Ask ten different sources how much water a single ChatGPT query uses, and you'll likely get ten different answers, ranging from a fraction of a milliliter to more than half a liter. That gap is at the center of a heated public debate that intensified after OpenAI CEO Sam Altman claimed 38,000 ChatGPT queries use about as much water as growing one California almond. We dug into the actual research behind the competing numbers to figure out who's closest to right.
The Claim That Started It
Altman's almond comparison, reported by Yahoo Finance, was meant to reassure the public that AI's water footprint is trivial. But when PolitiFact fact-checked the claim, it was rated "Mostly False" for combining a low estimate of ChatGPT's water use with a high estimate of almond water use, understating AI's footprint in the comparison.
Where the "Bottle of Water Per Prompt" Number Came From
The most widely cited high estimate traces back to a 2023 paper called "Making AI Less Thirsty" by researchers including Shaolei Ren at UC Riverside. In collaboration with the Washington Post in 2024, Ren's team estimated that generating a 100-word ChatGPT response could use approximately 519 milliliters of water, or roughly the volume of a standard water bottle. That figure went viral and was picked up by outlets including The Guardian and Forbes.
But there's an important catch: that estimate was based on GPT-4 running on 2024-era hardware. Ren himself has since told researchers the number is outdated. According to Forbes' 2026 reporting, Ren now puts the figure closer to 15 milliliters per GPT-4 prompt for more efficient, modern infrastructure, over 30 times lower than the original viral number.
Altman's Number vs. the Independent Researchers
Altman's own figure, published in his June 2025 blog post "The Gentle Singularity," puts the average ChatGPT query at about 0.000085 gallons, or roughly 0.32 milliliters of water. That's dramatically lower than even Ren's revised 15-milliliter estimate. The gap matters because, as one detailed breakdown from Promptslove points out, Altman never disclosed a methodology, scope, or whether his number includes the water used to generate the electricity powering the query, information that is standard in academic water-footprint studies.
That distinction, known as "direct" versus "indirect" water use, is at the heart of most of the confusion. Direct (or on-site) water is what's evaporated cooling the servers themselves. Indirect (or off-site) water is consumed generating the electricity that powers those servers, and according to the International Energy Agency, indirect use can account for roughly 60% of a data center's total water footprint. A number that only counts direct use, like Google's public per-query figures, will always look dramatically smaller than one that counts both.
Why Even Researchers Disagree With Each Other
Ren has publicly pushed back on comparisons that mix these two accounting methods. Speaking to The Verge, he said companies reporting only on-site figures against his combined on-site-plus-off-site figures are "hiding the critical information," arguing it sends the wrong message to the public. A separate benchmarking study, "How Hungry is AI?", found the spread across different models and workloads is enormous: less than 2 milliliters for efficient models, but over 150 milliliters for some heavier reasoning models processing longer responses.
The Bigger Picture: Total AI Water Demand
Individual query estimates aside, the scale question is where the real concern lies. A 2026 study covered by ScienceBlog projected that global AI infrastructure could require between 4.2 and 6.6 billion cubic meters of water withdrawal annually by 2027, a range the researchers compare to somewhere between four Denmarks and half of the United Kingdom's total annual water withdrawal. Separately, a Meta data center in Newton County, Georgia has been reported to use roughly 500,000 gallons of water per day, about 10% of the entire county's water consumption, according to the Lincoln Institute of Land Policy.
So, Is ChatGPT "Draining" Water Supplies?
The honest answer is that the picture is far messier than either side's talking points suggest. Sam Altman is likely right that the earliest viral claims (17 gallons per query) were wildly exaggerated and based on outdated 2023–2024 estimates. But his own comparison significantly understates AI's water footprint relative to independent academic research, and OpenAI has not published enough operational detail to verify his figures against Ren's or anyone else's. The clearest fact in the entire debate is the one nobody disputes: without standardized, transparent reporting from AI companies, the public will keep getting wildly different numbers depending on who's doing the math.
For more on where AI companies are directing their spending as demand for compute grows, see our report on why RAM prices are skyrocketing in 2026, and read our companion piece on how fast-moving tech infrastructure stories often get more complicated the deeper you dig.
Frequently Asked Questions
1. How much water does one ChatGPT query really use?
Estimates range enormously depending on methodology, from about 0.32 milliliters (Sam Altman's figure, direct use only) to roughly 15 milliliters (UC Riverside researcher Shaolei Ren's updated 2026 estimate, which includes indirect electricity-generation water). Older, now-outdated estimates put a single 100-word response at up to 519 milliliters.
2. Why do the numbers vary so much?
Mainly because "water use" can mean direct (on-site cooling) water only, or direct plus indirect (electricity-generation) water. Indirect use can account for roughly 60% of total water footprint according to the International Energy Agency, so leaving it out makes any figure look much smaller.
3. Is Sam Altman's almond comparison accurate?
PolitiFact rated it "Mostly False." It pairs a low estimate of ChatGPT's water use with a high estimate of almond water use, and independent estimates suggest the real ratio is closer to 1,000–10,000 ChatGPT queries per almond, not 38,000.
4. Has the original "bottle of water per prompt" claim been debunked?
It's outdated rather than wrong outright. The 519 mL figure was based on GPT-4 and 2024-era infrastructure; the same researcher who produced it, Shaolei Ren, now estimates a comparable modern query at roughly 15 mL, reflecting real efficiency gains in newer AI hardware.
5. Is AI's total water demand actually a concern?
At scale, yes, according to multiple independent studies. Global AI infrastructure water withdrawal is projected to reach billions of cubic meters annually by 2027, and individual data centers have been shown to consume a meaningful share of local community water supplies in some regions.