that bypass NVIDIA entirely. Hyperscalers can afford to sell GPU compute at a loss because they make money on storage, networking, and software. Neoclouds only sell GPU hours.

Power. The bottleneck has shifted from GPU supply to electricity. Data center facilities with adequate density for GPU clusters are scarce. A neocloud that owns GPUs but cannot find a facility with enough power and cooling to run them watches those GPUs depreciate while generating zero revenue.

Why neoclouds exist

Hyperscalers cannot build data centers fast enough. Microsoft announced $80 billion in AI data center CapEx for fiscal 2025. Amazon announced $100 billion for 2025. Google announced $75 billion. Demand is still growing faster than they can add capacity. CoreWeave's $66.8 billion backlog, much of it from the same hyperscalers that operate their own clouds, shows the gap. Even with that buildout, the top five hyperscalers still only directly own about half of all clustered Blackwell GPUs, with anchor neoclouds like CoreWeave holding most of the rest.

Neoclouds deploy faster. Building a data center takes 18-30 months, but a neocloud can lease existing space, rack servers, and bring capacity online in weeks.

References

1. CoreWeave, Q4 2025 earnings and 10-K annual report (February 2026)
2. Reuters, "AI cloud startup Lambda raises $480 million in new round, Nvidia among investors" (February 2025)
3. CoreWeave, S-1 registration statement filed with the SEC (March 2025)
4. American Compute internal pricing data: 2,326 pricing observations across 8 providers and 10 GPU models, July 2022 through March 2026
5. SiliconAngle, "Resetting GPU Depreciation: Why AI Factories Bend, But Don’t Break, Useful Life Assumptions" (November 2025)
6. TechCrunch, "Lambda raises $1.5 billion following multibillion-dollar Microsoft deal" (November 2025)
7. Reuters, "OpenAI’s biggest data center secures $11.6 billion funding" (May 2025)
8. CNBC, "CoreWeave raises $7.5 billion in debt financing" (May 2024)
9. AWS News Blog, "Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances" (June 2025)
10. Eaton, "2025 Data Center Progress Report" (2025)
11. Crusoe Energy, "Bringing Computing to Where the Energy Happens"