AI has a physical footprint: data centres consume electricity and water, and hardware requires resources to manufacture.
Where the Impact Comes From
- Training large models uses substantial computing power over weeks or months.
- Inference — running models for users — adds up because it happens constantly, and for widely used services it can exceed training.
- Cooling data centres uses energy and, in many facilities, water.
- Hardware manufacturing and disposal carry their own environmental costs.
Factors That Matter
- The carbon intensity of the electricity grid where computing runs.
- Data centre efficiency.
- Model size and how efficiently it's served.
- How much computing an application actually needs.
Reducing Impact
- Use the smallest model that meets requirements.
- Reuse pretrained models instead of training from scratch.
- Cache results and avoid redundant requests.
- Batch non-urgent work.
- Choose regions and providers with cleaner energy and published efficiency data.
- Use efficiency techniques such as quantisation and distillation.
- Avoid unnecessary experiments; track what's already been tried.
Measure and Report
Estimate the energy and emissions of significant training runs and high-volume inference. Providers increasingly publish data to support this.
Weigh Benefits
Environmental cost should be weighed against the value an AI system delivers — including cases where AI helps reduce emissions elsewhere, such as optimising energy use.