GPU Cluster Power Requirements — kW per Rack, MW per Cluster

Written and maintained by Haink's procurement and allocation advisory team · Updated September 2026

Manufacturer and OEM figures verified as of 19 September 2026 against nvidia.com, docs.nvidia.com, supermicro.com and hpe.com. Reviewed with each new system generation.

A Blackwell-generation GPU cluster draws about 1.8 kW per GPU at the system level, before networking, storage and cooling. NVIDIA rates the eight-GPU DGX B200 at about 14.3 kW maximum, and HPE rates the 72-GPU GB300 NVL72 rack at 132 kW nominal and advises data centers to provision 192 kW of busway per rack (figures checked 19 September 2026). On that basis, 1,000 GPUs need roughly 1.8–2.7 MW of IT capacity and 8,000 GPUs roughly 14–22 MW, before cooling and other facility overhead. Per rack, the range runs from about 29 kW for two eight-GPU servers to 130–155 kW for one NVL72 rack. In a 2025 survey by the Uptime Institute, 82% of operators said they ran no racks above 30 kW. The number that decides a project is rarely the total. It is how much of that power is contracted, energised and accepted at specific rack positions, and on which date.

How much power does one GPU actually need?

The planning number is the system, not the chip. A GPU's own rating leaves out the CPUs, memory, network adapters, storage and fans that share its chassis. Manufacturers publish a system figure, and that is the figure a data-center operator provisions against.

Per-GPU values are calculated from the published system figures. Two things follow from the table. The step from Hopper to Blackwell raised system power per GPU by about 40% (10.2 kW against 14.3 kW for the same eight GPUs). And NVIDIA's own product pages give no rack power figure for the NVL72 systems; the numbers come from OEM datasheets, which differ by vendor and by what they measure.

Nominal, peak and provisioned are three different numbers. HPE's QuickSpecs for the GB300 NVL72 state a thermal design power of "132 kW nominal", an electrical design peak "of approximately 155 kW", and a recommendation that data centers "provision their bus ways to be able to support 192 kW" (version 3, dated 8 September 2026). Cooling is sized on the first number, the electrical path on the last. A site that plans on nominal power alone is short by about 45% on the electrical side for this system.

How many kW per rack does a GPU cluster need?

It depends on the form factor and on how densely the operator lets the racks be filled.

• Eight-GPU servers. A DGX B200 occupies 10 rack units. Two per rack is about 28.6 kW, three is about 42.9 kW, four is about 57.2 kW, calculated from the 14.3 kW maximum. The rack count is a facility decision, not a server decision: the same cluster can be spread over more, lighter racks if the hall allows it.
• Rack-scale systems. An NVL72 is one rack and can't be split. Supermicro gives 125–135 kW operating power for the GB200 NVL72 and 132–140 kW for the GB300 NVL72. NVIDIA's contribution of the GB200 NVL72 design to the Open Compute Project, published 15 October 2024, specifies 120 kW of cooling capacity per rack.

Compare that with the installed base. The Uptime Institute Global Data Center Survey 2025 (published July 2025) found that 82% of responding operators had no racks above 30 kW. Two eight-GPU Blackwell servers already reach that line. A single NVL72 rack needs four to five times it. For most existing halls, Blackwell is not an upgrade of the racks already there. It needs a different room, or a room rebuilt for it.

How many MW does a 1,000, 4,000 or 8,000 GPU cluster need?

The table scales the published system figures. It shows compute load only: the GPU systems themselves, without the network fabric, storage, management servers or cooling, which are sized from the actual bill of materials and site design.

Calculated from NVIDIA's DGX B200 figure and HPE's GB300 NVL72 QuickSpecs, both checked 19 September 2026. NVL72 counts are rounded up to whole racks. These are sizing references, not a design: an operator's own figures for the specific configuration replace them as soon as they exist.

What sits on top of compute. Cooling and power conversion are added through power usage effectiveness (PUE), the ratio of total facility energy to IT energy. The same 1,000-GPU compute load of 1.79 MW becomes about 1.95 MW at a PUE of 1.09, which Google reports as its 2025 fleet average, and about 2.75 MW at 1.54, the weighted industry average in the Uptime Institute's 2025 survey. At 8,000 GPUs the spread is about 15.6 MW against 22.0 MW. Which end a project lands on is a property of the site, not of the GPUs.

PUE is an annual average, so it understates the grid connection. The connection has to carry the hottest hour of the year at full load, not the average. And redundancy, such as N+1 or 2N on the electrical path, adds equipment rather than load: it doesn't raise the consumption figure, but it raises what has to be built and commissioned before the first rack runs.

What does "available power" mean, and how is it proven?

The phrase is used for at least four different things. Only the last one runs GPUs.

A site can be large on the first line and have nothing on the last. A headline megawatt figure for a campus usually describes the first or second line. A GPU order needs the fourth, in the right number of positions, at the right density, on the delivery date.

Why a GPU supplier asks about it. Power is not only an engineering question in GPU procurement. The BIS counter-diversion guidance of 13 May 2025 lists as a red flag a data center that "does not or cannot affirm it has the infrastructure (e.g., power/energy, cooling capacity, or physical space needed to run servers containing advanced ICs)". It recommends that exporters request "a written attestation from the data center" that the end user is authorised to operate at the location and that the site has the infrastructure for the servers being exported. It also recommends evaluating data centers able to operate advanced-IC servers above 10 megawatts (guidance still published on bis.gov, checked 19 September 2026). A request for thousands of GPUs at a site that can't show power for them doesn't read as ambitious to a reviewer. It reads as a quantity the site doesn't explain. How the site fits into the end-user review is covered in how the end user, site and use are checked together.

Why does energisation set the delivery date?

Because GPUs can't run in a position that isn't live, and hardware that arrives early waits. For large clusters the electrical path usually has the longest lead of anything in the project.

The International Energy Agency's Energy and AI report (April 2025) put numbers on the constraint: "wait times for critical grid components such as transformers and cables have doubled in the past three years", new transmission lines "can take four to eight years in advanced economies", and "around 20% of planned data centre projects could be at risk of delays" unless grid risks are addressed. None of that is visible in a GPU quote.

Three practical consequences follow.

1. The binding date is the handover of positions, not the ship date. A delivery schedule that runs ahead of the operator's commissioning schedule produces hardware in storage, with warranty and depreciation running.
2. Large orders are phased to energisation. Each phase is sized to the positions that will be live when it arrives. How phases and tranches are built around this is covered in how each phase is tied to site energisation.
3. A slip in the site moves every later phase, whatever the state of supply. Why quoted hardware dates move is covered in NVIDIA GPU lead times.

What does an NVL72 rack require from the data center?

An NVL72 is delivered and accepted as one rack-scale system, and it can't be spread across lighter racks. That turns several facility details from preferences into conditions. The figures below come from the OEM documents cited, checked 19 September 2026; other OEMs' versions differ in detail.

Power feed. HPE's GB300 NVL72 uses eight power shelves, each fed at 400 V or 480 V and 60 A through IEC 60309 connectors, with top or bottom feed. Supermicro's GB200 and GB300 racks likewise use eight 1U shelves of 33 kW, 132 kW in total. The operator has to bring those feeds to the position and, following HPE's advice, size the busway for 192 kW per rack rather than the 132 kW nominal load.

Liquid loop. NVL72 systems are liquid-cooled at the chip. The heat has to leave through a coolant distribution unit (CDU). Supermicro lists an in-rack CDU of up to 250 kW, an in-row CDU of up to 1.8 MW serving up to eight racks, and a liquid-to-air sidecar of up to 200 kW for the GB300 NVL72; HPE lists a CDU rated for 1.3 MW supporting up to eight racks. The liquid-to-air option matters for sites without facility water, but it moves the heat into the room, so the hall's air system still has to remove it. The physics of direct liquid cooling is covered in liquid cooling for AI GPU servers.

Floor and space. HPE gives an approximate fully loaded weight of 3,300 lb, about 1,500 kg, for a rack with a footprint of 600 × 1,068 mm. That is roughly 2,300 kg per square metre over the rack's own footprint (calculated). HPE lists a height of 2,495 mm and Supermicro 2,236 mm, so door openings, lifts and raised floors along the delivery route have to be checked, not only the position itself.

What the operator has to commit to. In practice, an NVL72 order depends on the operator confirming in writing:

• the number of positions, the power per position and the busway rating;
• the liquid loop, the CDU arrangement and who operates and maintains it;
• floor loading along the route and at the position;
• the commissioning and handover date for each group of positions;
• that the named end user has the right to place and operate the hardware there.

The right to place the hardware. The last point is the one most often missing. A site with capacity is not a site where a given buyer may install hardware. That right comes from a colocation agreement, or at an early stage a letter of intent for capacity, that names the end user. For advanced GPUs it is also exactly what the BIS guidance asks the exporter to obtain an attestation for. Without it, the power figures above describe the building, not the project.

What should be settled before a GPU order is sized?

In order, because each answer constrains the next:

1. The system and its published power, including nominal, peak and provisioning figures where the OEM gives them.
2. The density the hall supports, in kW per rack, and whether it has liquid cooling or can add it.
3. Contracted capacity in writing, with the end user named, and the kW and density stated.
4. A commissioning and handover schedule for the positions, phase by phase.
5. The facility overhead at that site, so the total draw and the grid connection match.
6. The GPU quantity, which is the last number, not the first: it is what the energised positions can hold on each date.

The minimum site information a supply request starts from is listed in what is needed for a preliminary assessment.

Frequently asked questions

How much power does a 1,000-GPU cluster need?

About 1.8 MW for the Blackwell compute systems alone, calculated from NVIDIA's 14.3 kW maximum for an eight-GPU DGX B200 (checked 19 September 2026). For GB300 NVL72 racks, HPE's figures give 1.85 MW nominal and 2.69 MW of provisioned busway for 14 racks. Networking, storage and facility overhead come on top; at the Uptime Institute's 2025 average PUE of 1.54, the compute load alone becomes about 2.75 MW at the facility.

How many kW per rack does a B200 cluster need?

It depends on how many servers go in each rack. At 14.3 kW per DGX B200, two per rack is about 28.6 kW and four is about 57.2 kW. Rack-scale NVL72 systems run at 125–140 kW per rack according to Supermicro's datasheets (checked 19 September 2026).

What is the power draw of a GB300 NVL72 rack?

HPE's QuickSpecs (version 3, 8 September 2026) give 132 kW nominal, about 155 kW peak, and advise provisioning 192 kW of busway per rack. Supermicro gives 132–140 kW operating power. NVIDIA's product page gives no rack power figure (checked 19 September 2026).

Can an existing enterprise data center host Blackwell GPUs?

Sometimes, at low density. Eight-GPU servers can be spread at two per rack, about 28.6 kW, which is near the 30 kW line that 82% of operators in the Uptime Institute's 2025 survey did not exceed. NVL72 racks need 120–155 kW per position, direct liquid cooling and a floor rated for about 1,500 kg per rack, which most existing halls weren't built for.

Why does a GPU supplier ask about site power?

Because it is part of the end-user review. The BIS guidance of 13 May 2025 treats a data center that cannot affirm it has the power, cooling and space for the servers as a red flag, and recommends a written attestation from the data center. A quantity that the site can't power is a quantity the request can't explain.

Is PUE enough to size the grid connection?

No. PUE is an annual average. The grid connection and the electrical path are sized for peak load at the worst design conditions, and for the OEM's peak or provisioning figure rather than nominal power.

What a preliminary feasibility assessment checks alongside site power, and how Haink applies it before any supply request: Preliminary feasibility assessment →