Training an AI model draws attention, but running it to answer requests, called inference, is where most day to day cost and energy go once a model is in use. In 2026 much engineering effort has focused on making inference cheaper through techniques that shrink models and skip unnecessary computation without badly hurting quality.
These gains matter for the environment and for access, because cheaper inference lets more people and smaller organisations use capable models. Methods include trimming a model's size, reusing past computation and running only the parts of a model a given request needs. The quiet result is that the same quality of answer costs less energy than it did a year earlier.