Material Insights · Understanding the Material World

Large language models: Energy footprint

Energy in use

The energy it uses or converts while it works, and how efficiently

Consumption
Electricity for GPU servers that train and run models; AI servers took 15% of data-centre energy demand in 2024 [8]
Conversion
Turns electrical energy into computation; almost all of it ends as heat
Note
The IEA figure covers all AI servers, not only language models.

as of 2026-11-11

Energy links

The energy sources and technologies it uses, holds, converts, loses or gets back

Consumes
  • El Electricity AI servers used 24% of server electricity demand in 2024 [8].
Loses
  • Ht Heat The electricity used by AI chips ends as heat that data-centre cooling must remove.

as of 2026-11-11

Sources

  1. Data centers will use twice as much energy by 2030, driven by AI (IEA Energy and AI) Scientific American, 2025