This could have been AMD, but they choose to sabotage themselves with half-baked APIs that don't even support all of their GPUs. Hell, ROCm didn't even support Windows until recently.
Windows is horrendously annoying to implement anything on and completely absent from the HPC field. Maybe people are trying to do AI training things on it now but that also seems likely to be Linux boxes.
Shipping rocm releases that don't run on a bunch of their GPUs is just an absurd choice, there's no way to defend that one.
Everyone talks about the chips, but Nvidia's true competitive advantage is CUDA. Only recently has PyTorch added AMD support, and benchmarks I've seen have shown that AMD has a lot of work to catch up.
Frankly, it's shocking how bad they are at it.