Beyond the Myths: Understanding the Real Energy and Water Footprint of AI Inference

Public discourse around artificial intelligence often portrays AI systems as extremely energy- intensive and water-intensive, sometimes comparing a single user query to hours of household electricity or liters of water consumption. While these concerns are legitimate, such comparisons frequently rely on misleading analogies and poorly defined assumptions. This article provides a grounded, order-of-magnitude analysis of the energy and water foot- print of large language model (LLM) inference. Starting from the computational fundamentals, tokens, floating-point operations (FLOPs), and modern GPU efficiency, we estimate the marginal electricity consumption of a range of typical interactions, from short requests to detailed analyzes, and place it in context with everyday energy usage. We then examine how electricity consumption translates into water use across different scopes, including direct data center cooling (Scope 1), electricity generation (Scope 2), and hardware manufacturing (Scope 3). By clearly separating marginal per-query costs from infrastructure-level averages, and by ac- counting for recent advances in model efficiency and hardware performance, this work aims to clarify where AI’s environmental impact is real, where it is overstated, and why accurate sys- tem boundaries matter. Rather than dismissing environmental concerns, the article argues that meaningful discussions about AI sustainability require correct physical reasoning, transparent as- sumptions, and appropriate comparisons.