It's so weird that Cuda has latched on to the ML world so tightly. It's a field where literally everything else is all about open-sourcing and giving away ivory tower knowledge, and then there's so much BS that Nvidia makes you go through - I had to wait a few days to get "approved" to download CudNN, where else is that acceptable? AMD releases high-quality open-source drivers, has better performance/TDP, usually better performance/$, and isn't having security issues every month. It'd be great to see them start to compete with ML libraries as well, they're such a better citizen.
So no Nvidia cards is a feature for me.