[bangaladore]

Makes sense to me. Inference on these models will be a race to the bottom. Hosting inference themselves will be a waste of compute / dollar for them.