First focus on inference. It is a lot costlier to have GPUs for inference.