As the artificial intelligence boom drives a massive expansion of data centers across the globe, environmental concerns have shifted from electricity consumption to a new focal point: water. However, OpenAI CEO Sam Altman is pushing back against the narrative that generative AI is drinking the planet dry.
In a September 2026 interview on the Sources podcast, Altman addressed the mounting public backlash regarding AI water consumption. Responding to viral claims that a single ChatGPT query is equivalent to running a shower for six hours, Altman dismissed the notion as a “robust meme” that fails to hold up to scrutiny. To contextualize the scale of AI’s environmental footprint, he offered a striking comparison: producing a single California almond requires the same amount of water as processing 38,000 ChatGPT queries.
As local communities and environmental groups actively protest the construction of new hyperscale data centers in drought-prone areas, the debate over AI’s resource consumption has never been more contentious. Whether you are a daily ChatGPT user wondering about your digital carbon footprint or an industry observer tracking data center infrastructure, separating the facts from the viral memes is essential. We are breaking down the math behind the “almond” comparison, the evolution of server cooling technologies, and the reality of AI’s water usage in 2026.
The Math Behind the Almond Comparison
To understand Altman’s claim, we have to look at the exact numbers comparing agricultural water usage to modern AI inference calculations.
California grows approximately 80% of the world’s almonds, a notoriously water-intensive crop. According to a 2019 study by researchers affiliated with the US Geological Survey, it takes an average of 3.56 liters (nearly 1 gallon) of water to produce a single almond.
If we apply Altman’s ratio—38,000 queries per almond—that implies a single ChatGPT query consumes roughly 0.09 milliliters of water.
This figure is even lower than previous estimates provided by the industry. Earlier in the year, Altman stated that a single ChatGPT query uses approximately 0.32 milliliters of water. For comparison, Google recently reported that a median text prompt to its Gemini system uses just 0.26 milliliters of water—roughly the volume of five drops.
When viewed through this lens, agricultural water consumption dwarfs AI processing. As Altman noted on the podcast, “People that are scarfing down 12 almonds at a time don’t feel like they’re doing something horrible from a water perspective”.
Debunking the “Bottle of Water” Myth
If modern queries only use a fraction of a milliliter, where did the viral narrative that ChatGPT “drinks a bottle of water” per conversation come from?
The discrepancy lies in the rapid evolution of AI models and the difference between training a model and running it (inference).
The infamous “water bottle” statistic originated from a highly publicized independent study which estimated that older systems, like the GPT-3 model, consumed roughly 500 milliliters of water (a standard single-serving plastic water bottle) for every 10 to 50 queries.
However, that older calculation factored in a perfect storm of inefficiencies that have largely been engineered out of modern systems:
- Older Data Centers: It assumed the use of outdated, highly consumptive evaporative cooling towers.
- Power Grid Water: The 500 mL figure included the “indirect” water used by local power plants to generate the electricity required to run the servers.
- Less Efficient Hardware: Earlier AI models required significantly more compute power (watt-hours) per generated token than the highly optimized GPT-4o and GPT-5 models running today.
The Evolution of Data Center Cooling
Altman argues that the public perception of data centers is stuck in the past, heavily focused on outdated cooling mechanisms that are no longer the industry standard.
Servers generate massive amounts of heat, and for years, data centers relied on evaporative cooling—giant misters that sprayed water over hot pipes to carry heat away via evaporation. This method literally removes water from the local aquifer and releases it into the atmosphere.
According to Altman, modern hyperscale facilities have largely abandoned this practice. Many state-of-the-art data centers now utilize closed-loop liquid cooling systems or advanced air cooling, which continuously recycle a fixed amount of fluid without constantly drawing fresh water from the local municipality.
Altman claims that a modern, large-scale data center now uses an equivalent amount of water as a standard commercial office building, primarily accounting for routine human use like sinks and toilets.
Use the interactive calculator below to see how AI query volume stacks up against everyday water usage based on both older and modern efficiency metrics:
Community Pushback and the Real Problem
Despite the favorable math per query, the sheer scale of the AI rollout means the total environmental impact cannot be entirely dismissed.
With OpenAI processing billions of prompts every single day, even fractions of a milliliter compound into millions of liters. The primary issue is not the global water supply, but localized water stress.
A significant portion of new US data centers are being constructed in areas with cheap land and abundant solar power—often in desert regions like Texas, Utah, and Nevada, or drought-prone zones in California and New Mexico.
- In one notable instance in Fayette County, Georgia, a data center consumed 29 million gallons of water over 15 months.
- A recent Gallup survey revealed that 71% of Americans oppose the construction of a data center near their home, with local resource and water depletion cited as a primary concern.
While an individual query may be a drop in the bucket compared to eating a handful of almonds, a massive server farm drawing millions of gallons of water from a stressed local aquifer remains a highly valid concern for the communities living around them.
Frequently Asked Questions (FAQ)
Understanding ChatGPT’s Water Footprint
How much water does one ChatGPT query use?
Recent estimates from OpenAI and Google suggest that a standard AI query (prompt and response) uses between 0.26 and 0.32 milliliters of water on modern infrastructure. This is a massive improvement from older estimates, which placed the usage around 10 to 50 milliliters per query on legacy hardware.
Did Sam Altman say ChatGPT uses the same water as one almond?
Yes. In a September 2026 podcast interview, Sam Altman stated that 38,000 ChatGPT queries use the same amount of water required to produce a single almond in California. (An almond requires roughly 3.56 liters of water to grow).
Why do data centers need water?
Data centers house thousands of high-performance servers that generate immense heat. Water is used in cooling systems (like cooling towers or liquid cooling loops) to absorb and dissipate this heat, preventing the hardware from overheating and failing.
Are AI data centers draining local water supplies?
It depends entirely on the location and the cooling technology used. While modern data centers are highly efficient, facilities built in drought-prone or desert areas can place significant strain on local aquifers, sparking pushback from local residents.
Sam Altman’s “almond” comparison serves as a highly effective reality check against the hyperbolic memes surrounding AI’s environmental footprint. The data supports his core argument: when utilizing modern, closed-loop cooling infrastructure, the water required to process an individual ChatGPT query is microscopic, especially when compared to traditional agriculture. However, dismissing the issue entirely ignores the localized realities of the AI boom. As tech giants continue to build massive server farms in water-stressed regions, the industry will have to prove that its macro-level efficiencies translate into sustainable practices for the local communities hosting their hardware.
