While artificial intelligence is frequently criticized for its heavy water consumption, OpenAI CEO Sam Altman recently offered an unexpected point of comparison: the humble California almond.
During a September 1 appearance on the Sources Podcast, Altman claimed, “For every 38,000 ChatGPT queries, that is the same amount of water that is used in the production of a single almond in California.” He noted that while he wasn’t certain of the exact figures, the comparison was likely close.
Where did those numbers come from, and do they hold up to scrutiny?
Altman’s calculation appears to stem from a water-usage figure he shared in a 2025 blog post combined with a 2019 study on almond farming. In his post, Altman stated that a single ChatGPT query consumes just 0.000085 gallons—roughly 0.32 milliliters—of water. Because OpenAI keeps its overall resource consumption private and did not respond to inquiries for clarification, that figure remains unverified.
Meanwhile, the 2019 study estimated that producing one California almond requires 12 liters (or about 3.2 gallons) of water. Crucially, this calculation went beyond direct irrigation, factoring in natural rainfall and the water required to dilute agricultural pollutants. Dividing that 12,000-milliliter total by the 0.32 milliliters attributed to a ChatGPT query yields roughly 37,500—mirroring Altman’s claim.
However, experts point out that this comparison is flawed because it relies on an unverified, exceptionally low estimate for AI water use while overstating the freshwater footprint of almonds.
Comparing AI and Agriculture
On a macro scale, agriculture dwarfs the tech industry in water usage. U.S. farms consumed nearly 30 trillion gallons of water for crop irrigation in 2020 alone, compared to the roughly 228 billion gallons used by the nation’s data centers in 2023. Data centers rely on water for cooling to prevent overheating, though alternative cooling methods use less water while demanding more electricity—which itself often consumes water during generation.
Calculating the exact water footprint of an individual AI prompt is difficult, as it fluctuates based on data center efficiency, local climate, and the specific cooling infrastructure used.
“There is very little public information about how today’s ChatGPT systems operate and what resources they use,” notes Shaolei Ren, a researcher studying AI’s environmental footprint at the University of California, Riverside.
When looking closer at almonds, the 12-liter metric includes rainwater and “grey water” (water needed to dilute fertilizer runoff). The actual fresh water used directly for irrigation is closer to 6 liters (about 1.5 gallons) per almond—a finding supported by a separate 2021 study.
Adjusting the almond metric to reflect only direct freshwater usage cuts Altman’s comparison in half. Furthermore, more recent independent assessments of AI water consumption suggest that the prompt-to-almond ratio is significantly smaller than Altman suggested.
Depending on prompt length and the model used, independent research indicates that a standard conversational AI query actually consumes anywhere from 0.6 to 17 milliliters of water. Tools tracking AI energy impact, such as EcoLogits, estimate that drafting an email using newer models requires over 6 milliliters of water. Additionally, complex AI tasks—such as coding or multi-step reasoning—can demand more than ten times the energy of a basic chat, driving up water usage accordingly.
Because AI resource demands vary widely and OpenAI has not released comprehensive usage data, establishing a single baseline is difficult. However, current independent estimates suggest that anywhere from 1,000 to 10,000 conversational AI prompts—rather than 38,000—equal the freshwater footprint required to grow a single California almond.
Conclusion
While Sam Altman’s comparison makes for an arresting headline, it relies on an unverified, low estimate for AI water consumption paired with an inflated total for almond production. Factoring in actual direct freshwater usage for crops and updated estimates for AI queries reveals a much narrower gap, making the original claim misleading.