Inference is already cheap enough. It's basically free for the quantities people actually want to read.

LLMs only become expensive if you let them generate 1000 lines of code every 2 minutes.