“For AI inference workloads, customers do not care about NVIDIA's CUDA, as their primary requirement is a simple API.”