I think the “it’ll come crashing down” argument comes from cheaper inference, not more expensive. The frontier labs have to pay for these massive training runs. If inference gets cheap enough that you can run it yourself, why run it on their side? Open weights are (much) less than a year behind.