CUDA is much more accessible/documented/available than a lot of comparable products. FPGAs were even more closed, more expensive and had much worse documentation. Things like compute clusters were call-for-pricing, prepare to spend as much as a small house.

On the other hand, CUDA is closed enough the chips that run it aren't a commodity. If you want to download that existing ML project and run it on your AMD card - someone will have to do the leg work to port it.

That means they've been able to invest quite a lot of $$$ into CUDA, knowing the spending gets them a competitive advantage.