I think AMD is focusing on the "inference" side of ML, which doesn't really require CUDA or similar, and is what they believe a much larger market.
Time will tell if that strategy is going to pan out. Ceding the ML "training" market entirely to Nvidia is certainly a bold move