Aircloud Compute is now Lightyard. Aircloud continues as a private inference platformAircloud continues for inference

Economics

How much does a GPU cluster cost?

The total cost of a GPU cluster line by line, from servers and fabric to the team and cost of capital, with a five-year worked example for 8, 32 and 64 HGX B300 nodes.

All notesPublished Updated 5 min read
In this note
  1. Who owns it and who runs it
  2. What drives the total cost of a GPU cluster
  3. Worked example: 8, 32 and 64 nodes
  4. What the example leaves out
  5. When owning gets expensive

A GPU cluster built on HGX B300 needs about US$5.5M of capex at 8 nodes (64 GPUs), US$21M at 32 nodes and US$42M at 64 nodes, at mid-2026 street prices. Owning it for five years costs about three times what the servers cost, once colocation, power, support, staff and the cost of capital are counted.

The figures below use the assumptions from our HGX B300 cluster cost model, scaled to each size. Swap in your own quotes where you have them.

Who owns it and who runs it

There are a few common ways to get a dedicated cluster:

  • Buy and run it yourself. You pay for the servers, fabric and storage up front, lease colocation space, hire a team and carry the hardware on your balance sheet. The cost per hour depends on how fast you deploy and how busy you keep it.
  • Buy and have it built and run. You still own the hardware, but a contractor designs, sources and builds the cluster and operates it. The team line becomes a contract.
  • Lease the hardware. A lessor buys the servers and you pay monthly. You still need the space, the power and the team. Our GPU lease checklist covers the terms.
  • Reserve capacity. A provider owns and runs the cluster and you commit to it for a fixed term. You skip the capex, the site and the on-call team. See reserved clusters.

The rest of this note is about owning, because it's the option that needs a spreadsheet. Our buy or rent note compares the options.

What drives the total cost of a GPU cluster

Nodes

The servers are the largest single line and usually about a third of the five-year total. An 8-GPU HGX B300 server is reported at about US$430K to US$550K in 2026 (GPUsmith (opens in a new tab)). Quotes start around US$500K, so we use the upper part of that range and take US$525K. Rack-scale systems are priced per rack: a GB300 NVL72 sold for just under US$5.0M in August 2026 purchase orders (Data Gravity (opens in a new tab)), covered in GB300 NVL72 cost.

Fabric

Each GPU gets an 800 Gb/s link on InfiniBand or Ethernet, plus storage, in-band and out-of-band networks. Switches and optics come to about US$40K to US$70K per node (our estimate, based on NVIDIA's reference architecture (opens in a new tab)). We use about US$59K per node.

Storage

A parallel all-flash file system sized at 2 TB or more per GPU, following SemiAnalysis's (opens in a new tab) cluster scenarios. Storage grows more slowly than the GPU count, so we use US$0.4M, US$1.0M and US$1.8M for 8, 32 and 64 nodes.

Power and facility

An air-cooled HGX B300 node draws about 14 kW per system, and with network and storage a cluster needs about 15.6 kW provisioned per node. Wholesale colocation averaged about US$196 per kW-month in North American primary markets at the end of 2025 (CBRE (opens in a new tab)), and energy is metered on top at around 9 cents per kWh (EIA (opens in a new tab)). Over five years, colocation costs more than energy. Fit-out, freight and commissioning add a few hundred thousand dollars per site. Our note on building a data center for GPU clusters covers the site side.

Operations

Someone has to be on call around the clock, which takes about five people at one site no matter how many nodes there are. Add network, storage and vendor work and we get 5, 6.5 and 7.5 full-time staff for 8, 32 and 64 nodes, at about US$250K a year each loaded. Support contracts add 5% of hardware a year while under warranty and 10% after, and cluster software about US$6K per node a year.

Hardware also fails. Meta (opens in a new tab) saw about 0.135 hardware interruptions per GPU per year on H100s, which costs 1% to 3% of GPU-hours in swaps and restarts.

Depreciation and cost of capital

The capex is spent before the first useful hour, and the cluster is worth a fraction of it in five years. We credit 15% of the server cost as resale value at year five (an OEM TCO study (opens in a new tab) assumes zero) and charge 8% a year on the capex. Accounting schedules are in GPU depreciation and useful life.

Time before useful work

Between the purchase order and the first production job there are usually 6 to 18 months of leasing, fit-out, delivery and burn-in (deployment timeline). The middle case takes 9 months out of a five-year life. With 2% lost to failures, that leaves 83% of the hours as productive.

Worked example: 8, 32 and 64 nodes

HGX B300, air-cooled, in a third-party colocation hall, middle-case assumptions. Figures in US$ millions over five years.

Line8 nodes (64 GPUs)32 nodes (256 GPUs)64 nodes (512 GPUs)
Servers at US$525K4.2016.8033.60
Fabric and networks0.471.903.80
Storage0.401.001.80
Fit-out, freight, commissioning0.200.450.80
Spares (5% of servers)0.210.841.68
Capex5.4820.9941.68
Colocation (about 125, 500 and 1,000 kW)1.465.8311.66
Energy0.491.953.90
Support, years 1 to 51.776.9013.72
Team (5, 6.5, 7.5 FTE)5.787.528.68
Cluster software0.240.961.92
Insurance0.100.390.78
Residual value at year 5-0.63-2.52-5.04
Cost of capital at 8%1.485.6711.26
Five-year all-in cost16.247.788.6
Productive GPU-hours2.33M9.34M18.68M

The 32-node column matches the middle case in our B300 cost note, which also has low and high cases. The 8- and 64-node columns scale the same assumptions; storage, fit-out and team are our estimates for those sizes.

At 8 nodes the team is the biggest line, larger than the servers. At 64 nodes it drops to about a tenth of the total, and the servers, support and cost of capital dominate. Doubling from 32 to 64 nodes raises the total by about 85%, not 100%, because the team and storage grow more slowly than the GPU count.

These figures also assume every productive hour is used. Divide by the share you expect to use (0.8 or 0.6 is realistic for most owners) to see what an idle hour costs.

What the example leaves out

Building or converting a hall, which is a separate project measured in megawatts. Liquid cooling, which rack-scale systems such as GB300 NVL72 require. Optional software licences such as NVIDIA AI Enterprise. Tax effects of depreciation. Each of these adds to the total.

When owning gets expensive

If the cluster is under about 32 nodes, the fixed team cost makes owning expensive per hour. If you don't have a hall with spare power, the 6 to 18 months of build time cost money and model progress. If your capital costs more than 8%, the cost-of-capital line grows quickly.

Those are the lines we take on. If you own the hardware, we can build the cluster and run it, so the team and the site are a contract instead of a hiring plan. If you'd rather not own it, we reserve dedicated clusters from 8 HGX nodes (64 GPUs) or one NVL72 rack (72 GPUs), single-tenant, on a 60-month term. Most clusters are 32 to 128 nodes; larger ones are scoped case by case. Talk to us. For the hardware-specific numbers, see B300 cost and GB300 NVL72 cost.

Questions this note answers

Sources

  1. GPUsmith: NVIDIA HGX B300 system page (mid-2026 pricing, power bins) (opens in a new tab)
  2. Data Gravity: How much does an NVIDIA NVL72 cost? (August 2026) (opens in a new tab)
  3. NVIDIA DGX SuperPOD reference architecture: DGX B300, architecture (opens in a new tab)
  4. SemiAnalysis: How much do GPU clusters really cost? (2026) (opens in a new tab)
  5. CBRE: North America Data Center Trends H2 2025 (February 2026) (opens in a new tab)
  6. US EIA: Electric Power Monthly, average price by end-use sector (opens in a new tab)
  7. OEM study: on-premise vs cloud generative AI TCO (July 2026) (opens in a new tab)
  8. Meta: The Llama 3 Herd of Models (reliability section) (opens in a new tab)
Get started

Tell us where you're starting.

Capital, hardware, a site or a team: tell us what you have and when you need it running. We usually reply within a few hours and come back with a written plan.

sales@lightyard.ai