There's nothing stopping any of us from writing a better Nvidia driver btw. LLMs are very helpful with reverse engineering.

Being fast and being as easy to program as CUDA are two different things.