Almonds vs. AI: Debunking Water Guilt Myths

OpenAI CEO Sam Altman has sought to dispel the notion that everyday AI use, such as drafting emails with ChatGPT, is an environmental catastrophe. He stated on the Sources podcast that modern data centers use as much water for 38,000 queries as is needed to produce a single California almond, which requires approximately 1.1 gallons of water.

This assertion contrasts with earlier studies, such as one from UC Riverside, which estimated that a brief LLM conversation could use up to 500 milliliters of water. The discrepancy arises from differing methodologies, with Altman's figures focusing on direct cooling in modern facilities, while other analyses include water used for electricity generation and older hardware.

Critics argue that Altman's comparison, while perhaps accurate for single, lightweight prompts on new systems, overlooks the aggregate scale of AI usage, which processes billions of prompts daily. Furthermore, AI data centers often compete for water resources in arid regions, and more complex AI tasks consume considerably more energy, leading to indirect water usage.

Modern data centers are shifting from evaporative cooling to closed-loop liquid cooling, significantly reducing direct water loss. However, replacing water cooling with energy-intensive chillers increases electricity demand, which carries its own environmental footprint. The debate highlights the complexities of measuring AI's true water impact.