The 1/W Law: Routing Topology Dominates LLM Inference Energy Efficiency More Than Hardware Generation
arXiv 2603.17280·high signal
This paper (arXiv:2603.17280, March 18) derives the '1/W law': tokens per watt halves every time the effective context window doubles, meaning the routing topology — which determines which GPU services which context — is a more powerful energy lever than purchasing newer GPU generations. For practitioners running large-context agentic deployments, this implies architectural routing decisions outweigh hardware refresh cycles in operational cost. Directly actionable for infrastructure planning in multi-agent systems.