I thought that inference was now the majority of AI workload, not training.

in terms of raw token throughput and serving users it is, but in terms of R&D and long term lab capabilities, training is still the goose that lays the golden egg