They actually have already - PyTorch works straight up with ROCm, and so does Tensorflow. There is a little faffing to do to get it to work but they've made great progress in the last six months.
Except consumer GPUs don't officially support ROCm, despite consistent pressure from users for years. And there's no indication of when that situation will change.
CUDA is successful because the same software works on low-powered laptop GPUs and expensive datacenter GPUs.
CUDA is successful because the same software works on low-powered laptop GPUs and expensive datacenter GPUs.