InFactPlanner: A Framework for Sustainable Geo-Distributed LLM Data Centers

A recent paper on arXiv (2608.12915) presents InFactPlanner, a decision-support framework driven by trace data aimed at conducting what-if analyses for the sustainable deployment of AI data centers focused on LLM inference. This framework responds to evolving sustainability challenges, emphasizing the transition from one-off training to ongoing service, where infrastructure choices significantly affect energy consumption, carbon output, water usage, and service quality. InFactPlanner allows operators to evaluate deployment options prior to extensive infrastructure investments, thus circumventing the need for expensive and time-consuming direct measurements. It integrates query traces, hardware-model profiles, potential site setups, PUE/WUE metrics, renewable energy models, and fluctuating grid carbon intensity to forecast power, energy, carbon emissions, water consumption, latency, and server efficiency. The paper is categorized as a cross-type announcement on arXiv and aims to facilitate sustainable decision-making in LLM data center deployments, addressing the increasing environmental impact of AI inference.