[pbmonster]

Pretty sure that was scientists competing for 6 month training runs of new 100B+ parameter models, not coders burning through a couple of million tokens.