{"type":"post","id":"only-accepted-tokens-count-an-energy-accounting-of-af8f7c","slug":"only-accepted-tokens-count-an-energy-accounting-of-af8f7c","title":"Only Accepted Tokens Count: An Energy Accounting of Speculative Decoding Under Latency Targets","description":"A new paper measures speculative decoding as an energy and latency trade-off under production-like serving conditions, rather than treating it as a universal efficiency upgrade.","retrieval_nugget":"A new paper measures speculative decoding as an energy and latency trade-off under production-like serving conditions, rather than treating it as a universal efficiency upgrade. For teams operating LLM inference, the practical result is clear: batch size, draft length, workload, and accepted-token rate can reverse the expected savings.","published_at":"2026-10-31","updated_at":"2026-10-31","record_date":"2026-10-31","date_kind":"published_at","topics":["agent-economics","model-routing"],"entities":["SGLang","NVIDIA H100","NVIDIA H200","CUDA Graphs","EAGLE3","DFlash","DSpark"],"related_patterns":["agent-economics-moves-to-completed-work","harness-architecture-outlives-model-choice"],"novelty_label":"new_fact","what_changed":"New SGLang measurements quantify energy savings up to 77% at batch 1, but show savings shrink with batching and can turn into energy losses with long drafts or low acceptance.","prior_refs":["memory:ea6e7e73-49a1-80a4-a848-2a55d27b2300","memory:f29a971e-c84c-8716-a7d0-ddab4d1c468f","memory:b0babd5e-ff30-88b6-abe3-b5832dc7e4e2"],"source_url":"https://arxiv.org/abs/2610.40287","source_urls":["https://arxiv.org/abs/2610.40287"],"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:only-accepted-tokens-count-an-energy-accounting-of-af8f7c","status":"published","visuals":[],"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"routes":{"html":"https://newruntime.com/posts/only-accepted-tokens-count-an-energy-accounting-of-af8f7c/","markdown":"https://newruntime.com/posts/only-accepted-tokens-count-an-energy-accounting-of-af8f7c.md","json":"https://newruntime.com/posts/only-accepted-tokens-count-an-energy-accounting-of-af8f7c.json"},"next_reads":[{"type":"topic","path":"/topics/agent-economics/","url":"https://newruntime.com/topics/agent-economics/","title":"Agent Economics","reason":"Explore the agent-economics topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/agent-economics.md","json_url":"https://newruntime.com/topics/agent-economics.json"},{"type":"topic","path":"/topics/model-routing/","url":"https://newruntime.com/topics/model-routing/","title":"Model Routing","reason":"Explore the model-routing topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/model-routing.md","json_url":"https://newruntime.com/topics/model-routing.json"},{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/source-ledger.json","url":"https://newruntime.com/source-ledger.json","title":"Source Ledger","reason":"Inspect the public source evidence graph.","media_type":"application/json"}]}