some stuff is still CUDA-dependent (a lot of scientific code). other stuff is not (major ML frameworks and many math libraries). nobody said CUDA was obsolete, just that there are great options for many use cases that no longer rely on it.