AI Energy & Datacenters

Committed capacity, power sources, hardware efficiency, and model-level energy estimates for major AI operators. Data: kinnarigroup/energydata.

Committed capacity by operator

Power source mix

Where operators are sourcing power — by source type and status.

US AI datacenter capacity growth

Hardware efficiency by generation

Tokens per kWh improving across GPU generations — a proxy for compute efficiency gains.

Model energy estimates

Energy cost per million output tokens across frontier models.

Usage calculator

Estimate the energy footprint of a given token workload.

Measured inference efficiency (TokenPowerBench)

Empirically measured joules per token for open-weight models on H100 hardware. Batching is the biggest lever: throughput mode at batch=32 cuts energy per token by 3–4×.

Data tables