AI Data Center Power & Energy (2026)

The power ladder from one GPU to a whole data center: a single H100 draws 700W, a GB200 NVL72 rack ~120kW (over ten times a traditional rack), and large AI clusters reach hundreds of MW to GW scale. This table explains why AI is so power-hungry.

ItemPower/valueNotesBasis/source
NVIDIA H100 SXM5700 WHopper flagship GPUOfficial spec
NVIDIA H200 SXM~700 WHopper, 141GB HBM3eOfficial spec
NVIDIA B200 (air, HGX/DGX)1,000 WBlackwell, configurable TDPOfficial spec (air config)
NVIDIA B200 (liquid, GB200)1,200 WFull-spec liquid, ~500W over H100Media/GTC (configurable)
AMD Instinct MI300X750 W192GB HBM3, peak boardOfficial spec
AMD Instinct MI325X1,000 W256GB HBM3E, peak boardOfficial spec
Traditional rack (non-AI)~5–10 kW/rackGeneral enterprise/cloud rackIndustry typical (range)
GB200 NVL72 full rack~120 kW/rack72 Blackwell GPUs, liquid-cooledNVIDIA design value (field 130–132kW)
PUE — industry average1.54Power usage effectiveness, ~1 is bestUptime Institute 2024 survey
PUE — hyperscaler1.09–1.15Google fleet-wide TTM 1.09Operator self-reported
xAI Colossus (Memphis)~250 MW / 100k H100Single large training clusterMedia report (2024, fast-moving)

Method & sources

Retrieved 2026-08-28. **Three bases, do not mix**: (1) **Chip TDP** from official specs, highest confidence — but NVIDIA Blackwell is "configurable TDP": the same B200 is 1,000W in air-cooled HGX/DGX and 1,200W in liquid-cooled GB200, labeled by configuration; (2) **Rack power** — two incomparable bases: traditional (non-AI) racks ~5–10 kW; GB200 NVL72 is NVIDIA's ~120 kW design value (field deployments reported 130–132 kW); "typical H100 rack 30–40 kW" is derived (DGX H100 ~10.2kW official × servers per rack), not a single official figure; (3) **PUE** — industry average 1.54 (Uptime Institute) vs hyperscaler 1.09–1.15 (Google fleet-wide 1.09) are different populations, not interchangeable or averageable. **Cluster-scale power is the weakest (low confidence)**: xAI Colossus 250MW is a 2024 report; GW-scale figures are SemiAnalysis-style planning estimates that can be stale within months ("as reported early 2026"). Official chip-spec PDFs are binary and not machine-readable verbatim; figures reflect consistent multi-source citation of official specs. The rack multiple (AI ~12–17x traditional) is derived and shifts with the chosen 7–10kW / 120–132kW endpoints. Sources: Uptime Institute, Google efficiency, NVIDIA/AMD specs, TweakTown, SemiAnalysis.

Source: https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2024.GlobalDataCenterSurvey.Report.pdf

Retrieved: 2026-08-28

FAQ

What does "AI Data Center Power & Energy (2026)" cover?
It covers NVIDIA H100 SXM5, NVIDIA H200 SXM, NVIDIA B200 (air, HGX/DGX), NVIDIA B200 (liquid, GB200), AMD Instinct MI300X, AMD Instinct MI325X, Traditional rack (non-AI), GB200 NVL72 full rack, compared across: Item, Power/value, Notes, Basis/source.
What are the sources and methodology?
Retrieved 2026-08-28.
When was this data last updated?
The data was retrieved/updated on 2026-08-28.