This. When I was still doing infra for a large recommendation engine, the training was what got the big GPUs. Inference was on CPU or (later) ASICs (may have been FPGAs?).