AMD was but focused on hpc instead on ml. They have very large installations, but the card is tuned for fp32+ calculations not lower precisions like NVIDIA.
You provide a BLAS implementation, and your primary customer can take it from there.
As opposed to Nvidia's current market, which spans HPC to AWS/Azure/GCP to corporate datacenter to "I followed the directions on Reddit" personal machines.