Probably depends on the use-case: I mean, weren't the cryptocurrency people using AMD before they switched to custom FPGAs? But for sure, for the TensorFlow users, there doesn't seem to be a viable alternative to NVIDIA and there must be a reason why CUDA is more popular than OpenCL.
What I mean with ignoring consumer wishes is that they neither offer an open-source driver like AMD and Intel do and they do not even support others in doing so, like the Nouveau people. Subsequently, as a Linux user with NVIDIA hardware, you are either stuck with the buggy binary driver or the feature-incomplete Nouveau driver while AMD and Intel users have relatively stable and feature complete open-source drivers.
Why do you think Linus Torvalds likes Nvidia so much?
I am not sure what the status of AMD ROCm is, but I guess it will take a while until it becomes a viable alternative.
Here is the FFT example in ROCm: https://github.com/ROCm-Developer-Tools/HCC-Example-Applicat...
Compared to FFT in CUDA: https://github.com/drufat/cuda-examples/blob/master/cuda/fft...
1163 vs 216 lines of code! CUDA is really straightforward; ROCm is still low-level and incomprehensible (unless you are a C++ expert who happens to love low-level details).
This won't help current owners of Nvidia hardware (as myself), but you shouldn't blame RedHat for issues that obviously have been caused by Nvidia.