Full definition
Inference means running an already trained model to produce a response. Distinct from training, the creation of the model, a one-off but highly intensive event. In 2026, inference represents most of the energy consumed over a model’s life (around 60 to 90 %), because it repeats with every request and its volume is exploding. The real environmental issue of AI is therefore aggregate use, rather than training.
