InFactPlanner framework lets operators simulate sustainable LLM data centers before building

As LLM inference shifts from one-time training to continuous serving, infrastructure decisions increasingly dictate energy use, carbon emissions, water consumption, and service quality. But operators rarely have a way to test deployment alternatives without building expensive infrastructure first. InFactPlanner, developed by Nicoletta Tsiopani, Moysis Symeonides, George Pallis, and Marios D. Dikaiakos (University of Cyprus), is a trace-driven what-if analysis framework that simulates single-site and geo-distributed AI data centers before a single server is deployed.

The framework integrates real query traces, hardware-model profiles, candidate site configurations, PUE/WUE parameters, renewable generation models, and time-varying grid carbon intensity to estimate power draw, energy consumption, carbon emissions, water usage, latency, and server utilization. It abstracts low-level serving effects into configurable profiles, making it easy to rapidly compare hardware choices, model placement, renewable integration, routing policies, and geographic siting. The authors validated their energy accounting pipeline by reproducing reference LLM inference estimates with less than 10% deviation, then demonstrated scalability across multiple data centers and server counts. Their scenario analyses reveal that sustainability-optimal deployments often differ from latency-optimal ones, and that the carbon value of any deployment hinges strongly on the local grid mix. In short, InFactPlanner gives operators a practical tool to balance environmental impact and performance before committing billions to new AI infrastructure.