AI/ML is addicted to CUDA. AMD and even Intel have viable alternatives in terms of raw compute performance, but they'll never natively support CUDA, and that's what most of the software tooling is focused right now.