The race is to be the first to make a self-improving model (and have the infrastructure it will demand).

This is a winner-takes-all game, that stands a real chance of being the last winner-takes-all game humans will ever play. Given that, the only two choices are either throw everything you can at becoming the winner, or to sit out and hope no one wins.

The labs know that substantial losses will be had, they aren't investing in this to get a return, they are investing in it to be the winner. The losers will all be financially obliterated (and whoever sat out will be irrelevant).

I doubt they are sweating to hard though, because it seems overwhelmingly likely that most people would pay >$75/mo for LLM inference monthly (similar to cell phone costs), and at that rate without going hard on training, the models are absolute money printers.

[dead]