CUDA has been compatible with the open kernel modules for a while, in fact now that I looked it up, there's at least a handful of features that requires the open kernel drivers, including 1 CUDA specific one (DMABUF support):

https://download.nvidia.com/XFree86/Linux-x86_64/550.78/READ...

From TFA:

> ...so far there's no indications of major shifts around going open-source on the user-space driver side, especially around their walled CUDA compute garden.