I don't think inference is subsidized, it's the training. So what happens is, there's no new models anymore or are released slower.