If you think taking decades is an acceptable rate while others do it in a timely manner it's your own problem. For any normal user it's completely unacceptable and is the opposite of great (add to it, that even after decades of dragging their feet they only offer half cooked support and still can't even sort out upstreaming their mess). Garbage support is what it is.

AMD is notorious for not having ROCM support on in production currently sold GPUs, and horrendous bugs that actually make using the devices unusable.

I use AMD gpus on linux, I generally regret not just buying an Nvidia GPU purely because of AMDs lacklustre support for compute use cases in general.

Intel is still too new in the dGPU market to trust and on top of that there is so much uncertainty about whether that entire product line will disappear.

So at this point the CUDA moat makes is a non issue, on top of that what works works and keeps working, whereas with AMD I constantly wonder whether something will randomly not work after an update.

A timeline of decades for “features” your biggest consumers don’t care about is a reasonable tradeoff, even more so if actually pushing those features would reduce stability.