Inference is latency sensitive so the front end is still relevant.