This is flawed thinking - the models will exist even if every company who trained them fails, and then someone else will rehost them and offer them at inference rates. Inference is _astonishingly_ cheap.

If anything, prices would go _down_ as the rehosters wouldn't be saddled with debt from training models.