AI consumes energy in 2 phases: training and inference. The first has a finite cost per training run; the second keeps accumulating with use. One training run has a defined end. Inference follows demand, whether the model handles a few requests or millions. Microblog @antgrasso Training requires computation to develop the model. Once a training run is complete, the energy consumed during that run is a fixed total. The trained model can then be used repeatedly without being trained again for every request. But every request still requires computation. The more frequently the model is used, the more inference energy accumulates. The amount also depends on how much processing each request requires. A simple question and a complex analysis do not necessarily consume the same amount of energy. For organizations deploying AI, estimating energy consumption means considering both the training process and the workload generated throughout the life of the service. #AIEnergy #AIInfrastructure