OpenAI CEO Sam Altman recently pushed back against some of the strongest claims about AI's water footprint. During an appearance on the Sources podcast with Alex Heath, Altman argued that the amount of water used by modern AI infrastructure is often exaggerated and that the technology used to cool large data centers has changed significantly.
His most attention-grabbing statement involved an unusual comparison. Altman said that producing a single almond in California uses about as much water as 38,000 ChatGPT queries. However, he also made an important qualification: he was recalling the number from memory and did not have the calculation in front of him.
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.
That makes the claim interesting, but it should not be treated as a precise scientific measurement. The larger question is much more useful: How much water do AI data centers actually use, and are newer cooling technologies really changing the situation?
Why Do AI Data Centers Need Water?
AI systems run on large numbers of processors that generate heat while performing calculations. The more powerful the hardware becomes, the more important efficient cooling becomes.
Traditional data centers have used several cooling approaches. Some facilities rely heavily on cooling towers, where water evaporates as heat is removed. Other facilities use air cooling or hybrid systems that combine different methods depending on weather conditions.
AI is adding another challenge. High-performance GPUs and other accelerators can produce large amounts of heat in a relatively small physical space. This has encouraged data center operators to look beyond conventional air cooling.
Liquid cooling is one of the major technologies being developed for this purpose. Instead of relying entirely on cold air moving through a server room, liquid can be brought much closer to the components generating the heat.
What Did Sam Altman Actually Say?
Altman's comments were primarily a response to claims that using ChatGPT has an enormous hidden environmental cost because of water consumption.
He argued that the discussion often relies on older assumptions about data center cooling. According to Altman, modern large data centers can have water use comparable to an office building when considering ordinary uses such as sinks and toilets.
There is an important distinction here. Altman did not say that every data center uses exactly the same amount of water. He acknowledged that older evaporative cooling systems could use substantial quantities of water.
The disagreement is therefore not simply about whether AI uses water. It is about how much water is used, where it is used, what cooling system is installed, and which parts of the water footprint are included in the calculation.
The 38,000-Query Almond Claim Needs Context
The almond comparison became the headline because it is easy to understand. Instead of talking about liters, cooling towers, and data center infrastructure, it gives readers something familiar to compare with ChatGPT.
But the comparison has limitations.
Water used to grow an almond and water associated with an AI query are not measured in exactly the same way. Agricultural water use can involve irrigation and other parts of the growing process, while an AI water footprint can include direct cooling water and, depending on the methodology, water associated with electricity generation.
There is also no universal amount of water for every ChatGPT query. A short text request, a long reasoning task, an image-generation request, and other AI workloads can require different amounts of computing.
Data center location matters too. A facility operating in a cool climate can have different cooling requirements from one operating in a hot region. The design of the building, the hardware density, the electricity source, and the cooling architecture can all influence the final number.
Important: The 38,000-query figure should be treated as Sam Altman's stated estimate, not as a universal measurement of ChatGPT's water consumption.
Modern Data Centers Are Changing Their Cooling Systems
This is where Altman's argument has some important technical context.
Data center cooling has evolved considerably. Companies are developing systems that can reduce dependence on evaporative cooling, and some new facilities are being designed specifically around AI workloads.
Microsoft is one example. The company says it introduced a new data center design in 2024 that is optimized for AI workloads and uses a closed-loop, direct-to-chip cooling system. According to Microsoft, the design can operate with zero water consumption for cooling during operation because the liquid is recirculated rather than continually evaporated.
Read Microsoft's explanation of its water-efficient data center cooling technology here.
That does not mean every Microsoft data center is water-free. The company's existing infrastructure includes different cooling approaches, and conditions vary from one facility to another.
Still, the technology demonstrates why older assumptions about AI cooling should not automatically be applied to every new data center.
Liquid Cooling Could Become Increasingly Important
The growth of AI is pushing data centers toward higher computing density. More powerful processors generate more heat, and traditional air cooling can become less practical as the amount of heat produced by individual racks increases.
Liquid cooling can address this problem by moving heat away from chips more efficiently. In a direct-to-chip design, a cooling liquid circulates through components positioned close to the processors.
The liquid can then carry the heat away from the server and transfer it to another part of the cooling system.
This approach does not automatically mean that a data center has zero environmental impact. The system still needs energy, equipment, infrastructure, and maintenance. However, it can substantially change how much water is needed for cooling in the right environment.
NVIDIA Is Also Working on Water-Efficient AI Cooling
NVIDIA has been developing liquid-cooling infrastructure for high-density AI systems. Its approach is designed around the increasing heat output of modern AI computing hardware.
One of the key advantages of a closed-loop liquid-cooling system is that the same cooling liquid can circulate repeatedly. That is different from systems that depend on continuously evaporating water to remove heat.
NVIDIA has also discussed the use of higher-temperature liquid cooling and dry coolers in suitable environments. The goal is to reduce the need for evaporative cooling while still keeping high-performance AI hardware within acceptable operating temperatures.
See NVIDIA's explanation of liquid cooling for AI data centers here.
The exact results depend on the facility and climate, so claims about water savings should not be interpreted as a guarantee for every AI data center.
Amazon Shows Why the Numbers Can Be Complicated
Amazon's data center operations provide another useful example of why the AI water discussion needs more than a single headline number.
Amazon says its global data centers achieved a water-use effectiveness figure of 0.12 liters per kilowatt-hour of IT load in 2025. The company says this represented a 52% improvement compared with 2021.
Amazon also says its facilities rely on air cooling for most of the year, while water cooling is used during particularly hot conditions at some locations.
Read Amazon's latest explanation of its data center water efficiency efforts here.
Amazon reported that its global data center operations withdrew about 2.5 billion gallons of water during 2025. That figure is important because it demonstrates both sides of the debate: efficiency can improve significantly, but a huge global data center footprint can still involve substantial total water use.
Why Location Matters More Than a Global Average
One of the easiest mistakes in discussions about data center water consumption is to focus only on a global average.
Water is a local resource. A certain amount of water consumption may be manageable in one region but become controversial in another location experiencing drought, population growth, or pressure on municipal supplies.
This means two data centers with similar computing capacity could create very different local concerns.
A facility using a water-efficient cooling system in an area with abundant resources presents a different situation from a large facility using evaporative cooling in a water-stressed region.
For communities, the most useful questions are therefore local: How much water will the facility withdraw? Where does that water come from? How much is consumed rather than returned? Which cooling technology is being used? And how will those numbers change as the facility expands?
AI Efficiency Does Not Automatically Mean Lower Total Consumption
There is another issue that deserves attention: scale.
Technology companies can make individual AI tasks more efficient while the overall demand for AI continues to increase.
Imagine that a new generation of hardware requires less energy and water for each task. If the number of AI tasks grows several times faster, total resource consumption could still rise.
This is particularly relevant as AI expands beyond ordinary chatbots. Image generation, video generation, coding tools, AI agents, research systems, and enterprise applications can create workloads that are much more demanding than a simple text question.
For that reason, measuring the environmental impact of AI requires more than calculating the footprint of a single prompt.
Training AI Is Different From Answering a Prompt
Another important distinction is between AI training and AI inference.
Training involves running large amounts of computation for extended periods while a model learns from its training data. Inference is what happens when a user asks a model to generate an answer or perform a task.
These activities have different resource requirements.
A viral claim about one ChatGPT query does not tell us how much water was required to build and train the underlying model. Likewise, a large training footprint does not mean that every individual prompt has an equally large footprint.
A responsible discussion should keep these stages separate.
Should You Feel Guilty About Using ChatGPT?
For ordinary users, the answer is probably no.
It is reasonable to care about the environmental impact of technology, but turning every ChatGPT question into a personal environmental guilt trip misses the larger issue.
The bigger responsibility lies with companies designing and operating the infrastructure at enormous scale. Better hardware, efficient cooling, responsible facility locations, improved electricity sourcing, and transparent reporting can have a much larger effect than asking individual users to stop making ordinary queries.
That does not mean individual behavior is irrelevant. Using technology thoughtfully is always reasonable. But environmental discussions are more useful when they focus on measurable infrastructure rather than emotionally powerful comparisons.
What We Can Actually Learn From Sam Altman's Statement
Altman's comments do not prove that AI has no water problem. They do, however, highlight an important weakness in simplistic claims about AI and water.
There is no single number that accurately describes the water consumption of every AI query, every data center, or every AI company.
The technology is changing quickly. Cooling systems are becoming more sophisticated, and some new AI-focused data center designs are being built around closed-loop systems that can dramatically reduce cooling-water consumption.
At the same time, the total amount of AI infrastructure is growing rapidly. That means efficiency improvements need to continue if companies want overall resource consumption to remain under control.
The right question is not simply “Does AI use water?” It does. The better question is how much water a particular system uses, where that water comes from, and how efficiently the infrastructure operates.
Final: Sam Altman's argument deserves more nuance than the headline alone suggests
Modern AI data centers are not identical to older facilities, and cooling technology has advanced considerably. Microsoft's closed-loop approach, NVIDIA's work on liquid cooling, and Amazon's reported improvements in water efficiency all show that the industry is actively looking for ways to reduce resource consumption.
But that does not make concerns about water disappear.
Large data centers can still consume significant amounts of water, and their impact can be particularly important in regions where water resources are already under pressure. The rapid expansion of AI also means that improvements in efficiency must be measured against growing demand.
The 38,000-queries-per-almond comparison is therefore best understood as an attention-grabbing estimate rather than a final answer. Altman himself acknowledged that he was recalling the figure from memory.
The most accurate conclusion is somewhere in the middle: AI data center water fears may sometimes rely on outdated assumptions, but water consumption remains a legitimate infrastructure and environmental issue.
As AI continues to expand, transparency will matter just as much as efficiency. Users, communities, and regulators need reliable information about how much water data centers actually consume and how new cooling technologies are changing those numbers.
In the end, the debate is not really about whether one almond uses more water than thousands of ChatGPT queries. It is about whether the AI industry can keep scaling its computing infrastructure while becoming significantly more efficient with the resources it depends on.
0 Comments