How though? Is there any legal cases where this has been sucessfully done before? And even they couldn't NVIDIA drag their feet on it for years in court?

Personally I think a bigger risk than that is AI models getting good enough to legitimately just reverse engineer and open source CUDA directly.