How to Start a Neocloud: a Step-by-Step Guide for GPU Operators

How to Start a Neocloud

To start a neocloud, you pick a business model, secure powered space in a Tier III or Tier IV colocation facility and specify the hardware. You then sign a customer contract, finance the GPUs against that contract and build the operating layer that enterprise buyers audit before they commit. Each step produces the evidence the next one depends on, so the order matters as much as the steps.

Synergy Research forecasts that the neocloud market will approach $400 billion by 2031, from more than $25 billion of revenue in 2025 (April 2026). Q4 2025 alone brought in $9 billion of neocloud revenue, up 223% year over year.

As of September 2026, USD.AI had reviewed over 700 GPU financing deals representing about $25 billion of loan demand. Neoclouds make up the majority of qualified borrowers, so this guide follows the launch from the financing side.

It is written for founders, CFOs and infrastructure leads making a first deployment decision, and for operators adding a second site. At either stage, USD.AI's page on GPU financing for neoclouds covers how financing fits around your deployment.

Six Steps to Launch a Neocloud

The steps run roughly in this order because each one is evidence for the next. Lenders and customers review them together, so a signed contract needs powered space behind it and a purchase order needs a contract to repay it.

1. Choose the business model you will run

Operators sell compute from owned GPUs in three broad ways, and each one sets a different balance of value and exposure.

• Full-stack platform: You own the GPUs and build the software and orchestration layer customers use. This model earns the most value per MW and carries the most hardware-refresh and utilization exposure.
• Bare-metal GPU cloud: You own the hardware and rent servers or clusters with a lighter software layer, and customers run their own stack.
• Power and infrastructure host: You own the land, power and cooling, and customers bring their own hardware.

In our view, choose by what you can contract and finance as well as by margin, since the model sets the revenue a lender sees. Your revenue mix follows the same logic. On-demand or spot revenue doesn't disqualify a deal with USD.AI, but it affects pricing because it isn't contracted.

If you run the site but don't own the GPUs, the hardware owner is the borrower. That matters most for power and infrastructure hosts, whose customers own the servers and carry the financing.

2. Secure power and a colocation site

Power is the scarce input at large scale. Lawrence Berkeley National Laboratory estimates data centers could use 11.8% of US electricity by 2030, or 649 TWh (June 2026). Its full range spans 9.5–15.3%.

Texas shows how that demand reaches the grid. ERCOT's July 2026 update counted about 474 GW of large-load interconnection requests, roughly 90% of them data centers (June 2026). About 5.9 GW of large load was energized, and the gap between requests and energized load is the queue a new build joins.

New builds also wait on equipment. A National Laboratory of the Rockies report found large power transformers averaged about 30 months from order to receipt (May 2026). That figure covers the first half of 2025, up from 5–16 months in 2010, and it paces any site that needs a new substation.

We also treat switchgear as a long-lead item to plan around, and in our view a first deployment of 1–10 MW belongs in existing colocation. We see 1–5 MW of powered racks readily available (October 2026), so a middle-market operator can skip the multi-year build cycle.

Financed GPUs must sit in a Tier III or Tier IV facility. Tier III means concurrently maintainable, so components can be serviced without taking IT load offline. Tier IV adds fault tolerance, so a single equipment failure doesn't interrupt operations.

Before you sign, check power density per rack and the cooling type, air or direct liquid, against your hardware. Colocation providers that want to host financed deployments can join USD.AI's infrastructure partner program.

3. Specify the hardware and equipment schedule

Financing is sized on the full system, so name more than the accelerator. The equipment schedule, the itemized list of everything the loan buys, covers:

• The system model and builder, and GPUs per system
• CPUs, memory and storage
• Networking, including adapters, switches, cabling and an InfiniBand or RoCE fabric
• Power distribution, cooling and racks
• The delivery schedule and the site

Common configurations include 8-GPU HGX B200 and B300 servers, and DGX B200 and B300 systems are air-cooled at about 14 kW each (September 2026). Rack-scale NVL72 systems join 72 GPUs and 36 CPUs in one liquid-cooled rack, and the rack is the unit of purchase. On the AMD side, the MI350X is air-cooled and the MI355X uses direct liquid cooling. USD.AI's page on NVIDIA GB300 rack systems covers financing for the rack-scale option.

USD.AI's pricing research puts a complete 8-GPU NVIDIA B300 server at about $600,000 at Q4 2026 pricing, or about $75,000 per GPU. Networking, racks, installation and tax add to that total, so carry them as separate lines on the schedule. The NVIDIA B300 financing details page shows how that server is financed.

Storage and the network fabric deserve the same care as the GPUs, because they decide whether the GPUs sit idle. Storage IOPS, the input/output operations per second a storage system sustains, and fabric bandwidth set how fast data reaches each GPU.

4. Sign your first customer contract

The customer contract repays the loan, so it shapes every step after it. In the middle market, we most commonly see 3–4 year take-or-pay contracts from inference providers and non-US hyperscalers, matched to loan tenor (July 2026). Under take-or-pay, the customer pays for reserved capacity whether or not it uses it.

CoreWeave's 2025 annual report shows committed contracts made up over 98% of its FY2025 revenue. Its contracts carried a weighted-average duration of about five years and average prepayments of 15–25% of total contract value (December 2025). These are one company's FY2025 results, and they show what a contracted revenue base looks like at scale.

When you negotiate, aim for a term that matches the loan, fixed monthly take-or-pay payments and a creditworthy customer. The payments should cover debt service with room to spare, because that headroom is what absorbs a late invoice or a slow ramp.

Compute IQ, USD.AI's agentic underwriting system, reflects the same priority. The customer's credit and the terms of the offtake contract carry most of the review, because contracted revenue repays the loan. That is also why first deployments qualify with a strong customer and a well-drafted contract, before the operator has a long track record.

5. Finance the GPUs against the contract

GPU financing through USD.AI is asset-backed, secured by the hardware and underwritten on contracted cash flow. Your GPUs stay deployed and earning while they secure the loan. Each stage of the GPU financing process draws on the same application package:

• An approved hardware purchase order
• Proof of equity
• An executed colocation agreement
• An executed customer contract or evidence of on-demand revenue

Diligence adds an org chart showing owners of 10% or more, trailing 12 months of financial statements, and KYC/AML checks. Compute IQ then scores the customer, the contract and the operator on a 100-point scale, and that score sets financeability, pricing tier, size and structure. Our page on how Compute IQ scores deals explains a model built on more than $25 billion of GPU financing data points (October 2026).

Every financing originated through USD.AI closes on the same structure. A bankruptcy-remote special-purpose vehicle (SPV), an entity separate from your operating company, holds the GPUs, customer contracts and revenue accounts. Capital is held in escrow until the servers are independently verified as installed, and a three-month debt-service reserve is funded at closing. Every loan's hardware carries residual-value insurance, designed to reduce loss severity if the GPUs sell below their insured value.

Loans cover up to 80% of the deployment cost, with you contributing at least 20% as equity. Rates are fixed or floating, terms run 3–5 years with straight-line monthly amortization, and financing starts from about $5 million.

Illustrative example. Assume 20 eight-GPU B300 servers (160 GPUs) at about $600,000 each at Q4 2026 pricing, which comes to $12.0 million of server cost. That supports a loan of up to $9.6 million against at least $2.4 million of equity. On a three-year term, straight-line amortization repays about $267,000 of principal a month plus interest. The example covers server cost only, so networking, racks, installation and tax raise every figure.

6. Build the operating layer enterprise buyers audit

Enterprise buyers audit your controls before they sign, and the same controls support the loan file. SOC 2 is an independent audit of a provider's security controls. CoreWeave's 10-K states the company aligns with SOC 2 and ISO/IEC 27001 and requires its colocation providers to maintain them. Inside your own stack, that means tenant isolation, identity and access management (IAM) with single sign-on (SSO), and encryption key management.

Set service-level agreements (SLAs) with defined remedies and a support escalation path. Build observability across GPU utilization, job queues, storage throughput and network health. Telemetry monitoring of uptime and server health is part of every financing through USD.AI, so that data also feeds the loan's monitoring.

Uptime Institute's 2026 outage analysis found power remains the leading cause of impactful outages (May 2026). Uninterruptible power supply (UPS), transfer switch and generator failures lead that category. Per-site outage rates have declined five years running, so test that equipment before customers depend on it.

Where New Neoclouds Get Stuck

A single large customer can carry a launch, and that dependence doesn't fade with scale. Microsoft was about 67% of CoreWeave's FY2025 revenue, and CoreWeave's largest customer was 62% in 2024; those are one company's results. The fix is to add customers over time and match each contract's term to the loan it supports.

Ordering hardware before power, colocation and the contract are lined up strands capital in equipment with nowhere to run. Follow the six steps in order, since capital releases from escrow only after the servers are verified as installed.

Spot-only revenue leaves a deployment with income that is hard to forecast and hard to lend against. Contract part of the capacity and sell the rest on demand, since spot revenue is still financeable through USD.AI at a different pricing tier.

An underspecified equipment schedule names the GPUs and leaves networking, storage, power distribution and racks out of the deal. List the full system from the start so the financing covers everything the cluster needs to run.

Power and equipment delays hit new builds hardest, because transformer and switchgear lead times sit outside your control. Deploying a first cluster into existing colocation removes most of that exposure from your launch plan.

FAQ

How much does it cost to start a neocloud?

Hardware drives the cost: a complete 8-GPU NVIDIA B300 server costs about $600,000 at Q4 2026 pricing, before networking, racks and installation. Financing through USD.AI covers up to 80% of the deployment cost, with you contributing at least 20% as equity.

Can a first-time operator get GPU financing?

Yes. First deployments qualify with a strong customer and a well-drafted contract, and a track record counts toward the score alongside them.

How much power does a first neocloud deployment need?

Middle-market clusters run 1–10 MW, and in our view a first deployment at that size belongs in existing colocation. Check rack power density and cooling type against the hardware before you sign.

Do I need a customer before I buy GPUs?

For financing, yes. An executed customer contract or evidence of on-demand revenue is part of the application, alongside the purchase order, colocation agreement and proof of equity.

Launching with Financing in Place

A neocloud launch runs in sequence: business model, powered colocation, equipment schedule, customer contract, financing and the operating layer buyers audit. Once the customer contract is in hand, financing is the natural next step, and the documents you built along the way form the application. When your site and contract are lined up, Talk to USD.AI about financing your first deployment.