And a lot of GPU cycles happen before next token prediction, what's your point?