And although this standard obviously wants to stick within the existing UUID footprint, if you were say doing some IoT software that would run on billions of nodes simultaneously, just add another 32/64/whatever bits of random data and deal with the minor annoyance of longer ids and lack of UUID RFC compatibility. But even then, you can truncate and reformat these these to v4 UUIDs trivially without meaningfully impacting the collision resistance, for unsorted external ids in systems that need the compatability.
The actual big risk with this sort of scheme is vm initialization. You need to be sure the CSPRNG you're using is initialized, which can be slightly tricky in cloud environments with configuration/control layer stuff that's racy. Mess this up and two nodes hydrated from the same snapshot may overlap in sequence as they start generating ids, and of course the time component cannot be trusted to save you in this instance.
Even if one was not generating millions of UUIDs per second on average, the risk of spiky temporal distributions when generating UUIDs would still need to be considered.