“Vipul Ved Prakash estimates that AI inference workloads now constitute more than 50% of all GPU usage, a shift from training being the primary workload.”