IT's cute you think they're gonna do any full training of a model. As soon as they can extract cash from the machine, the better.

This is low effort thinking, and a low effort comment. They have a lot of cash. They do not think they have achieved a "city of geniuses" in a datacenter yet. They are racing against two high quality frontier model teams, with meta in the wings. They have billions of dollars in cash that they are currently trying to spend to increase their datacenter capacity.

Any compute time spent on inference is necessarily taken from training compute time, causing them long term strategic worries.

What part of that do you think leads toward cash extraction?