Abdeltawab Hendawi

Photo of Abdeltawab Hendawi
Abdeltawab Hendawi

Abstract Title

How Hungry is AI?

Benchmarking Energy, Water, and Carbon Footprint of LLM Inference
The rapid emergence of artificial intelligence has driven a sharp rise in global datacenter energy consumption, yet its true impact remains largely invisible due to limited public transparency. While most sustainability discussions have centered on the one-time cost of training open-source models, inference now accounts for nearly 90% of the total lifecycle energy use of large language models (LLMs), occurring billions of times each day. This study introduces the first infrastructure-aware framework to quantify the energy footprint of LLM inference by combining public API performance data with datacenter-specific environmental multipliers and probabilistic hardware estimation. Applying this framework to over 30 proprietary and open-source models reveals substantial discrepancies in energy footprint, with the most intensive systems consuming over 65× more electricity per query than the most efficient. When scaling a simple query that consumes 0.42 Wh to real-world usage, we estimate that inference alone requires annual electricity comparable to 35,000 U.S. homes. These results underscore the need for transparent energy reporting and standardized measurement frameworks to guide the sustainable development, deployment, and use of AI. 

Biography

Abdeltawab Hendawi is an Associate Professor in Computer Science and Data Science at the University of Rhode Island (URI). He is the Co-director of the AI-Lab at URI.  He received his PhD in Computer Science from the University of Minnesota (UMN). Then, he was awarded a four-year Post Doctoral Research fellowship with Prof. John Stankovic in Computer Science at the University of Virginia (UVA). His research interests are centered on AI and Big Data with a focus on smart cities and smart health related applications. His work has been recognized by several awards at top ACM and IEEE conferences. His research is sponsored by grants from NSF, TIDC, USDA, P2P, and MindImmune Inc.

Contact

hendawi@uri.edu