Would having their own distro even help? It seems like working around this would require forking from Linux at a lower level, and even that would only circumvent technical (rather than copyright) barriers.

There's an interesting discussion about the whole topic on the Phoronix forums about this. Some people claim that removing them and Nvidia's current behavior is a DMCA violation:

1. The kernel includes IP only licensed under GPLv2.
2. While a module linked against the kernel isn't necessarily a derived work which in turn would need to be licensed GPLv2 as well, there are specific interfaces that are meant for internal use and by their very nature would make your work derived if using them. These are the interfaces marked EXPORT_GPL_ONLY.
3. Using these interfaces with a module not licensed GPLv2, you taint the kernel and violate the licensing.
4. Removing the check, you aren't necessarily yet violating GPLv2, but you're removing a technical protection measure which is a violation of the DMCA.

It also raises the question why you'd remove checks that only prevent a possible GPLv2 violation if you're not violating GPLv2 anyways as Nvidia claims.

B) can't happen because of gpl. Even if it could, not many customers will move to an nvidia distro. ML people need good distros and good drivers.

If a hypothetical nvidia distro would speed up training by 10% but cause drop of productivity of humans of as small as 5%, no many will "buy" it. We can throw more hardware, people are the bottleneck nowadays

They already provide custom images for their Jetson modules, I think more NVIDIA distros are likely to happen one way or another.

Not to be contrarian, but b) could well be a full decade of work and numerous individual projects