Prefill is compute-bound, decode is memory-bound, and MoEs make everything more interesting: modelling LLM inference speed from compute, memory, and communication.