GPU Cluster Cost: Every Line Item in Owning a B300 Deployment

Key Takeaways

• Servers are the largest line, but a complete GPU cluster budget has about ten lines, and several have no public benchmark.
• One vendor's model puts the back-end network at 7.1%–8.5% of GPU hardware cost for a B300 cluster.
• Ten B300 servers draw about 140 kW, roughly $142,000–$166,000 a year in electricity at the August 2026 US industrial average and 1.2–1.4 design PUE (illustrative).
• Size any loan on all the eligible hardware on the purchase order; installation, taxes, software and other soft costs sit outside it.

Introduction

The first question most operators ask about GPU cluster cost is the server price. The budget that gets approved and financed is the whole deployment. A GPU cluster is a set of GPU servers networked to run as one system. The servers are only the first of many lines it puts on the books.

More companies are choosing to own that system. Comma.ai's CTO estimated the company spent about $5 million on its own data center, most of it on 600 GPUs. He put the cost of the same work in the cloud at $25 million or more, one example of why companies now own compute.

This guide is for operators who plan to own. If you're weighing rental instead, H100 rental prices are covered in a separate guide.

You'll find the full cost stack below, with each line marked by whether a public benchmark exists or a quote is required. A worked example then budgets 10 B300 servers, and the last section shows how financing fits the total.

What a GPU Cluster Costs: the Line-item Stack

A GPU cluster costs about $600,000 per 8-GPU NVIDIA B300 server at Q4 2026 pricing, based on USD.AI's own pricing research. That figure covers the server alone, before tax, so 10 servers come to about $6.0 million. A complete budget then adds networking, storage, racks, power and cooling, installation, colocation, electricity, software, taxes, insurance and the cost of financing.

The budget has about ten lines, and only some of them have a public benchmark. No acceptable public all-in figure exists per GPU or per MW, so the total has to be built from line items and quotes. The table below sets out each line, what it covers and where a benchmark comes from.

The hardware lines follow an equipment schedule, the itemized hardware list behind a purchase order. A financing review asks for one in six categories, and each maps to a line above:

• System model and builder: the server model and the company that builds it, which sets the server line.
• GPUs per system and number of systems: the GPU count, which drives networking, power and software.
• CPUs, memory and storage: the server configuration and any shared storage.
• Networking adapters, switches and cabling: the full network bill of materials.
• Power distribution and cooling: PDUs, cooling hardware and the cooling type the servers require.
• Racks, delivery schedule and site: rack count, when each shipment lands and which facility hosts it.

Servers: the Largest Line in a GPU Cluster Budget

A complete 8-GPU NVIDIA B300 server costs about $600,000 at Q4 2026 pricing, or about $75,000 per GPU, based on USD.AI's own pricing research. The figure is server-only and pre-tax. Networking, racks, installation and tax add to it, which is why the server price alone understates the cost of a GPU cluster.

The 8-GPU server is the standard building block for a cluster of this size. It carries eight Blackwell Ultra SXM GPUs on one HGX baseboard, NVIDIA's board that mounts the GPUs and links them with NVLink. System builders put that baseboard into x86 servers with their own CPUs, memory, drives and power supplies, so configurations and prices vary by builder.

Rack-scale systems such as the NVL72 are bought differently. They are sold as a whole rack, with the GPUs, CPUs, switches and liquid cooling designed together. That changes how the budget is built, because several lines that are separate for 8-GPU servers arrive in one rack price.

The NVIDIA B300 hardware reference covers the B300's specifications in detail. For the components inside each system and how they fit together, see how a GPU server is built.

Server quotes vary with configuration, so compare them on the same CPU, memory and drive specification. Confirm too whether each quote includes tax, since our $600,000 figure is pre-tax.

Networking: a Share of GPU Hardware, Priced by Design

Networking is usually modeled as a share of GPU hardware cost, and one vendor's model puts it at 7.1%–8.5% for B300. In a 2026 model, Hedgehog, a networking vendor, put the back-end network for a 1,024-GPU cluster at $3.26M–$3.90M, depending on cabling. That range is for its do-it-yourself design with NVIDIA switches, and it equals 7.1%–8.5% of GPU hardware cost for B300 and 8.0%–9.5% for B200.

The back-end network is the high-speed fabric that links GPUs across servers so they can train or serve as one system. It is built from switches, network interface cards (NICs) in each server and the optics and cabling between them. Hedgehog's total also includes a $1.65M gateway and firewall, about half of the low end of that range.

Read the figure as one vendor's model of one design. It measures the back-end network against GPU hardware cost alone, so it says nothing about the total cluster budget. It also comes from a 1,024-GPU design. Applying its percentage to a single server price would carry that design's assumptions into a much smaller deployment.

Your own network cost follows your design. The choice between InfiniBand and Ethernet, the switch tiers, the cable lengths and the use of copper or optics each move the number. Ask your system builder or integrator for a full network bill of materials priced against your rack layout.

Switches, NICs and cabling are hardware on the purchase order. Asset-backed lenders generally treat them as eligible hardware alongside the servers, so a complete network quote also sizes part of the loan.

Power, PUE and Colocation: the Recurring Costs

Once a cluster is installed, its recurring costs are electricity and colocation. Ten B300 servers draw about 140 kW of IT load before facility overhead. These two lines run for as long as the servers are in service. They set the monthly cash cost of keeping the cluster online.

Colocation contracts handle power in one of two ways. Some bundle electricity into the per-kW rent, and others meter it separately at a pass-through rate. Check which applies before you add the lines below, so you don't count power twice.

Power draw and PUE

NVIDIA rates the DGX B300 at about 14 kW of power consumption. The DGX B300 user guide lists 14.5 kW. NVIDIA puts the DGX B200 at about 14.3 kW maximum. An 8-GPU B300 server from another builder will differ, but NVIDIA's DGX figures are a reasonable proxy for budgeting.

The facility uses more power than the servers alone. Power usage effectiveness (PUE) is total facility power divided by IT power. A PUE of 1.3 means 0.3 kW of cooling and overhead for every kW of servers. Crusoe built its Abilene Stargate campus to an annualized design PUE of 1.2–1.4, and its 2026 projects target 1.1–1.25.

Those are one operator's design targets, and measured averages at your facility may differ. Ask your colocation provider for the design PUE and any measured figures. The worked example below uses Crusoe's 1.2–1.4 range as a benchmark.

Electricity prices

The US industrial average electricity price was 9.64 cents per kWh in August 2026, according to preliminary EIA data. State rates vary widely, and the table shows four common data center states.

These are averages across industrial customers. A colocation contract may pass power through at a different rate, with its own markup or tariff structure. Use the contract rate once you have it, and use the EIA figure as a starting point.

Colocation rents

CBRE expects average rental rates for 250–500 kW requirements to be "above $215 per kW per month" in 2026. That is the closest public benchmark for a deployment the size of a small cluster.

For large requirements, CBRE's H1 2026 Northern Virginia report says asking rates for 10-plus-MW requirements "rose to $160 to $185 per kW." Colocation vacancy there "fell to 0.24% in H1," which leaves tenants little room to negotiate. Treat these as large-requirement benchmarks only.

A 10-server deployment at about 140 kW of IT load sits below both bands. You'll need quotes from colocation providers, and the quote should state whether power is bundled or metered.

Some operators build their own space instead of leasing it. CBRE puts demanding data center builds at $14–16 million per MW, a different scale of project from leasing a few racks. Data center financing sits outside a GPU loan and needs its own plan.

Software, Taxes and Insurance

NVIDIA AI Enterprise costs $4,500 per GPU for a one-year subscription, if you need it. Taxes depend on whether your colocation facility qualifies for a state exemption, and insurance has no public rate.

NVIDIA AI Enterprise is priced at $4,500 per GPU per year as of September 2026, including Enterprise Business Standard Support. It is optional. Many operators run their own software stack, so budget it only if you or your customers need it.

Sales and use tax on servers varies by state and by whether an exemption applies. The Texas Comptroller describes a temporary 6.25% state sales and use tax exemption for qualifying data centers, though local taxes are still due. Qualifying requires a $200 million investment over five years and 20 qualifying jobs.

Virginia's exemption runs through June 30, 2035, according to the Virginia Economic Development Partnership. It requires $150 million of new investment, 50 jobs and wages of 150% of the prevailing average wage. Colocation tenants join through the facility's memorandum of understanding (MOU) with the state and a Participation Certificate and Agreement.

A 10-server deployment at about $6 million falls far below either state's investment threshold on its own. A small deployment's exemption therefore depends on whether its colocation facility qualifies and admits tenants. Ask the facility early, and talk to your tax adviser before you count any exemption in the budget.

Insurance on GPU hardware has no public benchmark rate, so it needs a quote from an insurer or broker. Every loan financed through USD.AI carries property and casualty insurance once escrow is released. Budget the premium as a recurring line from the day the hardware is installed.

The Line Items You'll Need Quotes For

Four lines in a GPU cluster budget have no acceptable public figure, and neither does an all-in cost per GPU or per MW. Each line needs a quote from the party that will supply it. Blog estimates and broker percentages for these lines rest on assumptions you can't check. We leave them out and rely on the quotes a financing review asks to see.

• Storage: quoted by the OEM, system builder or a storage vendor, sized to your datasets and checkpoint needs.
• Racks, PDUs, liquid cooling and coolant distribution units (CDUs): quoted by the system builder, integrator or colocation provider, depending on who supplies each part.
• Installation, burn-in and acceptance testing: quoted by the integrator or system builder that racks and commissions the servers.
• Spares and return merchandise authorization (RMA) allowance: quoted by the OEM or system builder, covering spare parts and warranty returns.
• All-in cost per GPU or per MW: no single party quotes it, so build it from the other quotes.

A CDU moves liquid coolant between the facility's water loop and the racks, and you need one only if the servers are liquid-cooled. An RMA is the manufacturer's process for taking back failed parts, so the allowance covers spares while repairs are out.

The cooling type follows the server design. The DGX B300 is air-cooled, while NVL72 rack-scale systems are liquid-cooled and need CDUs and facility water. Confirm the cooling type before you request colocation quotes, since it decides which facilities can host the cluster.

Worked Example: Budgeting 10 B300 Servers (80 GPUs)

This illustrative budget puts 10 B300 servers at about $6.0 million, with electricity at about $142,000–$166,000 a year if billed at the US industrial average. Every other line needs a quote or depends on the deployment's details. The example uses the DGX B300 as a proxy for power draw.

The power line shows its math in four steps:

1. IT load: 10 servers × about 14 kW = about 140 kW.
2. Facility load: 140 kW × 1.2–1.4 PUE = 168–196 kW.
3. Annual energy: 168–196 kW × 8,760 hours = 1,471,680–1,716,960 kWh.
4. Annual cost: 1,471,680–1,716,960 kWh × $0.0964 ≈ $141,900–$165,500, or about $142,000–$166,000 a year.

Using the user guide's 14.5 kW per server would raise each figure slightly. In Texas at 7.21 cents per kWh, the same load would cost less, and in Georgia at 13.05 cents it would cost more. If your colocation rent bundles power, this line folds into the colocation quote.

The software line applies only if you subscribe: 80 GPUs × $4,500 = $360,000 a year. The example leaves networking at quote required on purpose. Hedgehog's percentage describes a 1,024-GPU design, and a 10-server network should be priced from its own bill of materials.

How Financing Fits the GPU Cluster Budget

Asset-backed GPU loans are sized on the eligible hardware on the purchase order. For 10 B300 servers, that starts with the $6.0 million server line and grows with networking and racks. Eligible hardware generally means the servers plus networking, owned racks, PDUs and cabling. Soft costs sit outside the loan and are funded by the operator: installation, labor, shipping, taxes, software subscriptions and data center build-out.

Financing through USD.AI covers up to 80% of the deployment cost, known as loan-to-value, with the operator contributing at least 20% as equity. The loan runs 3–5 years at fixed or floating rates, with straight-line amortization, meaning equal monthly principal payments. The operator funds a three-month debt-service reserve at closing, cash set aside to cover loan payments. Financing starts at about $5 million, with larger facilities available.

The loan term should sit inside the hardware's useful life, so the debt is repaid while the servers still earn revenue. Check that the term you choose leaves room between the final payment and the end of that life.

In an illustrative case on the servers alone, a $6.0 million server line supports a loan of up to $4.8 million, with at least $1.2 million of equity. Principal comes to about $133,000 a month over three years, or $80,000 a month over five.

Networking and other eligible hardware add to both the loan and the equity. That is why the loan is sized on deployment cost, and why a complete network and rack quote matters before you apply. The soft costs listed above come on top, funded from the operator's own capital.

Loan capital is held in escrow and released once its conditions are met: the servers are installed, burned in and independently verified online. Plan your cash for vendor deposits and soft costs with that sequence in mind. For the full process, see financing a GPU cluster step by step.

Conclusion

The cost of a GPU cluster starts with the servers, at about $600,000 per B300 server at Q4 2026 pricing. Networking adds a share of GPU hardware cost, and power and colocation add recurring monthly costs. Software, taxes and insurance follow, and storage, racks, cooling and installation need quotes.

Build the budget line by line, quote every unpriced line and size the loan on the full deployment. As of September 2026, USD.AI had reviewed over 700 GPU financing deals. Each financing starts with the same pieces in place: customer revenue, a hardware purchase order, colocation and the equity contribution.

GPU financing through USD.AI covers new deployments and GPUs already in service. Every deal is scored against published underwriting standards on a 100-point scale, as described on the Compute IQ scoring page. When your purchase order and quotes are ready, bring them to the first conversation.

Talk to USD.AI

FAQ

How much does a GPU cluster cost?

A GPU cluster costs about $600,000 per 8-GPU NVIDIA B300 server at Q4 2026 pricing, based on USD.AI's own pricing research. Ten servers come to about $6.0 million before networking, other hardware, colocation, power, software and tax.

What is a GPU cluster?

A GPU cluster is a set of GPU servers networked to run as one system. Operators use clusters for large training jobs and for inference at scale.

How much power does a B300 server use?

NVIDIA rates the DGX B300 at about 14 kW, and its user guide lists 14.5 kW. Ten servers draw about 140 kW of IT load before PUE adds facility overhead.

What does colocation cost per kW?

CBRE expects average rents for 250–500 kW requirements to be above $215 per kW per month in 2026. In Northern Virginia, asking rates for 10-plus-MW requirements rose to $160 to $185 per kW in H1 2026. Smaller deployments need quotes.

Is it cheaper to own or rent a GPU cluster?

It depends on how fully you use the GPUs and on the terms of your offtake contract or cloud agreement. Comma.ai's CTO estimated about $5 million for its own data center, mostly 600 GPUs, against $25 million or more in the cloud.

How much of a GPU cluster can be financed?

Financing through USD.AI covers up to 80% of the deployment cost, with the operator contributing at least 20% as equity. The loan is sized on the eligible hardware on the purchase order, so soft costs such as installation and taxes sit outside it.