Starting a Neocloud in 2026
Starting a neocloud follows a predictable sequence: raise capital, lease space in a data center, order GPU servers, wait for delivery, rack and cable the hardware, onboard customers, and bill your first GPU-hour. Each stage has its own constraints: capital structure, hardware lead times, power density, staffing, and customer acquisition.
Neocloud launch timeline (click each stage for details)
What it costs to start
Hardware is the biggest cost. In bills of materials (BOMs) and deployment quotes we've reviewed, an 8-GPU server (B200 or B300 generation) lands in the $250,000-$400,000 range as of early 2026. [1]Based on our review of real BOMs and industry pricing (2025-2026) In those same quotes, GPUs account for roughly 60-70% of total hardware spend. Our estimate for a 576-GPU cluster (a group of servers networked together to work as one system), 72 servers with InfiniBand networking and supporting infrastructure, is roughly $36 million. [2]Based on review of multiple 576-GPU B300 cluster deployment quotes (early 2026) A smaller 16-GPU deployment (2 servers plus networking) starts around $1 million on the same basis. [1]Based on our review of real BOMs and industry pricing (2025-2026)
Hardware is only part of the capital need. On top of the hardware shown below, a neocloud also pays for:
- Colocation (renting space, power, and cooling in someone else's data center)
- Staff (infrastructure engineers, platform engineers, sales)
- Software and licensing (scheduling, monitoring, billing systems)
- Insurance (property, business interruption, transit)
These costs vary widely by provider, region, and scale, so we have not included them in the hardware estimates below.
Startup capital by scale
GPUs take 2 weeks to 3-6 months to arrive depending on allocation. [3]Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026) Once delivered, they need to be racked, cabled, tested, and connected. Customers need to be onboarded. Months can pass between signing a purchase order and generating the first dollar.
Lambda raised $480 million in equity in February 2025 [4]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/ and then $1.5 billion more in November 2025. [5]TechCrunch, "Lambda raises $1.5 billion following multibillion-dollar Microsoft deal" (November 2025)https://techcrunch.com/2025/11/lambda-raises-1-5-billion/ Nscale closed a $1.4 billion GPU-backed loan in February 2026. [6]Nscale, "$1.4bn Delayed Draw Term Loan Backed by GPUs" press release (February 2026)https://www.nscale.com/press-releases/nscale-signs-1-4bn-delayed-draw-term-loan Even at those fundraising levels, these companies are still scaling up.
How to raise the capital
Two sources fund neoclouds: equity (selling ownership to investors) and GPU-backed debt (borrowing from lenders with hardware as collateral). Most use both.
Equity comes from venture capital, private equity, or public markets. Lambda's $480 million Series D at a $2.5 billion valuation (February 2025) is a typical mid-stage raise. [4]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/ CoreWeave raised $1.1 billion in Series C equity at a $19 billion valuation in 2024, [7]CNBC, "Nvidia-backed GPU cloud provider CoreWeave is worth $19 billion" (May 2024)https://www.cnbc.com/2024/05/01/nvidia-backed-gpu-cloud-provider-coreweave-is-worth-19-billion.html then went public in March 2025.
Debt is typically GPU-backed lending, most commonly structured as private credit. CoreWeave's $2.3 billion facility in August 2023 was the first GPU-backed credit facility at this scale. [8]Reuters, "CoreWeave raises $2.3 billion in debt collateralized by Nvidia chips" (August 2023)https://www.reuters.com/technology/coreweave-raises-23-billion-debt-collateralized-by-nvidia-chips-2023-08-03/ That was followed by $7.5 billion from Blackstone and others in May 2024. [9]CNBC, "CoreWeave raises $7.5 billion in debt financing" (May 2024)https://www.cnbc.com/2024/05/coreweave-debt-financing/ The hardware itself serves as collateral: if the borrower defaults, the lender seizes the GPUs.
| Term | Illustrative range |
|---|---|
| Loan-to-value | 60-70% |
| Loan term | 2-4 years |
| Interest rate | 8-12% |
| Amortization | 25-30% per year |
| Collateral | GPU hardware + customer contracts |
Illustrative GPU-backed loan terms as of early 2026, based on market commentary and public deal disclosures rather than standardized published loan sheets. [10]AltStreet, "GPU Depreciation & Obsolescence Risk" reference guide (January 2026)https://altstreet.investments/reference/risk/gpu-depreciation
Contracted revenue makes borrowing easier. CoreWeave's $66.8 billion backlog (the total value of signed but not yet delivered contracts) helps explain why it has been able to borrow at this scale. [11]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's $500 million facility from Macquarie was structured as a GPU financing vehicle secured by the hardware and supported by its cash flow. [12]Reuters, "Lambda secures $500 mln loan with Nvidia chips as collateral" (April 2024)https://www.reuters.com/technology/lambda-secures-500-mln-loan-with-nvidia-chips-collateral-2024-04-04/
The firms lending at this scale include Blackstone, Apollo, Blue Owl, PIMCO, Magnetar, Macquarie, and Brookfield. [9]CNBC, "CoreWeave raises $7.5 billion in debt financing" (May 2024)https://www.cnbc.com/2024/05/coreweave-debt-financing/ [8]Reuters, "CoreWeave raises $2.3 billion in debt collateralized by Nvidia chips" (August 2023)https://www.reuters.com/technology/coreweave-raises-23-billion-debt-collateralized-by-nvidia-chips-2023-08-03/ [12]Reuters, "Lambda secures $500 mln loan with Nvidia chips as collateral" (April 2024)https://www.reuters.com/technology/lambda-secures-500-mln-loan-with-nvidia-chips-collateral-2024-04-04/ [13]Crusoe, "$750 million credit facility from Brookfield" press release (June 2025)https://crusoe.ai/newsroom/crusoe-secures-usd750-million-credit-facility-from-brookfield-to-accelerate GPU-backed debt yields 8-12% for lenders, well above the ~5% available on investment-grade corporate bonds as of 2025. [14]Georgetown Financial Policy Institute, "The Explosive Growth of Private Credit" (2025)https://finpolicy.georgetown.edu/wp-content/uploads/2025/06/The-Explosive-Growth-of-Private-Credit.pdf
What hardware to buy
Most neoclouds buy NVIDIA GPU servers from OEMs (original equipment manufacturers) like Dell, HPE, Supermicro, and Lenovo. The standard unit is an 8-GPU server built on NVIDIA's HGX baseboard, a reference board that packages 8 GPUs with high-speed NVLink connections between them. [15]NVIDIA, "HGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/hgx/ In BOMs we've reviewed, a complete server runs $250,000-$400,000 for the current B200/B300 generation. [1]Based on our review of real BOMs and industry pricing (2025-2026)
| Component | Typical cost |
|---|---|
| HGX GPU board (8 GPUs) | $200K-$300K+ |
| CPUs, RAM, NVMe drives | $15K-$40K combined |
| Network adapters (8x) + DPUs | $12K-$18K combined |
| PSUs, chassis, cooling | $3K-$8K |
Illustrative B200/B300 component pricing as of early 2026, based on BOMs and quotes we've reviewed. [1]Based on our review of real BOMs and industry pricing (2025-2026)
In the deployment quotes we've reviewed, networking adds 10-25% on top of server costs. [1]Based on our review of real BOMs and industry pricing (2025-2026) A 576-GPU cluster in that range uses roughly 12 InfiniBand switches, 18 Ethernet switches, 40 racks, and 1,500 fiber runs. [2]Based on review of multiple 576-GPU B300 cluster deployment quotes (early 2026)
Lead times and allocation
GPU lead times peaked at 8-11 months in late 2023 and shortened to 3-4 months by early 2024. [16]Tom's Hardware / UBS, "Wait times for Nvidia's AI GPUs ease to three to four months" (February 2024)https://www.tomshardware.com/tech-industry/artificial-intelligence/wait-times-for-nvidias-ai-gpus-eases-to-three-to-four-months-suggesting-peak-in-near-term-growth-the-wait-list-for-an-h100-was-previously-eleven-months-ubs Availability as of early 2026 depends on the GPU model and your relationship with the OEM. Smaller buyers often get pushed toward a VAR (value-added reseller), a middleman with pre-negotiated OEM allocations. [3]Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026) At hyperscale volumes (tens of thousands of GPUs), buyers can bypass OEMs and work directly with an ODM (original design manufacturer) like Quanta. Meta used this approach for its 24,576-GPU cluster to avoid OEM markups. [17]Pytorch to Atoms, "Meta's 24k H100 Cluster Capex/TCO and BoM Analysis" (May 2024)https://pytorchtoatoms.substack.com/p/metas-24k-h100-cluster-capextco-and
Where to put them
Neoclouds lease from colocation providers like Equinix, QTS, and CyrusOne. Lease terms, pricing, and deposit requirements vary widely by provider, region, and deployment size. [18]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
The constraint is power density. GPU clusters draw 40-80 kW per rack, well above the 5-15 kW range typical for enterprise racks. [18]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 Liquid-cooled training racks go higher: NVIDIA's GB200 NVL72, a 72-GPU rack, draws 120-132 kW. [19]NVIDIA, "GB200 NVL72" (2024)https://www.nvidia.com/en-us/data-center/gb200-nvl72/ Many older data centers have 8-15 kW rack limits and air-only cooling. They cannot support GPU workloads regardless of their tier certification.
A neocloud that owns GPUs but cannot find a facility with enough power and cooling watches those GPUs depreciate at $50,000-$75,000 per server per year (assuming a $300,000 server and a 4-6 year depreciation schedule [20]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/) while generating zero revenue.
Not every neocloud leases from a traditional provider. Crusoe originally deployed GPU clusters at stranded natural gas sites, converting gas that would otherwise be flared into electricity. [21]Crusoe Energy, "Bringing Computing to Where the Energy Happens"https://crusoe.ai/energy/ It has since expanded to grid-connected facilities, including a large campus in Abilene, TX serving Oracle and OpenAI's Stargate project. [22]Data Center Dynamics, "Crusoe's Abilene data center campus officially live, serving Oracle and OpenAI's Stargate" (September 2025)https://www.datacenterdynamics.com/en/news/crusoes-abilene-data-center-officially-live-serving-oracle-and-openais-stargate/
What you need besides hardware
Software stack
Customers expect to spin up GPU capacity on demand, see how much they're using, and get billed automatically. CoreWeave built its platform on Kubernetes, a system that schedules and manages workloads across servers. [23]CoreWeave, "Kubernetes Management for GenAI" (accessed March 2026)https://www.coreweave.com/products/coreweave-kubernetes-service Lambda offers GPU capacity through a self-service dashboard. [24]Lambda, "Lambda Cloud" product page (accessed March 2026)https://lambdalabs.com/service/gpu-cloud AI training customers often expect Slurm, the job scheduler that most academic and research clusters use.
Team
The first hires are infrastructure engineers who can rack, cable, and bring up GPU servers.
Beyond hardware: a platform engineer to build and maintain the software platform, a sales lead who can close six- and seven-figure enterprise contracts, and someone managing finance and vendor relationships. CoreWeave had roughly 880 employees at the end of 2024, the year before it went public. [25]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
Compliance and insurance
Most enterprise customers expect SOC 2 Type II certification, an audit that verifies security controls over at least six months. Colocation providers handle physical security, but the neocloud is responsible for logical access controls, change management, and incident response.
GPU-backed lenders require insurance on the collateral. A $36 million hardware deployment needs property insurance covering fire, water, theft, and equipment breakdown. Business interruption insurance covers lost revenue during extended outages. Transit insurance covers hardware in shipment, a common point of loss. [3]Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026)
How to get customers
Two pricing models dominate. CoreWeave uses 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 they use the capacity or not. Its contracted backlog hit $66.8 billion at the end of 2025, roughly 13x annual revenue. [11]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 demand, raising and lowering rates as supply shifts. [26]American Compute internal pricing data: 2,326 pricing observations across 8 providers and 10 GPU models, July 2022 through March 2026
Published H100 SXM rates range from roughly $1.35/hr on brokered bulk platforms to $6.16/hr on CoreWeave, with actual contract rates lower than listed prices. [26]American Compute internal pricing data: 2,326 pricing observations across 8 providers and 10 GPU models, July 2022 through March 2026
The take-or-pay model is easier to finance. Lenders value signed customer commitments as collateral alongside the GPUs themselves. An on-demand neocloud with no contracted revenue has to convince lenders that cash flow will cover loan payments without any guaranteed income.
The biggest cloud providers are also neocloud customers. Microsoft, Amazon, and Google are each spending $75-100 billion a year building their own AI data centers, [27]Reuters, "Microsoft plans to invest $80 billion on AI-enabled data centers in fiscal 2025" (January 2025)https://www.reuters.com/technology/artificial-intelligence/microsoft-plans-spend-80-bln-ai-enabled-data-centers-fiscal-2025-cnbc-reports-2025-01-03/ [28]CNBC, "Amazon expects to spend $100 billion on capital expenditures in 2025" (February 2025)https://www.cnbc.com/2025/02/06/amazon-expects-to-spend-100-billion-on-capital-expenditures-in-2025.html [29]Reuters, "Alphabet plans massive capex hike" (February 2025)https://www.reuters.com/technology/google-parent-alphabet-misses-quarterly-revenue-estimates-2025-02-04/ but demand for GPU capacity is growing faster than they can build. They rent from neoclouds to fill the gap. CoreWeave's $66.8 billion backlog is largely from these buyers.
Customer concentration is the tradeoff. Microsoft accounted for 67% of CoreWeave's 2025 revenue. [11]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/ Servers are not always running at full capacity. Contract rates are lower than the published prices listed above. Both of these reduce how much revenue each server actually generates.
What the economics look like
CoreWeave's FY 2025 financials are the most detailed look at neocloud economics available. The company reported $5.1 billion in revenue with a 60% adjusted EBITDA margin, meaning 60 cents of every dollar earned remained after operating expenses but before depreciation and interest. [11]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 two largest costs below that line are depreciation and interest. Depreciation depends on how many years a neocloud spreads out the cost of its servers (between 4 and 6 years in practice), and the difference between those schedules adds up to tens of millions per year. Interest on CoreWeave's over $21 billion in debt cost $1.2 billion in 2025. After both, CoreWeave reported a net loss of $1.2 billion.
CoreWeave's existing operations generate more cash than they consume, but the company keeps investing billions in new hardware to fill its backlog. That reinvestment is what drives the reported loss.
What kills neoclouds
Customer concentration. Microsoft accounted for 67% of CoreWeave's 2025 revenue [11]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 is building its own AI accelerator, Maia. If its largest customer reduces external GPU demand, revenue shrinks while debt payments stay fixed.
Hyperscaler price pressure. AWS cut pricing on its H100 cloud instances by up to 45% in June 2025. [30]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-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/ Google and Amazon are deploying proprietary accelerators (TPUs, Trainium) that bypass NVIDIA entirely. Hyperscalers can sell GPU compute at a loss because it feeds their broader cloud platform, where they make money on storage, managed services, and data transfer. Neoclouds only sell GPU hours.
Hardware obsolescence. NVIDIA ships a new GPU architecture roughly every two years. Each generation increases power per GPU: a B200 draws up to 1,000W compared to 700W for an H100 SXM. A neocloud that refreshes its fleet needs both the capital for new servers and a facility that can support the higher power density. One that stays on older hardware risks losing customers to providers with newer GPUs.
Power scarcity. The bottleneck has shifted from GPU supply to electricity. A neocloud that buys GPUs before securing adequate colocation space has idle hardware depreciating with no revenue to offset it.
References
- Based on our review of real BOMs and industry pricing (2025-2026)
- Based on review of multiple 576-GPU B300 cluster deployment quotes (early 2026)
- Based on industry experience and conversations with GPU hardware buyers, sellers, and brokers (2025-2026)
- Reuters, "AI cloud startup Lambda raises $480 million in new round, Nvidia among investors" (February 2025)
- TechCrunch, "Lambda raises $1.5 billion following multibillion-dollar Microsoft deal" (November 2025)
- Nscale, "$1.4bn Delayed Draw Term Loan Backed by GPUs" press release (February 2026)
- CNBC, "Nvidia-backed GPU cloud provider CoreWeave is worth $19 billion" (May 2024)
- Reuters, "CoreWeave raises $2.3 billion in debt collateralized by Nvidia chips" (August 2023)
- CNBC, "CoreWeave raises $7.5 billion in debt financing" (May 2024)
- AltStreet, "GPU Depreciation & Obsolescence Risk" reference guide (January 2026)
- CoreWeave, Q4 2025 earnings and 10-K annual report (February 2026)
- Reuters, "Lambda secures $500 mln loan with Nvidia chips as collateral" (April 2024)
- Crusoe, "$750 million credit facility from Brookfield" press release (June 2025)
- Georgetown Financial Policy Institute, "The Explosive Growth of Private Credit" (2025)
- NVIDIA, "HGX Platform" (accessed March 2026)
- Tom's Hardware / UBS, "Wait times for Nvidia's AI GPUs ease to three to four months" (February 2024)
- Pytorch to Atoms, "Meta's 24k H100 Cluster Capex/TCO and BoM Analysis" (May 2024)
- Eaton, "2025 Data Center Progress Report" (2025)
- NVIDIA, "GB200 NVL72" (2024)
- SiliconAngle, "Resetting GPU Depreciation: Why AI Factories Bend, But Don’t Break, Useful Life Assumptions" (November 2025)
- Crusoe Energy, "Bringing Computing to Where the Energy Happens"
- Data Center Dynamics, "Crusoe's Abilene data center campus officially live, serving Oracle and OpenAI's Stargate" (September 2025)
- CoreWeave, "Kubernetes Management for GenAI" (accessed March 2026)
- Lambda, "Lambda Cloud" product page (accessed March 2026)
- CoreWeave, S-1 registration statement filed with the SEC (March 2025)
- American Compute internal pricing data: 2,326 pricing observations across 8 providers and 10 GPU models, July 2022 through March 2026
- Reuters, "Microsoft plans to invest $80 billion on AI-enabled data centers in fiscal 2025" (January 2025)
- CNBC, "Amazon expects to spend $100 billion on capital expenditures in 2025" (February 2025)
- Reuters, "Alphabet plans massive capex hike" (February 2025)
- AWS News Blog, "Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances" (June 2025)
Frequently Asked Questions
How much does it cost to start a neocloud?
Hardware is the biggest cost. An 8-GPU server (B200 or B300 generation) lands in the $250,000-$400,000 range as of early 2026. A 576-GPU cluster, 72 servers with InfiniBand networking and supporting infrastructure, is roughly $36 million. A smaller 16-GPU deployment (2 servers plus networking) starts around $1 million.
How do neoclouds raise capital?
Two sources fund neoclouds: equity (selling ownership to investors) and GPU-backed debt (borrowing from lenders with hardware as collateral). Most use both. The debt is asset-backed. The hardware itself serves as collateral: if the borrower defaults, the lender seizes the GPUs.
Why is colocation hard for GPU clusters?
The constraint is power density. GPU clusters draw 40-80 kW per rack, well above the 5-15 kW range typical for enterprise racks. Liquid-cooled training racks go higher: NVIDIA's GB200 NVL72, a 72-GPU rack, draws 120-132 kW. Many older data centers have 8-15 kW rack limits and air-only cooling. They cannot support GPU workloads regardless of their tier certification.
What kills neoclouds?
Microsoft accounted for 67% of CoreWeave's 2025 revenue and is building its own AI accelerator, Maia. If its largest customer reduces external GPU demand, revenue shrinks while debt payments stay fixed. AWS cut pricing on its H100 cloud instances by up to 45% in June 2025. The bottleneck has shifted from GPU supply to electricity.
Coverage creates a minimum value for what your GPUs are worth at a future date. If they sell below the floor, the policy pays you the difference.
Learn how it works →