In this note
A GB300 NVL72 rack sells for roughly US$3.7M to US$6.5M, depending on the source and on what the quote includes. Purchase orders reviewed in August 2026 came in just under US$5.0M, about US$69,000 per GPU.
Owning the rack for five years costs a lot more than buying it. In our middle case it's about US$13.6M per rack (US$8.7M to US$20.4M across our cases) once you count the team, support, failures, software, the months before it does useful work, insurance, resale value and the cost of capital. The rack is about 37% of that.
NVIDIA doesn't publish a list price. The rack prices below are public reporting, everything else is our estimate, and every line is in a table so you can drop in your own quotes.
How much does a GB300 NVL72 cost per rack?
| Source | Rack price | Per GPU (÷ 72) | What it covers |
|---|---|---|---|
| Loop Capital analyst estimate, via GPUsmith (opens in a new tab) | US$3.7M to US$4.0M | US$51K to US$56K | Training-class rack, derived from a large GB300 order |
| Data Gravity (opens in a new tab), purchase orders (August 2026) | Just under US$5.0M, about US$32K more with a 250 kW in-rack CDU | About US$69K | Small orders, as transacted |
| Tom's Hardware (opens in a new tab) source (March 2026) | US$6.0M to US$6.5M | US$83K to US$90K | Inference-optimized configuration |
The same Tom's Hardware report puts GB200 NVL72 at US$2.8M to US$3.4M, and its source warns that many of the quotes going around are ODM prices without a proper warranty. Before you compare totals, check what support term, spares and integration each quote includes.
Volume moves the price too. Data Gravity expects hyperscale buyers to do better than the purchase orders it saw, and a buyer of four racks won't get a hyperscaler's price for the same bill of materials. We use US$3.7M, US$5.0M and US$6.5M for the low, middle and high cases.
What's in a GB300 NVL72 rack
From NVIDIA's reference architecture (opens in a new tab) and an OEM product guide (opens in a new tab):
- 18 compute trays, each with two GB300 superchips (two Grace CPUs and four Blackwell Ultra GPUs), four 800 Gb/s ConnectX-8 ports and one BlueField-3 DPU
- 9 NVLink switch trays that put all 72 GPUs in one NVLink domain
- 6 or 8 power shelves of 33 kW each, on a 50 V DC busbar
- Cold plates, manifolds and quick disconnects for direct-to-chip liquid cooling
Nearly all of the price is silicon. Data Gravity puts the 72 GPU modules at about US$3.4M and the compute trays at about 87% of the rack's value, with NVIDIA content above 92% once you include the NVLink fabric. Morgan Stanley values the liquid-cooling parts at US$49,860 per rack, as reported by Tom's Hardware (opens in a new tab). Our GB300 NVL72 page covers what the rack asks of a site.
What you pay for on top of the rack
The model is for a buyer with four racks in a colocation hall that already supports liquid cooling. Shared items are split four ways.
Scale-out network. Each rack has 72 compute-fabric links at 800 Gb/s, 36 links at 400 Gb/s for storage and in-band management, and a 1 GbE out-of-band network. NVIDIA's design uses 144-port switches, eight leaves for every eight racks, and every link needs optics. We use US$250K, US$300K and US$500K per rack.
Storage. NVIDIA's guideline for an eight-rack unit is 90 GB/s of read and 45 GB/s of write at the standard tier, which means an all-flash parallel file system where much of the cost is software licensing. SemiAnalysis (opens in a new tab) assumes 2 TB per GPU in its large pre-training scenario and about 12 TB in its multimodal research one, so a rack needs roughly 150 to 860 TB. We use US$150K, US$350K and US$600K per rack.
Fit-out, freight and commissioning. Even a liquid-ready hall leaves you paying for the CDU and its water connection, power distribution to the shelves, freight, integration and acceptance. Data Gravity's version of these line items, with rack-attached networking included, takes a rack from just under US$5.0M to about US$5.7M deployed. We use US$150K, US$250K and US$400K, on top of the network line.
Spares. A pool of compute trays, PSUs, NICs, optics and quick disconnects, at 3%, 5% and 7% of the rack price.
Colocation and energy. Wholesale colocation for 250 to 500 kW in North American primary markets averaged US$195.94 per kW-month at the end of 2025, with vacancy at 1.4% (CBRE (opens in a new tab)). You pay on provisioned kW. OEM specifications list 132 to 135 kW per rack with peaks up to 155 kW, so we provision 140 to 155 kW at US$180 to US$250. 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)), and we use 7 to 12 cents.
Costs most build cases leave out
Team
A cluster serving production needs someone on call at all hours. A week has 168 hours and a person works about 40, so a round-the-clock rotation that allows for leave and training takes about five people at one site, whether you run one rack or four. Add part-time shares of network, storage, security, and vendor and facilities work, and we get 5.5, 6.5 and 7.5 full-time equivalents for a four-rack site. Each rack carries a quarter of that, at a loaded US$200K, US$250K and US$300K a year.
For scale, 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 compensation cost 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. Hiring each specialist takes three to six months by our estimate, so the team starts before the racks arrive; we pay it from halfway through deployment.
The same core team could run more racks, so this line falls per rack as you grow.
Support and maintenance
DGX GB300 (opens in a new tab) comes with three years of business-standard hardware and software 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.
Years four and five are post-warranty, at 8%, 10% and 12% a year. An OEM TCO study from July 2026 (opens in a new tab) uses 12% of system cost a year. Hardware here means the rack plus its share of network and storage: US$4.10M, US$5.65M and US$7.60M in the three cases.
Failures and RMA turnaround
While pre-training Llama 3 405B, 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 about 39 a year across four racks if Blackwell Ultra fails at H100 rates.
Each failure takes a four-GPU tray out for days while it's swapped from spares or sent back under RMA, and the job restarts from its last checkpoint. With two to ten days of turnaround and one to three hours of restart per event, four racks lose about 1% to 3% of their GPU-hours. That's the range we use. Nobody has published comparable failure data for Blackwell Ultra yet, and if your jobs need all 72 GPUs in a rack, one failed tray idles the whole rack until it's replaced, so you'd lose more. Our note on SLAs and goodput covers how to contract around this.
Time before useful work
Between the purchase order and the first production job come the colo search and lease, fit-out, delivery, installation, burn-in and acceptance. In leased colo space that takes 6 to 18 months (see the AI cluster deployment timeline). We take 6, 9 and 12 months out of a fixed five-year life that starts at the order. Capital goes out at the order, colocation and the team start halfway through deployment, and energy starts when the rack 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, our estimate, counting nine compute nodes per rack.
NVIDIA AI Enterprise (opens in a new tab) is optional, and we leave it out of the headline. It lists at US$4,500 per GPU for one year, or US$18,000 per GPU for a five-year subscription. NVIDIA's licensing guide (opens in a new tab) says it was part of the DGX software bundle on Hopper but has to be bought separately for Blackwell DGX systems, while the DGX GB300 product page still lists it in the software stack, so ask what your quote includes. At the five-year price it's about US$1.3M per rack.
Insurance, residual value and cost of capital
Insurance is 0.3%, 0.4% and 0.5% of hardware value a year, our estimate.
A five-year-old rack still has resale value. NVIDIA launched DGX A100 at US$199,000 (opens in a new tab) in 2020, and in October 2026 Tom's Hardware (opens in a new tab) reported A100 servers selling for up to US$82,000. That's about 40% of the launch price after six years, in a very tight market. The OEM TCO study assumes zero. We credit 25%, 15% and 10% of the rack price at year five. Decommissioning (de-install, drive sanitization, freight) comes out of that credit; we don't model it separately.
The capital goes out before the first useful hour. We charge 6%, 8% and 10% a year on capex as a level five-year annuity from the order, after deducting the present value of the residual. It's the interest you'd pay financing the purchase with a five-year loan and clearing the balance with the resale.
Five-year cost per GB300 NVL72 rack
Per rack, for a buyer with four racks in a third-party colocation hall. Team, tooling, spine switches and storage are split four ways.
| Line, five years | Low | Mid | High |
|---|---|---|---|
| Rack | US$3.70M | US$5.00M | US$6.50M |
| Network and storage share | US$0.40M | US$0.65M | US$1.10M |
| Fit-out, freight, integration | US$0.15M | US$0.25M | US$0.40M |
| Spares (3%, 5%, 7% of the rack) | US$0.11M | US$0.25M | US$0.46M |
| Capex | US$4.36M | US$6.15M | US$8.46M |
| Colocation (140, 150, 155 kW at US$180, 210, 250 per kW-month, for 57, 55.5, 54 months) | US$1.44M | US$1.75M | US$2.09M |
| Energy (120, 130, 140 kW; PUE 1.2, 1.3, 1.4; 7, 9, 12 cents per kWh; for 54, 51, 48 months) | US$0.40M | US$0.57M | US$0.82M |
| Support, years 1 to 3 (3%, 5%, 8% of hardware a year; often bundled) | US$0.37M | US$0.85M | US$1.82M |
| Post-warranty support, years 4 and 5 (8%, 10%, 12% a year) | US$0.66M | US$1.13M | US$1.82M |
| Team share (5.5, 6.5, 7.5 FTE over 4 racks) | US$1.31M | US$1.88M | US$2.53M |
| Cluster management and monitoring | US$0.14M | US$0.27M | US$0.45M |
| Insurance | US$0.06M | US$0.11M | US$0.19M |
| Residual value at year 5 (25%, 15%, 10% of the rack) | -US$0.93M | -US$0.75M | -US$0.65M |
| Cost of capital (6%, 8%, 10%) | US$0.92M | US$1.66M | US$2.81M |
| Five-year all-in cost per rack | US$8.7M | US$13.6M | US$20.4M |
| Optional: NVIDIA AI Enterprise, five-year list | US$1.30M | US$1.30M | US$1.30M |
Team, support, cost of capital and colocation together come to more than the rack does.
How many useful hours you get
72 GPUs over five years is 3.15 million GPU-hours. Deployment time and failures take some of those away before utilization comes into it.
| Hours | Low | Mid | High |
|---|---|---|---|
| Months before useful work | 6 | 9 | 12 |
| GPU-hours lost to failures and RMA | 1% | 2% | 3% |
| Productive GPU-hours | 2.81M (89%) | 2.63M (83%) | 2.45M (78%) |
To get your own cost per GPU-hour, divide the five-year total by the productive hours, then by the share of them you expect to use. The denominator moves the answer as much as any cost line. Colocation, staff and capital keep running through every month of delay and every idle hour, so each one raises the cost of the hours you do use. Few owners use every productive hour; 60% to 80% is closer to what most will see.
Hyperscalers land much lower on the same rack. They pay less for racks, power and capital, and spread their teams over tens of thousands of GPUs.
Numbers for a memo
- Rack price: US$3.7M to US$6.5M reported, just under US$5.0M in August 2026 purchase orders; US$51K to US$90K per GPU.
- Capex per rack with its share of network, storage, fit-out and spares: about US$4.4M to US$8.5M.
- Five-year all-in cost per rack, with team, support, software, insurance, cost of capital and residual value: about US$8.7M to US$20.4M, US$13.6M in the middle case.
- Productive GPU-hours per rack over five years: 2.45M to 2.81M before utilization.
- NVIDIA AI Enterprise, if you need it: about US$1.3M more per rack 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 work; three to six months to hire each specialist.
- Not included: adding liquid cooling to a hall that lacks it, and tax effects. Depreciation schedules are in GPU depreciation and useful life, and the same model for 8-GPU nodes is in HGX B300 cluster cost.
Buying, building or reserving GB300 NVL72
NVL72 racks are sold through NVIDIA's OEM and ODM partners, by the whole rack, and arrive with the liquid-cooling, power and fabric work above still to do. For how rack cost compares with 8-GPU nodes at different cluster sizes, see how much a GPU cluster costs.
Most of the lines above are work, not hardware: the site, the fit-out, the team and the months before the first job. If you own the racks, we can design, source and build the cluster and operate it for you. If you'd rather not own it, we reserve whole GB300 NVL72 racks on a 60-month term, single-tenant, in the region you choose. The contract includes an exit option, on terms set out in your proposal. Talk to us.
Questions this note answers
NVIDIA publishes no list price. Public reporting runs from US$3.7M to US$4.0M (a Loop Capital analyst estimate) up to US$6.0M to US$6.5M (a Tom's Hardware supply-chain source), and purchase orders reviewed in August 2026 came in just under US$5.0M.
Data center GB300 is sold as part of a rack, not as a single chip. Dividing the reported rack prices by 72 gives about US$51,000 to US$90,000 per GPU, before networking, storage and the facility.
72 Blackwell Ultra GPUs and 36 Grace CPUs, in 18 compute trays joined by 9 NVLink switch trays. Each GB300 superchip pairs one Grace CPU with two GPUs, and each compute tray holds two superchips.
On this note's assumptions, about US$8.7M to US$20.4M per rack, US$13.6M in the middle case, for a buyer with four racks in a colocation hall. That counts the team, support, failures, software, deployment time, insurance, cost of capital and resale value. The rack itself is about 37% of the middle case.
Yes, through NVIDIA's OEM and ODM partners, sold by the whole rack. It needs a hall with facility water and about 132 to 155 kW per rack.
Sources
- Tom's Hardware: NVL72 rack prices for GB200, GB300 and Vera Rubin (March 2026) (opens in a new tab)
- Data Gravity: How much does an NVIDIA NVL72 cost? (August 2026) (opens in a new tab)
- GPUsmith: NVIDIA GB300 NVL72 system page (Loop Capital estimate, power) (opens in a new tab)
- Tom's Hardware: Morgan Stanley on GB300 NVL72 liquid cooling content (November 2025) (opens in a new tab)
- NVIDIA DGX SuperPOD reference architecture with DGX GB300 (opens in a new tab)
- NVIDIA DGX GB300 product page (support and software) (opens in a new tab)
- OEM product guide: GB300 NVL72 rack-scale AI (opens in a new tab)
- OEM study: on-premise vs cloud generative AI TCO (July 2026) (opens in a new tab)
- OEM GB300 NVL72 rack specifications (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)
- SemiAnalysis: How much do GPU clusters really cost? (April 2026) (opens in a new tab)