“The process of AI inference is increasingly being disaggregated into two distinct steps: prefill and decode.”