How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference
Summary
Study quantifying environmental footprint of 30 LLM models during inference, focusing on energy, water, and carbon emissions disparities.
Key quotes
"O3 and DeepSeek-R1 consume over 33 Wh per long prompt, over 70x more than GPT-4.1 nano."
"Claude-3.7 Sonnet ranks highest in eco-efficiency."
"GPT-4o queries rival electricity use of 35,000 U.S. homes annually."
"Eco-efficiency depends on deployment infrastructure, not just model design."
"Unchecked AI scaling threatens sustainability without structural changes."
This paper introduces a novel framework for benchmarking the energy, water, and carbon footprint of large language models (LLMs) during inference. The study evaluates 30 state-of-the-art models, including proprietary and open-source options, by combining API performance data with region-specific environmental multipliers. Key findings emphasize that while some models like Claude-3.7 Sonnet achieve higher eco-efficiency, others like o3 and DeepSeek-R1 exhibit disproportionately high resource consumption. The research also underscores the significance of deployment infrastructure, such as hardware and cooling systems, in shaping environmental impact. Additionally, a case study on GPT-4o demonstrates that even daily interactions can collectively impose a substantial environmental burden, comparable to large-scale human activities.