Nvidia have less moat for inference workloads since inference is modular. AMD would be mistaken to go after training workloads but that's not what they're going after.