What's in the rack
NVIDIA's GB300 NVL72 joins 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single rack, linked so they behave as one large accelerator[1]. Vendors such as Lenovo, Supermicro, HPE and Dell build and ship the racks, and their published figures differ, so always work from your vendor's datasheet.
How much power a rack uses
There's no single number. Each figure below means something different, and only your vendor's configured electrical envelope should be designed to.
| Figure | Value | What it means |
|---|---|---|
| Lenovo, thermal design power | 135 kW[2] | Design figure, not a measured average |
| Lenovo, peak | Up to 155 kW[2] | Depends on workload and power behaviour |
| Supermicro, operating range | 132 to 140 kW[3] | One vendor's operating band |
| Vertiv reference design | Up to 142 kW per rack[4] | What the infrastructure is built for |
| HPE, supported power | Up to 198 kW per rack[5] | Supported envelope; confirm with HPE |
| NVIDIA power shelves, nameplate | 8 × 33 kW = 264 kW[7] | Installed capacity with redundancy, not draw |
For comparison, NVIDIA's own diagram of the earlier GB200 NVL72 labels it at about 120 kW[9].
Inside the rack, power shelves take three-phase AC (346 to 480 V between phases) and turn it into 50 V DC for the trays[8]. That busbar voltage has nothing to do with your grid connection, which in the UK is usually 11 kV stepped down to 400 V.
From rack power to grid supply
Grid connections are agreed in kVA, not kW. To convert, divide real power by the power factor: 140 kW at a power factor of 0.95 is about 147 kVA[11]. Then add everything that isn't a GPU rack: coolant pumps and dry coolers, air handling for the 10% of heat that still goes to air, UPS losses, lighting and controls, and headroom for the hottest day of the year.
PUE, the ratio of total facility energy to IT energy, describes that overhead across a year[12], but a site's connection has to cover the worst half hour, not the average. That's why we size supply from a full design rather than a multiplier.
How many racks fit your spare import?
Agreed capacity minus your highest half-hourly demand. A rough figure is fine.
7 racks504 GPUs
1,050 kW of IT on about 1,510 kVA, in one Hexi pod. The next rack needs about 1,700 kVA.
Hexi engine estimate for a reference UK site, including cooling, UPS losses and a 40°C heatwave. Your site check confirms it.
In the Hexi engine's reference design, one GB300 rack needs about 350 kVA of supply, four need about 930 kVA, seven about 1.5 MVA and eight about 1.7 MVA, all checked against a 40°C heatwave.
Cooling
A GB300 rack is cooled directly at the chip. Lenovo puts the split at about 90% of heat to liquid and 10% to air[2]: GPUs, CPUs and NVLink switches are liquid cooled, while network optics, storage and power distribution boards still need air[2]. A whole hall with support racks shows more air: Vertiv's GB300 reference design rejects 78% of heat to liquid and 22% to air[10].
Flow and temperature
The warmer the coolant you supply, the more of it the rack needs. Lenovo's table runs from 59 litres a minute at 25°C to 177 litres a minute at 45°C, per rack[6]. Warmer water lets dry coolers do more of the work without chillers, which is why the supply temperature is a design decision, not a detail.
Loops and coolant
Racks sit on their own clean loop, fed by a coolant distribution unit (CDU) that exchanges heat with the facility loop outside. HPE's CDU handles up to 1,300 kW at a 4°C approach, enough for up to eight racks[5]. Lenovo recommends deionised water in the rack loop, with a glycol mix as an option, and ships racks dry[2].
The room still matters
Air around the rack has to stay between 10°C and 35°C and 20% to 80% humidity, non-condensing[6].
Size and weight
HPE's populated rack is 600 mm wide, 2,298 mm tall (48U) and 1,068 mm deep, or up to 1,710 mm with extensions[5]. It weighs about 3,300 lb, roughly 1,500 kg[5]. Floors, ramps and doors on the delivery route all have to take that, and vendors list floor loading as part of site readiness[3].
What this means for a site
Most existing buildings weren't designed for 140 kW in 0.6 m² of floor, a liquid loop and 1.5 tonnes per rack. That's why a Hexi pod puts the racks in factory-built halls designed around them: identical halls with their coolant units, liquid loop and air handling, on a prepared base next to the site's existing supply. The Pod 1 specification lists everything included, and the configurator shows it in 3D for any rack count.
Questions people ask
- How much power does a GB300 NVL72 rack draw?
- It depends on the vendor's configuration and the workload. Published figures range from 132 to 140 kW operating (Supermicro, via Cisco) to 135 kW TDP with peaks up to 155 kW (Lenovo). Vertiv designs its reference infrastructure for up to 142 kW per rack.
- Is the quoted rack power real draw or nameplate?
- Usually neither is a measured average. TDP and operating ranges are design figures; the eight 33 kW power shelves in NVIDIA's reference rack add up to 264 kW of nameplate output, which is redundancy headroom, not what the rack uses.
- Is a GB300 rack fully liquid cooled?
- Mostly. Lenovo puts the split at about 90% of heat to liquid and 10% to air, because network optics, storage and power distribution boards are still air cooled. The room still needs air handling.
- How much site supply does a GB300 pod need?
- In the Hexi engine's Pod 1 design, seven racks (1.05 MW of IT) need about 1.5 MVA of import, including cooling, UPS losses and a 40°C heatwave. Four racks need about 930 kVA.
- How heavy is a GB300 NVL72 rack?
- HPE quotes about 3,300 lb populated, roughly 1,500 kg, in a 600 mm wide, 48U rack. Floors and delivery routes have to be checked for the actual configuration.
Sources
- 1. GB300 NVL72. NVIDIA, accessed 11 October 2026.
- 2. Lenovo NVIDIA GB300 NVL72 rack scale AI. Lenovo Press, updated 30 August 2026.
- 3. Supermicro NVIDIA GB300 NVL72 datasheet. Cisco, updated 15 September 2026.
- 4. Vertiv cooling and power reference architecture for GB300 NVL72. Vertiv, 11 June 2025.
- 5. Accelerate AI at scale with NVIDIA GB300 NVL72 by HPE. Hewlett Packard Enterprise, 2025.
- 6. GB300 NVL72 environmental specifications. Lenovo, accessed 11 October 2026.
- 7. DGX SuperPOD GB300 reference architecture. NVIDIA, accessed 11 October 2026.
- 8. Power shelves. Flex, accessed 11 October 2026.
- 9. Multi-node NVLink systems partition guide. NVIDIA, 2 July 2025.
- 10. Reference design 028. Vertiv, August 2025.
- 11. How do I convert kW to kVA?. Schneider Electric, modified 15 December 2025.
- 12. Does the spread of direct liquid cooling make PUE less relevant?. Uptime Institute, 5 November 2021.
This guide is general information, not legal, engineering or financial advice. Hexi engine figures are estimates until a firm offer. Spotted an error? Email louis@hexi.industries and we'll correct it.