You\u2019ve probably heard that using artificial intelligence consumes a lot of water. But how does it measure up with snacking on almonds?\u00a0<\/p>\n
OpenAI CEO Sam Altman made the comparison in a recent interview. OpenAI is the company that makes ChatGPT, the popular AI chat app.<\/p>\n
\u201cFor every 38,000 ChatGPT queries, that is the same amount of water that is used in the production of a single almond in California,\u201d he said in a Sept. 1 episode of the Sources Podcast<\/a>.\u00a0<\/p>\n Altman prefaced that by saying he isn\u2019t sure of the exact number, and it \u201cmight be wrong, but it\u2019s close.\u201d\u00a0<\/p>\n How close is it?\u00a0<\/p>\n Altman appears to be drawing from information about ChatGPT average water use that he published in 2025<\/a> on his blog, and a 2019 study measuring the water use of California almond farming.\u00a0<\/p>\n Altman said in that blog post that the average ChatGPT query uses 0.000085 gallons of water \u2014 about 0.32 milliliters, or 1\/15th of a teaspoon. Altman\u2019s post is the only source we found for that number. OpenAI does not publicize details about its overall water consumption, and the company did not respond to PolitiFact\u2019s request for more information.\u00a0\u00a0<\/p>\n A 2019 study<\/a> found that the process of growing a single California almond requires 3.2 gallons, or 12 liters, of water. That study counted more than just the direct water used to grow almonds. It also included rainfall and water quality impact in the measurement. If you compare those two numbers \u2014 dividing 12 liters (or 12,000 milliliters) for an almond by 0.32 milliliters for a ChatGPT query \u2014 you get 37,500, close to Altman\u2019s figure.\u00a0<\/p>\n But Altman\u2019s figure is misleading. It overstates the amount of water consumed by almond agriculture, and uses a relatively low ChatGPT water use estimate that is unconfirmed and disputed.<\/p>\n How AI and almonds stack up<\/p>\n Agriculture uses vastly more water than the AI industry and data centers overall. Nationally, farms consumed nearly 30 trillion gallons<\/a> of water for crop irrigation in 2020, the most recent data available. Meanwhile, data centers nationwide used about 228 billion gallons<\/a> of water in 2023. Many AI data centers<\/a> use water for cooling systems to keep computers from overheating, while other systems use very little water<\/a> but more electricity to keep the computers cool. Water is also used in some forms of electricity generation, so the power needed to run the data centers also often consumes water.<\/p>\n But the total water consumed by a single ChatGPT message is hard to measure and depends on many factors like the type of cooling and power used by a data center and the energy efficiency measures the company takes.<\/p>\n \u201cThere is very little public information about how today\u2019s ChatGPT systems operate and what resources they use,\u201d said Shaolei Ren, a University of California, Riverside, professor who researches AI\u2019s environmental footprint.<\/p>\n For almonds, the 12-liter-per-almond estimate includes rainwater and \u201cgrey water,\u201d which is a measure of what\u2019s needed to dilute pollutants such as fertilizer. About 6.3 liters of the total came from fresh water used for irrigation. A 2021 study<\/a> produced a similar fresh water estimate of 5.8 liters.\u00a0<\/p>\n That means a rough estimate of an almond\u2019s direct water consumption is closer to 6 liters \u2014 or 1.5 gallons \u2014 per almond.\u00a0<\/p>\n That already cuts Altman\u2019s comparison in half. The information available about water usage for an AI prompt further undercuts his point.\u00a0<\/p>\n AI\u2019s water demands vary too much to make a solid per-chat assessment. But based on some estimates, you\u2019d need anywhere from 1,000 to 10,000 conversational AI prompts to equal the 6 liters of fresh water used to grow one California almond.<\/p>\n Altman\u2019s 2025 blog post gave no explanation of what was included in the 0.000085 gallon measurement, and OpenAI hasn\u2019t published specific data.\u00a0<\/p>\n \u201cIn general, resource use can vary substantially depending on factors such as the data center location, AI model, prompt and output lengths, inference mode (e.g., reasoning versus non-reasoning), and the efficiency measures used to serve the model,\u201d Ren said.<\/p>\n A Sept. 1 study estimated the total water consumption, including electricity generation, for ChatGPT and Google\u2019s Gemini AI models and determined<\/a> that ChatGPT\u2019s 4o model used between 0.6 and 17 milliliters \u2014 1\/8th of a teaspoon and 3.5 teaspoons, respectively\u2014 per prompt, depending on the length of the input.<\/p>\n