Every AI feature has a cost that rarely appears in the plan: electricity. The IEA's 2025 *Energy and AI* report puts numbers on it, and they're large.
The figures
- Global data-centre electricity demand more than doubles by 2030 to ~945 TWh — more than Japan's entire consumption today.
- AI-optimised data-centre demand more than quadruples by 2030.
- AI was 5–15% of data-centre power recently; could hit 35–50% by 2030. US demand alone exceeds 400 TWh by 2030.
AI's marginal cost feels like a per-token line item. At scale, it's a power-grid problem.
Why builders should care
- Inference is recurring; at volume, efficiency is a feature.
- Smaller models win: Stanford notes a model matching 2022's flagship now runs with ~142× fewer parameters.
- Architecture decides the bill — caching, retrieval, and routing easy work to cheap models cut energy and cost together.
Sources
- IEA — Energy and AI
- Stanford HAI — 2025 AI Index
Written by ivector
Start a project →