Size the power and cooling for an AI GPU deployment in seconds. Pick a GPU, set the count and your facility PUE, and get IT power, facility power, cooling load and annual energy cost — for H100, H200, B200, B300 and the GB200 NVL72 rack.
Power is the first constraint on any AI data center. Before you can rack a single GPU you need to know how much electrical capacity to provision and how much heat you will have to reject. This calculator turns a GPU choice and a count into the four numbers that drive the electrical and mechanical design.
It starts from the accelerator's rated board power (TDP) and works outward to the facility. Each step is a simple, labelled multiplier so you can see exactly where the numbers come from and adjust the assumptions to match your site.
Modern AI accelerators are dense. An H100 or H200 draws up to 700 W, a Blackwell B200 pushes 1,000 W air-cooled and 1,200 W liquid-cooled, and a GB200 superchip pairs two Blackwell GPUs with a Grace CPU at up to 2,700 W. A single GB200 NVL72 rack concentrates 72 GPUs into roughly 120 kW — an order of magnitude beyond a traditional enterprise rack, and the reason direct-to-chip liquid cooling is now standard for the highest-density deployments.
PUE decides how much of your utility bill actually reaches the GPUs. At a 1.1 PUE a liquid-cooled hall spends almost every watt on compute; at a 1.6 PUE a legacy air-cooled room burns 60% extra on cooling and distribution before a single token is generated. Across a multi-megawatt cluster running continuously, that gap is millions of dollars a year and megawatts of grid capacity you have to secure up front.
Because nearly all of that electrical power becomes heat, cooling load tracks IT power almost one-to-one. Getting these figures right early is what lets you lock the longest-lead items — transformers, switchgear, UPS and cooling plant — before they gate the whole build. That is exactly the sizing this tool is built to accelerate.
Rated board power for the accelerators and rack systems seeded into the calculator. Figures are published TDPs used for planning; confirm exact specs against the configuration you deploy.
| Model | Architecture | GPUs / unit | Rated power |
|---|---|---|---|
| NVIDIA H100 SXM | Hopper | 1 | 700 W |
| NVIDIA H200 SXM | Hopper | 1 | 700 W |
| NVIDIA B200 (air-cooled) | Blackwell | 1 | 1 kW |
| NVIDIA B200 (liquid-cooled) | Blackwell | 1 | 1.2 kW |
| NVIDIA B300 (Blackwell Ultra) | Blackwell Ultra | 1 | 1.4 kW |
| NVIDIA GB200 superchip (2×B200 + Grace) | Blackwell | 2 | 2.7 kW |
| NVIDIA GB200 NVL72 (full rack, 72 GPU) | Blackwell | 72 | 120 kW |
An H100 SXM has a rated board power (TDP) of about 700 W. A full 8-GPU HGX/DGX H100 server draws roughly 10 kW once you add the CPUs, memory, NICs, fans and power-supply losses — which is why this calculator applies a system-overhead factor on top of raw GPU TDP.
A GB200 NVL72 rack integrates 72 Blackwell GPUs and 36 Grace CPUs and draws on the order of 120 kW per rack, with integrated direct-to-chip liquid cooling. At a 1.4 PUE that is roughly 168 kW of facility power per rack before redundancy.
PUE (Power Usage Effectiveness) is total facility power divided by IT power. A PUE of 1.4 means for every 1 kW the servers draw, the facility uses 0.4 kW more on cooling, distribution and losses. Efficient liquid-cooled AI halls reach 1.1–1.25; legacy air-cooled rooms run 1.5–1.8.
Essentially all electrical power drawn by the servers is converted to heat, so the thermal load your cooling system must reject is approximately equal to the IT power. A 100 kW GPU load needs roughly 100 kW of heat rejection capacity.
It is a fast planning estimate using published TDPs and standard assumptions, accurate enough to scope power, cooling and budget. Final electrical and mechanical design depends on your specific servers, rack layout, ambient conditions and utility. Our engineers confirm the real numbers when you request a quote.
Our solutions engineers size power, cooling and compute together and confirm real lead times — no payment, no commitment, quotes back in about one business day.
Our engineers scope power, cooling and compute together. No payment, and quotes come back in about a business day.