How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference

TL;DR

The paper defines an infrastructure-aware framework to benchmark the environmental footprint of LLM inference at the prompt level across 30 models in real data-center deployments. It integrates API performance, hardware power characteristics, and infrastructure multipliers (PUE, WUE, CIF) with Gaussian copula modeling and cross-efficiency DEA to quantify energy, water, and carbon efficiency, delivering a dynamic dashboard for ongoing monitoring. Through GPT-4o and GPT-5 case studies, the authors illustrate per-query and adaptive-routing impacts, revealing substantial variability driven by deployment infrastructure and prompting a call for systemic sustainability measures. The work provides a standardized methodology for accountability and policy guidance in AI deployment, while acknowledging limitations and outlining avenues for future lifecycle analyses and multi-modal workloads.

Abstract

This paper introduces an infrastructure-aware benchmarking framework for quantifying the environmental footprint of LLM inference across 30 state-of-the-art models in commercial datacenters. The framework combines public API performance data with company-specific environmental multipliers and statistical inference of hardware configurations. We additionally utilize cross-efficiency Data Envelopment Analysis (DEA) to rank models by performance relative to environmental cost and provide a dynamically updated dashboard that visualizes model-level energy, water, and carbon metrics. Results show the most energy-intensive models exceed 29 Wh per long prompt, over 65 times the most efficient systems. Even a 0.42 Wh short query, when scaled to 700M queries/day, aggregates to annual electricity comparable to 35{,}000 U.S. homes, evaporative freshwater equal to the annual drinking needs of 1.2M people, and carbon emissions requiring a Chicago-sized forest to offset. These findings highlight a growing paradox: as AI becomes cheaper and faster, global adoption drives disproportionate resource consumption. Our methodology offers a standardized, empirically grounded basis for sustainability benchmarking and accountability in AI deployment.

Figures (8)

• Figure 1: (Left) Mean energy consumption of GPT-4o and GPT-4o mini across providers and GPU types, measured by output size. (Right) Distribution of TPS (averaged across output sizes)
• Figure 2: Visual overview of the AI sustainability dashboard.
• Figure 3: Energy consumption per model across three prompt sizes (Wh, log‐scale).
• Figure 4: Water consumption and carbon emissions per model.
• Figure 5: (Top Left) Per-query and daily energy consumption of GPT-4o. (Top Right) Estimated total annual energy usage of GPT-4o in 2025. (Bottom Left) The estimated 2025 annual water consumption of GPT-4o. (Bottom Right) The estimated 2025 annual carbon emissions of GPT-4o.