Look beyond
the headline.
Training and everyday use create different demands. The location of a data centre, the electricity mix and the way a model is deployed all change its footprint. An efficiency claim needs a boundary before it needs a headline.
- 01Separate a model’s training footprint from ongoing inference.
- 02Ask where electricity, water and equipment come from.
- 03Compare the proposed service with a realistic alternative.
What the figure
can tell us.
Reported emissions from training GPT-3. One model, one training run—not the footprint of every AI system.
Read Patterson et al., 2021 ↗Demand is growing. The future is a projection.
Global data-centre electricity demand: 415 TWh in 2024, projected to reach about 945 TWh in 2030. This includes data centres serving uses beyond AI.
IEA · Energy and AI (2025) ↗Read the source for its methods and limits. A published estimate, observed result and future projection answer different questions. The research figures describe this topic; they are not learner scores.
A benefit.
A consequence.
What it can give
More efficient models and cleaner electricity can reduce emissions per task. AI can also support environmental monitoring and help identify changes that would otherwise be difficult to see.
What it can cost
More use can outweigh those efficiencies. Cooling draws on water resources, chips carry an embodied footprint, and the communities nearest infrastructure may bear costs far from its users.
The judgement depends on context. Which effects matter most, who experiences them, and what evidence would change the decision?