Surely rented GPUs would be used for inference, which runs anywhere and needs to scale much larger than a training run.