In 2023, inference consumed one-third of AI compute. Now it's two-thirds. Training stays concentrated in massive clusters. Inference fragments across specialized hardware everywhere. Watch where the actual computational work migrates during technology maturation.