Every time you send a message to ChatGPT, a data center somewhere processes your request. That takes electricity. And in most data centers, cooling the servers that run that electricity requires water.
The question of exactly how much water has produced a wide range of claims some alarming, some dismissive. The honest answer sits somewhere in between, and it depends entirely on what you are measuring.
The Quick Answer
Here is what the main sources say about ChatGPT’s water use per query:
| Source | Estimate Per Query | What It Measures |
| Sam Altman / OpenAI (June 2025) | ~0.3 ml (1/15 of a teaspoon) | Direct on-site cooling only |
| UC Riverside Study (Li et al., 2023) | ~10–25 ml | On-site cooling + some indirect use |
| Washington Post / UC Riverside analysis | ~519 ml for a 100-word email | Full lifecycle including electricity generation |
| NIAIS / European estimates | ~5–9 ml | Hybrid cooling systems |
The range is wide because each estimate draws the boundary differently. None of them are wrong. They are just measuring different things.
Training vs. Inference: Why the Distinction Matters
There are two separate phases where AI models consume water, and most people conflate them.
Training is the one-time process of building the model. It is intensive and happens once per model version. According to the UC Riverside study “Making AI Less Thirsty” by Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren, training GPT-3 in Microsoft’s US data centers directly consumed approximately 700,000 liters of freshwater enough to manufacture around 320 Tesla electric vehicles.
Inference is what happens every time you use ChatGPT. Each query triggers a relatively small compute task. This is where the per-query figures come from.
Training water costs are a one-time sunk cost per model. Inference costs are cumulative and grow with every user and every message sent.
What OpenAI Says
In June 2025, OpenAI CEO Sam Altman published a blog post called The Gentle Singularity his first public statement on ChatGPT’s resource use. He stated that the average ChatGPT query uses approximately 0.000085 gallons of water, which equals roughly 0.32 milliliters, or about one-fifteenth of a teaspoon.
That is a small number. But there are important caveats Altman did not address:
- The figure only accounts for direct, on-site water cooling at OpenAI’s data centers.
- It does not include indirect water use the water consumed by power plants generating the electricity those data centers run on.
- OpenAI did not publish the methodology behind the figure, so independent researchers treat it as a plausible floor, not a complete accounting.
Independent analysis by Data Center Dynamics noted that what qualifies as an “average query” was not defined, and whether more intensive tasks like deep research or image generation were excluded is unknown.
What Independent Research Says
The most widely cited academic work is the 2023 paper Making AI Less Thirsty from researchers at the University of California, Riverside and the University of Texas at Arlington. Their methodology accounts for both direct cooling water at data centers and the indirect water embedded in electricity generation.
Their inference estimate: roughly 500 ml (one standard water bottle) per 20–50 queries for ChatGPT at the time of publication, when the underlying model was still GPT-3.
That works out to approximately 10–25 ml per individual query depending on query length and data center location.
A subsequent analysis published via the Washington Post found that generating a single 100-word email using GPT-4 consumed approximately 519 ml when the full lifecycle was included electricity generation, cooling, and all related infrastructure. That figure is from the same UC Riverside research group and uses the same full-lifecycle methodology.
It is worth noting that these figures were derived for earlier model versions. Newer hardware and cooling infrastructure have improved efficiency, which is why Altman’s 2025 figure is significantly lower.
Why the Numbers Vary So Much
This is the part most articles skip, and it is the most important part.
- Scope of measurement Direct water use covers only what evaporates at the data center. Full-lifecycle water use adds the water consumed by power plants to generate the electricity feeding the data center. The indirect figure is typically much larger than the direct figure. The Lawrence Berkeley National Laboratory’s 2024 report estimated that US data centers consumed approximately 17.4 billion gallons of water directly through cooling in 2023 and an additional 211 billion gallons indirectly through their electricity supply.
- Location Data centers in hot, arid regions use more evaporative cooling and therefore more water. The same model running on servers in a cooler climate consumes less. The UC Riverside paper noted that training costs would have been three times higher in Microsoft’s Asian data centers than in the US.
- Model version GPT-3 and GPT-4 have different compute requirements. More capable models generally require more resources per query, though hardware efficiency improvements can offset this.
- Query complexity A one-sentence factual question uses far less compute than a multi-step research task or image generation request. Neither Altman’s figure nor most academic estimates clearly define the query type they model.
- Cooling technology Older evaporative cooling systems consume more water. Newer liquid cooling and direct-to-chip cooling methods significantly reduce on-site water use. Microsoft announced next-generation data center designs in 2024 capable of running AI workloads with near-zero on-site evaporative water.
What It Looks Like at Scale
The per-query number is small. The scale is not.
OpenAI reported approximately 1 billion queries per day across ChatGPT as of December 2024. Applying Altman’s own 0.3 ml figure to that volume gives roughly 300,000 liters of direct water use per day equivalent to around 1,000 US households.
Applying the full-lifecycle estimate of 5–10 ml per query raises that to approximately 5–10 million liters per day.
Morgan Stanley projected that AI-related data centers could consume more than 1 trillion liters annually by 2028 an elevenfold increase from 2024 levels across all AI systems, not just ChatGPT.
The UC Riverside paper projected global AI demand would require between 4.2 and 6.6 billion cubic meters of water withdrawal in 2027 more than the total annual withdrawal of four to six countries the size of Denmark, or roughly half the United Kingdom’s annual total.
Putting the Numbers in Context
| Comparison | Water Used |
| Single ChatGPT query (Altman’s figure) | ~0.3 ml |
| Single ChatGPT query (full lifecycle) | ~5–10 ml |
| 100-word email via ChatGPT (full lifecycle) | ~519 ml |
| Standard Google search | ~0.2–0.5 ml |
| One cup of coffee | ~200 ml to grow the beans |
| 1 kg of beef | ~15,000 liters to produce |
AI queries are not among the most water-intensive activities humans engage in. But the combination of scale, geographic concentration in water-stressed regions, and rapid growth makes it a legitimate planning concern particularly for local water systems near major data center clusters.
What Is Being Done
The industry is not standing still on this.
Microsoft announced in December 2024 that its next-generation data center designs can operate with near-zero on-site evaporative water use by using chip-level cooling instead of traditional evaporative systems. Google and Microsoft have also explored scheduling compute-intensive AI tasks during cooler hours or routing them to regions with more abundant water supplies to reduce peak demand.
Whether these improvements scale fast enough to offset the growth in total AI usage is a separate question. The efficiency per query may improve while total consumption rises due to volume growth a pattern known in energy economics as the rebound effect.
What Remains Uncertain
It is worth being direct about what we do not yet know.
OpenAI has not published a full methodology for its 0.3 ml figure. The UC Riverside estimates are based on publicly available data center efficiency metrics that may not reflect OpenAI’s specific infrastructure. The figures for newer models like GPT-4o and GPT-5 have not been independently peer-reviewed at the time of writing. And query complexity which significantly affects compute demand is rarely controlled for in published estimates.
The honest position is that the per-query figure is somewhere between Altman’s 0.3 ml and the full-lifecycle estimates of 5–10 ml, with the actual number depending on infrastructure, location, query type, and what you choose to include in the accounting.
Key Takeaways
- ChatGPT uses between 0.3 ml and ~10 ml of water per query, depending on the measurement scope.
- OpenAI’s official figure (0.3 ml) covers direct cooling only. It does not include water used in electricity generation.
- The UC Riverside study estimates roughly 500 ml per 20–50 queries using a full-lifecycle methodology.
- Training GPT-3 consumed approximately 700,000 liters a one-time cost per model, not a per-query cost.
- At 1 billion daily queries, even the lowest estimate means hundreds of thousands of liters consumed every day.
- The numbers vary widely because different studies measure different things. Understanding scope is more useful than comparing raw figures.
- New cooling technologies are reducing per-query water use, but total AI water consumption continues to rise with usage volume.
FAQ
How much water does one ChatGPT query use?
Between approximately 0.3 ml and 10 ml depending on what is measured. OpenAI’s CEO stated 0.3 ml in June 2025, covering direct on-site cooling only. Full-lifecycle estimates that include electricity generation water use are higher, typically in the 5–10 ml range.
Does ChatGPT use more water than a Google search?
Most estimates suggest yes. A standard Google search uses roughly 0.2–0.5 ml of water. A ChatGPT query uses more compute and typically more water, particularly when indirect electricity water costs are included.
How much water did it take to train ChatGPT?
Training GPT-3 the model underlying the original ChatGPT is estimated to have consumed approximately 700,000 liters of freshwater, according to the UC Riverside study. This is a one-time cost per model version, not an ongoing per-query cost.
Why do different sources give such different numbers?
Because they measure different things. Some count only direct data center cooling. Others include the water used to generate the electricity powering the data centers. Location, model version, query complexity, and cooling technology all affect the outcome.
Is ChatGPT’s water use a serious environmental concern?
At the per-query level, the numbers are small. At the aggregate level given ChatGPT’s scale, the rapid growth of AI usage globally, and the geographic concentration of data centers in water-stressed areas it is a legitimate concern for water resource planning, even if it is not comparable to agriculture or heavy industry.
What is OpenAI doing about water use?
OpenAI has not published detailed commitments on water use. Microsoft, which provides much of OpenAI’s infrastructure, announced next-generation data center designs in late 2024 aimed at near-zero on-site evaporative water consumption. Google has also disclosed water efficiency targets in its annual environmental reports.
Conclusion
The water ChatGPT uses per query is genuinely small. Altman’s 0.3 ml figure is plausible for direct cooling alone, and even the more comprehensive academic estimates of 5–10 ml are not alarming in isolation.
The more relevant question is what happens when you multiply small numbers by enormous scale, sustained growth, and geographic concentration in already water-stressed regions. That is where the concern becomes legitimate not because any single query is costly, but because the infrastructure supporting billions of them is expanding faster than the water systems hosting it can comfortably absorb.
The debate between OpenAI’s official figure and the academic estimates is partly a measurement disagreement and partly a scope disagreement. Both can be true simultaneously. The honest position is to understand what each number includes and to hold all of them with appropriate uncertainty.
References
- Li, P., Yang, J., Islam, M. A., & Ren, S. Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models (UC Riverside / UT Arlington, 2023) arxiv.org/abs/2304.03271
- Altman, S. The Gentle Singularity (OpenAI blog, June 2025)
- Lawrence Berkeley National Laboratory 2024 Data Center Energy Usage Report (US Department of Energy)
- Morgan Stanley AI data center water projections (2024)
About the Author
I’m Sanwal Zia, an SEO strategist with more than six years of experience helping businesses grow through smart and practical search strategies. I created Optimize With Sanwal to share honest insights, tool breakdowns, and real guidance for anyone looking to improve their digital presence. You can connect with me on YouTube, LinkedIn, Facebook, Instagram, or visit my website to explore more of my work.
Disclaimer
All information published on Optimize With Sanwal is provided for general guidance only. Users must obtain every SEO tool, AI tool, or related subscription directly from the official provider’s website. Pricing, regional charges, and subscription variations are determined solely by the respective companies, and Optimize With Sanwal holds no liability for any discrepancies, losses, billing issues, or service-related problems. We do not control or influence pricing in any country. Users are fully responsible for verifying all details from the original source before completing any purchase.