I also think the space of products that involve training per-customer models is quite large, much larger than might be naively assumed given what is currently o

I also think the space of products that involve training per-customer models is quite large, much larger than might be naively assumed given what is currently out there. It may be true that inference is 100x larger than training in terms of raw compute. But I think it very well could be that inference is only 10x larger, or same-sized. And besides, you can look at it in terms of sheer inputs and outputs. The size of data yet to be trained on is absolutely enormous. Photos and video, multimedia. Absolutely enormous. Hell, we need giant stacks of H100's for text. Text! The most compact possible format!