The cost accounting here is the most rigorous I've seen on this topic. Reading the baseline is most of the skill, and the DistServe/Mooncake/Dynamo comparison table alone is worth the read. One thing the piece prices in milliseconds and megabytes but not in compliance terms: that KV cache crossing the wire between prefill and decode pools is a recoverable representation of the prompt, not an opaque blob. On a regulated deployment where the two pools might not sit in the same trust zone, that handoff is a data-residency question as much as a latency one. Does kv_load_failure_policy=fail also need to guarantee the request doesn't silently retry across a boundary it isn't cleared to cross?