Engineering illustration · reviewed October 11, 2026
Power requirements depend on the GPUs, server design, networking, storage, cooling, and workload. NVIDIA specifies a maximum system power of 10.2 kW for an eight-GPU DGX H100. As an illustration, 1,000 H100 GPUs in 125 of those systems would have a combined maximum server power of 1.275 MW. That excludes external networking, storage, and facility overhead such as cooling; it is not a measurement of a particular cluster’s electricity use.
How to read this finding
Maximum rated power is different from average consumption. Megawatts measure power at a point in time; megawatt-hours measure energy over a period. A system drawing 1 MW continuously for one hour uses 1 MWh. GPU power ratings alone do not describe the whole facility, and an announced gigawatt of future site capacity is not proof of installed chips. This index does not meter electricity use, so its GPU counts cannot establish a global AI power total or a universal conversion between gigawatts and GPUs.
Citation estimates use inputs through September 1, 2026; fleet counts have an October 1, 2026 cutoff. The research-topic sample covers January 1 to June 1, 2025. See each answer for its period and limitations.
Benaich, Nathan. “How much power does a GPU cluster need?” State of AI Report Compute Index. Web page updated 2026-10-11; data periods as specified above.
