CUDA

CUDA is the toolkit Nvidia built so that programmers could use graphics cards for ordinary maths rather than only for drawing pixels. Nearly every AI framework was written against it, and nearly every published training recipe was developed on it. The chips are fast, but CUDA is the reason the ecosystem sits where it does.

That is why competitors cannot win on hardware specifications alone. AMD’s answer, ROCm, has to be good enough that researchers are willing to rewrite and re-debug work that already runs on Nvidia, which is a much higher bar than being cheaper.