For inference alone they get far better utilization of the hardware because they can keep it fully loaded with maximum concurrency. Other than that they get to collect a lot of data for training.