Signal Beyond Energy: When Sustainability Dimensions Reshape LLM Serving Decisions

Summary

LLM serving carries environmental costs across energy consumption, carbon emissions, water use, and biodiversity loss, but these dimensions have mostly been evaluated separately, leaving it unclear when they lead to different optimization choices. The authors present PRISM, a unified framework for characterizing and optimizing LLM serving jointly across all four dimensions. Their analysis finds a fundamental split: computing configuration determines energy consumption, while where and when serving is deployed determines its carbon, water, and biodiversity impacts. Under a fixed deployment and operational-only accounting, all four dimensions preserve the same energy-based ranking of configurations, but deployment rankings can diverge across dimensions, and embodied (manufacturing-related) impacts can break this consistency once they exceed a lifecycle crossover point. PRISM identifies these crossover conditions, quantifies the regret of optimizing for one dimension while ignoring others, and balances all four. In regional-routing experiments, PRISM cut the median worst-case regret by 50.2% compared with the strongest single-dimension baseline.