And Sam's (correct) point is that that doesn't work. If everyone tried to run the same amount of inference locally with the same size models that they're currently doing, the shortage would be even worse.