How Big Is a Fish? Making Sense of GenAI’s Energy Use

This article examines the environmental impact of generative AI by analyzing recent research papers from the UN, Google, Mistral, UC Riverside, and Epoch AI. It compares energy consumption metrics for AI queries, such as ChatGPT and Gemini, highlighting challenges in interpreting median values and system boundaries. The author questions the transparency of reported figures, noting differences in model sizes, hardware (TPUs vs. GPUs), and inclusion of overhead. The piece also discusses improvements in energy efficiency over time and raises concerns about scaling energy use for complex tasks. Overall, it provides a critical perspective on understanding GenAI's true energy footprint.