When most folks are using CUDA indirectly via library abstractions, swapping it out eventually becomes easy and their moat fades away. CUDA alternatives are in

When most folks are using CUDA indirectly via library abstractions, swapping it out eventually becomes easy and their moat fades away. CUDA alternatives are innovating pretty fast, and the ever present threat of TPUs and things like Apple's Neural Engine aren't helping them. They should've been building better partnerships and giving more margin away for integration and OEM instead of being greedy. But that wouldn't have juiced their short term profits.