A singular hardware stack for training AI models (GPUs) and a different hardware stack for inference (RDUs) makes the most sense for many AI use cases today.