GPU & Compute Power Load Calculator
Estimate the raw IT power an AI cluster draws: chip TDP × utilization × count, plus server overhead for CPUs, memory, and networking.
chips × TDP × utilization
with server overhead
32-GPU rack estimate
at this utilization, 8,760 h
Results update live as you type. For planning and field-check estimates — always verify against applicable standards and equipment ratings.
How we calculate this →How AI compute load is estimated
The standard data center formula — server power × server count ÷ 1,000 — needs one adjustment for AI facilities: the GPU dominates. A single H100 draws up to 700 W, and Blackwell-generation parts push past 1,000 W per chip, so the honest way to size an AI hall is to start from accelerator TDP, scale it by realistic utilization, and multiply by chip count. Training clusters routinely sustain 70–90% of TDP for weeks; inference fleets are burstier and often average far less.
TDP alone understates the server, though. CPUs, HBM and system memory, NICs, NVLink switches, and chassis fans add roughly 25–45% on top of GPU draw in dense AI servers, which is why this calculator carries an explicit overhead slider rather than burying the assumption. The output here is IT load only — what the racks themselves consume. Cooling and electrical losses sit on top of it, which is exactly what the cooling load and PUE calculators on this site handle next.
Two practical notes from the field: utilities and interconnection queues care about your peak coincident demand, not your average, so size interconnection requests from near-full utilization; and per-rack density is the number that drives your cooling architecture decision — above roughly 40–50 kW per rack, air cooling stops being the economical answer.
Related reading
Frequently asked questions
What is GPU TDP and is it the same as actual power draw?
TDP (thermal design power) is the sustained power a chip is rated to draw under heavy load, not a hard ceiling and not an average. Real workloads fluctuate with utilization, so actual draw over time is typically TDP multiplied by an average utilization factor rather than TDP itself.
How much power does an NVIDIA H100 use?
The H100 SXM variant is rated at up to 700 W, while the PCIe variant is rated at 350 W. Newer Blackwell-generation chips like the B200 and GB200 push past 1,000 W per GPU.
Why add server overhead on top of GPU power?
CPUs, HBM and system memory, NICs, NVLink switches, and chassis fans all draw power beyond the GPUs themselves — typically 25–45% on top of GPU draw in dense AI servers — so ignoring overhead understates the real IT load a rack pulls from the facility.