This is a great question. To my understanding the industry consensus is that for the big three providers, energy spend on inference had already surpassed training by summer last year, and the former's share only keeps increasing. The problem is that there's no hard data in public.

What we need to do here is write an article that makes a wild claim in either direction ("99% is inference!"), post it on HN, and wait for the comments to roll in that prove it right or wrong.