If you've spent any time on social media, you know the internet is abuzz right now with the claims of Open AI CEO Sam Altman. On the Sources Podcast with Alex Heath, Altman claimed that it takes 38,000 ChatGPT inquiries to use as much water as it takes to produce one almond in California. If you use AI at home or manage technology for a business, the number sounds reassuring. But does that claim stand up to a fact check?
Admittedly, Altman said he was recalling the calculation from memory and that it could be wrong.
The problem? The comparison cannot be verified. Water estimates change sharply based on what researchers count, where the servers operate and a host of other variables.
Key Takeaways
- Sam Altman's claim that 38,000 ChatGPT inquiries use as much water as producing one California almond is an unsupported, highly uncertain estimate.
- Using the commonly cited estimate of 3.56 liters per almond, the claim implies about 0.094 milliliters of water per ChatGPT inquiry.
- There is no universal water-per-query figure because usage varies by model, prompt length, response length, data-center location, cooling system, and electricity source.
- Different accounting methods produce very different comparisons, including estimates of about 11,125 requests per almond and ranges from roughly 85 to 5,500 requests.
- AI has real infrastructure and environmental costs, but reliable comparisons require transparent sources, consistent boundaries, and reproducible calculations.
Fact check: 38,000 ChatGPT Inquiries per California Almond
Verdict: treat the claim as an unsupported, highly uncertain estimate.
Sam Altman's comparison claimed that growing one almond required about as much water as 38,000 ChatGPT queries. If you start with an almond estimate of 3.56 liters, or 3,560 milliliters, the implied amount is tiny:
3,560 milliliters divided by 38,000 inquiries equals about 0.094 milliliters per inquiry.
That figure may be possible under a narrow set of assumptions. It's an estimate of facility-level water consumption under a particular accounting scope, not a universal result. Still, you can't independently confirm it without knowing the model, data-center locations, cooling systems, local weather, electricity sources, and whether the calculation counts only water used at the facility.
What Altman said, and what his wording leaves out
Altman didn't present a study or data to back up his claim. He said the number came from memory, acknowledged it might be wrong, and said he believed it was close. That's a public claim, not a documented measurement.
A reliable comparison needs clear boundaries. Does it count water directly consumed by cooling equipment? Does it include water used by power plants to produce electricity? Does it include construction, chip manufacturing, or AI model training? Without those answers, a ratio of 38,000 ChatGPT inquiries is a vague claim at best.
The almond number isn't fixed either
The commonly repeated California almond estimate is about 3.56 liters per almond. But even that number isn't a hard and fast rule. Other sources estimate 3.2 gallons of water, or about 12 liters, per almond. Agriculture always includes many variables, including weather, drought conditions, soil conditions and so much more.
Water footprint studies can include irrigation, rainfall, processing, and regional water stress in different ways. Change the almond figure, and the number of AI requests in the comparison changes with it.
How much water does one ChatGPT query really use?
The simple truth is, there is no universal water-per-query number. A two-word prompt and a long research request don't require the same computing work. ChatGPT queries can also vary with token consumption, model version, response length, and traffic levels.
Location matters just as much. Data centers in hot, dry areas may use different cooling systems than those in cooler climates. Electricity generation adds another layer because power plants may withdraw or consume water.
What the leading academic estimate measured
The best-known research in this area comes from UC Riverside and the University of Texas at Arlington. Its authors include Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren. Their paper, Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models, first appeared as a 2023 preprint and later appeared in Communications of the ACM in 2025.
The researchers estimated that a short large-language-model conversation of roughly 10 to 50 responses could consume about 500 milliliters of water under modeled conditions. That estimate describes modeled water consumption, not every current ChatGPT exchange. It depends on modeled infrastructure, locations, cooling assumptions, and electricity conditions.
Why different methods produce different answers
Some online discussions cite ranges from roughly 85 to 5,500 requests per almond under different cooling assumptions, including evaporative cooling. A range without a named model, time period, and method does not prove much. It does show why you should reject claims that present one exact number as universal.
Public disclosures from data centers remain too limited for outside reviewers to recreate the 38,000 ChatGPT inquiries calculation. CalMatters' review of the claim reached the same core conclusion: the lack of facility-level data reflects limited transparency.
Why the almond comparison can mislead you
An almond is a physical crop with a seasonal, regional agricultural water footprint. A ChatGPT request is a digital service whose resource use changes with data centers, infrastructure, and demand. Combining them is tricky business to say the least.
You also need to watch the water terms. They aren't interchangeable.
Water withdrawal is water taken from a source. Water consumption is water that isn't returned to the same watershed, often because it evaporates. A claim can sound lower or higher depending on which measure it uses.
Potable water means drinkable water. Recycled water has been treated for another use, such as industrial cooling. Closed-loop cooling systems are one design that may reduce or change on-site water use, but they don't eliminate every environmental cost. Indirect water use comes from producing the electricity that runs servers. The power grid's generation sources affect that water use. A company could report low on-site water use while omitting electricity-related water.
The missing evidence you should look for
Before accepting any AI-water comparison, look for the details that would allow another researcher to repeat it:
- The exact almond study and the definition of its water footprint.
- The AI model, the prompt length, and the average response length.
- The data centers included, their climate conditions, and their cooling technology.
- The electricity mix and the time period covered.
- Whether the estimate covers only inference, or also model training and hardware production.
Sparse disclosures create the same consumer problem you see with a vague home-repair estimate. You can't compare claims if the underlying scope is hidden. For local context, review how data centers can affect utility bills, including possible pressure on municipal water systems and power supplies during peak water demand.
What this claim does and doesn't prove
There is plenty of uncertainly around AI's environmental impact and resource cost right now. Data centers use electricity, land, equipment, and sometimes significant amounts of water, especially at large scale.
A simplified almond ratio cannot measure the AI industry's total impact.
How ChatGPT compares with other AI platforms on water use
You can't fairly rank ChatGPT, Google Gemini, Microsoft Copilot, Claude, or open-source models by water use today. No major provider has supplied enough comparable, product-level water data from its data centers for an independent apples-to-apples scorecard.
Providers can handle billions of queries through hyperscale data centers, but scale alone doesn't show that every request has the same footprint. A platform may use different models and facilities for different requests. Your own use also affects computing demand. A short rewrite, a large spreadsheet analysis, and image generation don't share the same resource profile.
ChatGPT, Gemini, Copilot, and Claude have different strengths
ChatGPT, from OpenAI, is widely used for general writing, coding help, analysis, and brainstorming. It can't provide a verified, facility-level water audit for each response.
Gemini works closely with Google services and supports multimodal tasks involving text and images. Its model choice, task requirements, and hosting environment can affect resource use. Public reporting doesn't provide a consistent water metric for each Gemini request.
Microsoft Copilot fits into Microsoft workplace products and can help with documents, meetings, spreadsheets, and coding workflows. Its footprint depends on the model used, task length, cloud infrastructure, cooling system, and electricity supply.
Claude is often chosen for long documents, writing, and careful analysis. It can't disclose proprietary cooling information or provide a complete water total for a single conversation.
Open-source models give organizations more control over model choice and hosting. They don't automatically use less water. Server location, cooling design, model size, and electricity supply still shape much of the impact.
Why no chatbot can give you a complete water audit
A chatbot can summarize research or calculate a ratio. It usually can't access real-time water use across proprietary data centers, detailed cooling records, or a consistent lifecycle method shared by all providers.
For an important purchasing, policy, or business decision, an AI answer leaves out a lot of important variables. No matter how sophisticated AI has become, it is no substitute for doing your own homework. That is exactly why TrustDALE vets every one of its Certified Partners with our rigorous 7 point investigative process.
A practical way to fact-check viral AI claims
You don't need specialized equipment to spot a weak environmental claim. You need a clear chain from the quote to the calculation.
- Find the original statement, speaker, date, and full context. Clips often cut out qualifications.
- Ask for the source calculation. A valid claim identifies its study, data period, and accounting method.
- Check the units yourself. For ChatGPT queries, convert gallons, liters, and milliliters before accepting a ratio.
- Compare at least two credible estimates, then identify why they differ.
- Check the boundaries closely. Verify which data centers and facility locations are included. Confirm whether the number covers cooling water, electricity-related water, model training, or only one type of request.
- Look for independent university, government, or utility evidence before repeating the figure.
Avoid social posts and infographics that repeat a number without tracing it to primary research. When a claim relies on an exact ratio but hides the math, the precision is marketing, not proof. Credible fact-checking requires transparency about sources, assumptions, and reproducible math.
Frequently Asked Questions
Is it true that 38,000 ChatGPT inquiries use as much water as one almond?
The claim has not been independently verified. It appears to be a rough estimate based on undisclosed assumptions about water use, data centers, cooling, electricity, and the almond's water footprint.
How much water does one ChatGPT inquiry use?
There is no universal amount. Water use can vary with the model, prompt and response length, data-center location, cooling technology, and whether indirect water use from electricity is included.
How much water is used to produce one California almond?
A commonly cited estimate is about 3.56 liters per almond, but other figures are also reported. The totals differ because studies may count irrigation, rainfall, processing, and regional water stress in different ways.
Why do estimates of AI water use differ so much?
Researchers use different accounting boundaries and model different infrastructure, climates, cooling systems, and electricity sources. Some estimates count only on-site cooling water, while others include water used to generate electricity or support the broader AI system.
Does the uncertainty mean AI has no environmental impact?
No. Data centers use electricity, land, equipment, and sometimes substantial amounts of water, particularly at large scale. The uncertainty means that a precise comparison such as 38,000 inquiries per almond should not be treated as a proven universal measurement.
The Evidence Matters More Than the Almond Analogy
The 38,000 ChatGPT inquiries per California almond comparison remains an unverified estimate, not a proven fact. Its implied figure is 0.094 milliliters per request, but the facts to support that claim remain unavailable.
Other reported estimates produce very different results. AI water use is a real infrastructure issue, and its environmental impact deserves scrutiny. Responsible fact-checking requires transparent methods and comparable boundaries. Disclosures from OpenAI and other companies would be more useful than a catchy comparison.