“There is no "CUDA lock-in" for AI inference workloads, as it is simple for users to switch between different hardware providers.”