In this note
An 8-GPU HGX B300 server sells for about US$500,000 to US$550,000 in 2026, or US$63,000 to US$69,000 per GPU. Add the fabric, storage, facility work and spares, and a 32-node cluster (256 GPUs) needs US$19M to US$24M of capex.
Owning it for five years costs about US$48M in our middle case, with US$35M in the low case and US$63M in the high one. The total counts the team, support, failures, software, insurance, residual value, cost of capital and the months before the cluster does any useful work. The servers are about a third of it.
The server prices are public reporting. The rest is our estimate, and every line is in a table so you can swap in your own quotes.
How much does a B300 server cost?
NVIDIA doesn't publish a price for HGX B300 or DGX B300. These are the figures in circulation:
| Source | System | Reported price | Per GPU |
|---|---|---|---|
| GPUsmith (opens in a new tab), mid-2026 | HGX B300 8-GPU server from OEMs | US$430K to US$550K+ | US$54K to US$69K |
| OEM TCO study (opens in a new tab), prices as of June 2026 | 8 × B300 server | US$785,607 (list) | US$98K |
The street figures are within about 10% of each other. The OEM figure is a list price before discounts. Quotes we see start around US$500K, so we use the upper part of the reported range: US$500K to US$550K, with US$525K as the middle case.
Quotes move with the CPU and DRAM, local NVMe, the NICs (eight ConnectX-8 at 800 Gb/s is normal), the chassis and the support term. NVIDIA (opens in a new tab) lists a DGX B300 at about 14 kW, with 2.1 TB of GPU memory and two BlueField-3 DPUs, and includes three years of business-standard support. Our HGX B300 page covers what the platform asks of a site.
B300 cluster price: what 32 nodes need besides the servers
Compute fabric
256 GPUs means 256 links at 800 Gb/s, on InfiniBand or Ethernet. NVIDIA's DGX B300 reference architecture (opens in a new tab) uses 8 leaf and 4 spine switches for a 72-node scalable unit. Scaled down to 32 nodes, a non-blocking two-tier fabric is about four leaves and two spines (our estimate). The optics often cost as much as the switches.
SemiAnalysis priced an H100 node (opens in a new tab) at US$270K, or about US$318K with its share of NVIDIA's reference networking, before storage. That's about US$48K of network per node. At 800 Gb/s we assume US$40K to US$70K per node for the compute fabric. With the storage, in-band and out-of-band Ethernet networks on top, the line is US$1.4M, US$1.9M and US$2.5M in the low, mid and high cases.
Storage
NVIDIA's guideline for a 72-node unit is 80 GB/s of aggregate read and 40 GB/s of write at the standard level (storage architecture (opens in a new tab)). SemiAnalysis's (opens in a new tab) cluster scenarios assume 2 TB per GPU for large-model pre-training and about 12 TB for multimodal research. For 256 GPUs that's 0.5 to 3 PB of all-flash parallel file system, and most of what you pay for it is software licensing. We use US$0.5M, US$1.0M and US$2.0M.
Fit-out, spares, space and power
Fit-out, freight, integration and commissioning come to US$0.25M, US$0.45M and US$0.65M (estimate). Spares are 3%, 5% and 7% of the server cost.
32 nodes plus network and storage need close to 500 kW provisioned. Wholesale colocation for 250 to 500 kW in North American primary markets averaged about US$196 per kW-month at the end of 2025, with vacancy at a record 1.4% (CBRE (opens in a new tab)). Energy is usually metered on top. The US industrial average was 9.64 cents per kWh in August 2026 (EIA (opens in a new tab)).
Air or liquid cooling for HGX B300
GB300 NVL72 only comes liquid-cooled; HGX B300 gives you the choice. OEMs (opens in a new tab) sell 8U air-cooled HGX B300 systems and liquid-cooled 4U and 2-OU systems. GPUsmith (opens in a new tab) lists the GPU at 1,100 W air-cooled and up to 1,400 W liquid-cooled.
| 32 nodes | Air-cooled (8U or DGX 10U) | Liquid-cooled 4U |
|---|---|---|
| Nodes per rack | 4 | 8 |
| Compute racks | 8 | 4 |
| Power per rack | ~56 kW (NVIDIA RA, 4 × DGX B300) | ~110 kW (estimate, 8 × ~14 kW) |
| GPU power bin | 1,100 W | Up to 1,400 W |
| Facility water | No | Yes, or a liquid-to-air sidecar |
| PUE assumed | 1.3 to 1.5 | 1.15 to 1.3 |
| Fit-out estimate | US$150K to US$400K | US$300K to US$700K |
At about 415 kW of IT load, a PUE of 1.22 instead of 1.4 saves roughly 650 MWh a year. At 9 cents per kWh that's about US$60K a year, or US$0.25M over the 51 months the middle case runs, which is about what liquid adds to the fit-out. On cost it's close to even. The decision comes down to whether the hall has facility water, how much floor space you have and whether you want the 1,400 W bin. The model below assumes air.
Costs most build cases leave out
Team
A production cluster needs someone on call around the clock. A week has 168 hours and one person works about 40, so a round-the-clock rotation with room for leave and training takes about five people at one site, whatever the cluster size. Add part-time shares of network engineering, storage, security, and vendor and facilities management, and we get 5.5, 6.5 and 7.5 full-time equivalents for 32 nodes, at US$200K, US$250K and US$300K a year loaded.
The BLS median for network architects (opens in a new tab) was US$134,050 in May 2025, and wages were 70.0% of private-industry employer compensation costs in June 2026 (BLS (opens in a new tab)). That puts the median near US$190K loaded, and people who run GPU fabrics usually cost more. Each specialist takes three to six months to hire (our estimate), so the team is on payroll before the nodes arrive. The model pays it from halfway through deployment.
The team doesn't double when the cluster does. We think the same core team could run 64 nodes with one more person.
Support and maintenance
DGX B300 includes three years of support, and many OEM quotes bundle a warranty. We still carry 3%, 5% and 8% of hardware a year for years one to three, for support you buy or an SLA upgrade on top. Drop the line if your quote already covers it. Post-warranty support in years four and five is 8%, 10% and 12% a year; the OEM TCO study (opens in a new tab) uses 12% of system cost a year for maintenance.
Hardware here means servers, fabric and storage: US$17.9M, US$19.7M and US$22.1M in the three cases. Insurance uses the same base. Spares and residual value are on the servers only.
Failures and RMA turnaround
During Llama 3 405B pre-training, Meta (opens in a new tab) recorded 419 unexpected interruptions in 54 days on 16,384 H100s. About 78% were hardware, and faulty GPUs and HBM3 memory alone accounted for 220. That's about 0.135 hardware interruptions per GPU per year, or around 35 a year on 256 GPUs at H100 rates.
Each one takes an 8-GPU node out while the part comes from spares or goes back under RMA, and the job restarts from its last checkpoint. With two to seven days per swap and one to three hours per restart, that's about 1% to 3% of GPU-hours, the range the model uses. Nobody has published comparable failure data for B300 yet.
Time before useful work
Months pass between the purchase order and the first production job: colo lease, fit-out, delivery, install, burn-in and acceptance. In leased colo space that's usually 6 to 18 months (AI cluster deployment timeline). We take 6, 9 and 12 months out of a fixed five-year life, because the next GPU generation arrives on schedule either way. Capital is committed at the order. Colocation and the team are paid from halfway through deployment, energy only once the cluster runs.
Software
Cluster management and monitoring (scheduler, observability, security tooling and their support) is US$3K, US$6K and US$10K per node a year (estimate).
NVIDIA AI Enterprise (opens in a new tab) is optional, so it's outside the headline. It lists at US$4,500 per GPU per year, or US$18,000 per GPU for five years. Hopper DGX systems bundled it, but NVIDIA's licensing guide (opens in a new tab) says Blackwell DGX systems need licences bought separately. Five years at list for 256 GPUs is about US$4.6M.
Insurance, residual value and cost of capital
Insurance is 0.3%, 0.4% and 0.5% of hardware a year (estimate).
Servers still sell at year five. NVIDIA launched DGX A100 at US$199,000 (opens in a new tab) in 2020. In October 2026 Tom's Hardware (opens in a new tab) reported A100 servers in China selling for up to US$82,000, about 40% of that launch price, though export controls keep demand for older GPUs unusually high there. The OEM TCO study assumes zero. We credit 25%, 15% and 10% of server cost. Decommissioning (de-install, drive sanitization, freight) comes out of that credit; we don't model it separately.
The capex is paid before the first useful hour, so the model treats it as a five-year loan at 6%, 8% and 10%, with the residual value as the final balloon payment. The cost of capital line is the interest on that loan.
B300 cluster cost over five years
Air-cooled, in a third-party colocation hall. The low case pairs the lowest costs with the highest residual value. All figures in US$ millions.
| Line, five years | Low | Mid | High |
|---|---|---|---|
| Servers (32 × US$500K, 525K, 550K) | 16.00 | 16.80 | 17.60 |
| Fabric and Ethernet networks | 1.40 | 1.90 | 2.50 |
| Storage | 0.50 | 1.00 | 2.00 |
| Fit-out, freight, integration, commissioning | 0.25 | 0.45 | 0.65 |
| Spares (3%, 5%, 7% of servers) | 0.48 | 0.84 | 1.23 |
| Capex | 18.63 | 20.99 | 23.98 |
| Colocation (480, 500, 520 kW at US$180, 210, 250 per kW-month, from mid-deployment) | 4.92 | 5.83 | 7.02 |
| Energy (380, 415, 450 kW; PUE 1.3, 1.4, 1.5; 7, 9, 12 cents per kWh; once running) | 1.36 | 1.95 | 2.84 |
| Support, years 1 to 3 (3%, 5%, 8% of hardware a year; often bundled) | 1.61 | 2.96 | 5.30 |
| Post-warranty support, years 4 and 5 (8%, 10%, 12% of hardware a year) | 2.86 | 3.94 | 5.30 |
| Team (5.5, 6.5, 7.5 FTE, from mid-deployment) | 5.23 | 7.52 | 10.13 |
| Cluster management and monitoring | 0.48 | 0.96 | 1.60 |
| Insurance (0.3%, 0.4%, 0.5% of hardware a year) | 0.27 | 0.39 | 0.55 |
| Residual value at year 5 (25%, 15%, 10% of servers) | -4.00 | -2.52 | -1.76 |
| Cost of capital (6%, 8%, 10%) | 3.94 | 5.67 | 7.97 |
| Five-year all-in cost | 35.3 | 47.7 | 62.9 |
| Optional: NVIDIA AI Enterprise, five-year list | 4.61 | 4.61 | 4.61 |
256 GPUs have 11.2 million GPU-hours in five years. Deployment time and failures leave fewer productive ones.
| Hours | Low | Mid | High |
|---|---|---|---|
| Months before useful work | 6 | 9 | 12 |
| GPU-hours lost to failures and RMA | 1% | 2% | 3% |
| Productive GPU-hours | 9.99M (89%) | 9.34M (83%) | 8.70M (78%) |
Team, colocation, support and cost of capital add up to US$25.9M in the middle case, well over the US$16.8M the servers cost.
The denominator matters as much as any line. Each month of delay and each point of idle time raises the cost of a useful hour, because colocation, staff and capital keep running regardless. Few owners use every productive hour, so plan around 60% to 80% of them when you divide the total.
B300 vs GB300 cost
Run on the same assumptions, owning HGX B300 comes out slightly cheaper per GPU than GB300 NVL72. Our GB300 NVL72 cost model works it through for a four-rack site.
| HGX B300 | GB300 NVL72 | |
|---|---|---|
| Unit you buy | 8-GPU node | 72-GPU rack |
| NVLink domain | 8 GPUs | 72 GPUs |
| Cooling | Air or liquid | Liquid |
| Hardware per GPU | US$63K to US$69K (our range) | US$51K to US$90K (reported) |
| Five-year all-in, middle case | US$47.7M for 256 GPUs | US$13.6M per 72-GPU rack |
The GB300 premium buys the 72-GPU NVLink domain, so a model too big for one node can run tensor- and expert-parallel inside the rack instead of across the network. If your models fit in 8 GPUs, or your hall can't take liquid cooling, we'd pick HGX B300. It's the cheaper way to run Blackwell Ultra.
B300 cost per GPU and per cluster, for a memo
- HGX B300 server: about US$500K to US$550K in 2026 (the upper part of the US$430K to US$550K reported), or US$63K to US$69K per GPU. One OEM's list price is US$785,607.
- 32-node, 256-GPU cluster capex with fabric, storage, fit-out and spares: about US$19M to US$24M.
- Five-year all-in cost, with team, support, software, insurance, cost of capital and residual value: about US$35M to US$63M, US$48M in the middle case.
- Productive GPU-hours over five years: 8.7M to 10.0M before utilization.
- NVIDIA AI Enterprise, if you need it: about US$4.6M more at five-year list.
- Team: about five people for round-the-clock cover at one site, 5.5 to 7.5 with network, storage, security and vendor management, and three to six months to hire each specialist.
- Not included: building a new hall, and tax effects. Depreciation schedules are in GPU depreciation and useful life.
How to buy B300, or have it built and run
Buying means an OEM or NVIDIA partner quote for HGX or DGX B300 servers, plus the fabric, storage, colocation lease and team covered above. For the same breakdown at 8 and 64 nodes, see how much a GPU cluster costs.
If you own the servers but not the team, we can design, source and build the cluster and operate it on our software. If you'd rather not carry the capex, we reserve dedicated HGX B300 clusters from 8 HGX nodes (64 GPUs) or one NVL72 rack (72 GPUs), single-tenant, on a 60-month term. The contract includes an exit option, on terms set out in your proposal. Talk to us.
Questions this note answers
Mid-2026 reporting puts an 8-GPU HGX B300 server at about US$430,000 to US$550,000. One OEM's July 2026 TCO study prices its 8-GPU B300 server at US$785,607, a list price before discounts.
B300 is sold inside 8-GPU HGX systems rather than as a card. A whole server works out to about US$54,000 to US$69,000 per GPU at reported street prices, or US$63,000 to US$69,000 on the US$500,000 to US$550,000 we use.
On this note's assumptions, a 32-node, 256-GPU HGX B300 cluster needs about US$19M to US$24M of capex and costs about US$35M to US$63M to own over five years, or US$48M in the middle case, once team, support, software, insurance, cost of capital and residual value are counted.
Per GPU, slightly, on the same assumptions. The GB300 premium buys the 72-GPU NVLink domain and needs a liquid-cooled hall. If your models fit in 8 GPUs, HGX B300 is the cheaper way to run Blackwell Ultra.
No. OEMs sell air-cooled 8U HGX B300 systems as well as liquid-cooled 4U and 2-OU versions. The liquid-cooled chassis allow denser racks and the higher 1,400 W GPU power bin.
Through an OEM or NVIDIA partner, as 8-GPU HGX or DGX B300 servers, usually with the fabric, storage and integration on the same quote. Lead times, a colocation lease and a team come with it.
Sources
- GPUsmith: NVIDIA HGX B300 system page (mid-2026 pricing, power bins) (opens in a new tab)
- OEM study: on-premise vs cloud generative AI TCO (July 2026) (opens in a new tab)
- NVIDIA DGX B300 product page (opens in a new tab)
- NVIDIA DGX SuperPOD reference architecture: DGX B300, architecture (opens in a new tab)
- NVIDIA DGX SuperPOD reference architecture: DGX B300, storage architecture (opens in a new tab)
- OEM NVIDIA accelerated systems (HGX B300 air- and liquid-cooled) (opens in a new tab)
- SemiAnalysis: AI Neocloud Playbook and Anatomy (October 2024) (opens in a new tab)
- SemiAnalysis: How much do GPU clusters really cost? (2026) (opens in a new tab)
- CBRE: North America Data Center Trends H2 2025 (February 2026) (opens in a new tab)
- US EIA: Electric Power Monthly, average price by end-use sector (opens in a new tab)
- Meta: The Llama 3 Herd of Models (reliability section) (opens in a new tab)
- NVIDIA AI Enterprise licensing guide: pricing (September 2026) (opens in a new tab)
- NVIDIA AI Enterprise licensing guide: licensing (DGX bundling) (opens in a new tab)
- US BLS: Employer Costs for Employee Compensation, June 2026 (opens in a new tab)
- US BLS: Computer network architects, median pay (opens in a new tab)
- NVIDIA: DGX A100 launch at US$199,000 (May 2020) (opens in a new tab)
- Tom's Hardware: A100 server prices in China (October 2026) (opens in a new tab)