Exactly. Coding for inference is solved. CUDA is no longer a moat.