Why AI Uses So Much Water
You type a few words into your mobile phone, and in the blink of an eye, artificial intelligence (AI) creates a beautiful image. For example, a trend has emerged in recent days involving the creation of photographs in the style of the 1980s. The process has become so easy that we rarely think about the complexity behind it.
However, fulfilling that single “command” or instruction consumes a huge amount of water and electricity behind the scenes. It is natural to ask: when someone uses AI on a smartphone while sitting at home, not even a drop of water is consumed in their home. So why do AI companies say that they need so much water? In fact, to address this water consumption, they are even considering placing data centers beneath the sea.
Let us find out the answers to these questions.
Suppose you type into an AI tool such as Ollama or Midjourney: “Create an image of a sunset on top of a mountain.” Within a few seconds, the instruction travels through the internet to a distant data center.
Data centers are essentially enormous computer factories. Thousands of powerful servers work there continuously, day and night. Those servers create your image. Your phone or laptop has almost no role in the actual image-generation process.
The problem is that when thousands of servers perform complex calculations simultaneously, they become extremely hot. If the heat is not removed properly, these expensive machines could burn out. A huge amount of water is required to keep these hot servers cool.
Most large data centers use a method called evaporative cooling to cool their servers. It is somewhat like the way our bodies cool down through sweating. On a hot day, we feel cooler when our sweat dries in the air. Data-center cooling works in a similar way.
Cool water is circulated around the hot servers. The water absorbs the servers’ heat and becomes warm. It is then sent through the air and specialized systems inside a cooling tower. Some of the water evaporates and escapes into the air. Because of this evaporation, the remaining water cools down again. The cooled water is then sent back to the servers.
How Much Water Does AI Use?
According to an analysis published in September 2025 by the US-based engineering magazine IEEE Spectrum, a 2023 study estimated that generating a typical 150- to 300-word piece of text with the GPT-3 model consumed approximately 17 milliliters of water at an average data center in the United States.
Of that amount, slightly less than 2 milliliters was used directly for cooling. The rest was used in generating the electricity required to run the system.
The report did not provide a specific estimate for image generation. Different analyses have produced different figures for the amount of water required to create an image. Some estimates based on energy-consumption data suggest that generating one image uses approximately 23 milliliters of water on average. Other analyses have placed the figure between 5 and 60 milliliters, or even higher.
The difference is mainly due to the method used for calculation. The final figure depends greatly on whether the estimate includes only the water used to cool the servers or also the water required to generate the electricity.
By comparison, an ordinary Google search uses far less water. Various estimates place its water consumption at between 0.2 and 0.6 milliliters.
These figures may seem insignificant when considered individually. But think about this: billions of people around the world are generating images and text with AI every day. Where does all that seemingly small water consumption ultimately lead?
According to the Washington-based policy research organization Environmental and Energy Study Institute (EESI), a medium-sized data center consumes approximately 110 million gallons of water per year—roughly equivalent to the annual water demand of 1,000 households. Large data centers can require as much as 5 million gallons of water a day.
This water consumption is not limited to cooling servers. Data centers require enormous amounts of electricity, much of which still comes from fossil-fuel-powered or thermal power plants. Those power plants also need large quantities of water to cool their equipment.
In other words, the more electricity AI consumes, the more water it indirectly uses.
Where Does This Water Come From?
Data centers generally obtain water from local supply systems, rivers, or groundwater sources. A large proportion of this water is potable freshwater.
As a result, in many places, data centers are placing severe pressure on water allocated for local residents’ daily needs and for agriculture. The situation becomes even worse during the summer.
Facing criticism, major technology companies such as Microsoft, Google, and Amazon are turning to alternative water sources.
In its sustainability report, Microsoft said that it had increased its use of treated wastewater in places such as Texas, Quincy in Washington State, California, and Singapore.
Treated wastewater is water used by households or factories that has been scientifically processed at a wastewater-treatment facility and made suitable for reuse. Although it is not suitable for drinking, treated wastewater is safe for the environment and effective for cooling data-center servers.
Microsoft claims that approximately 99 percent of the water used by its data center in Singapore is now non-potable water.
Amazon Web Services (AWS), the technology division of Amazon, announced in June 2025 that it planned to use treated wastewater at more than 120 data centers in the United States by 2030. The company claims that this could save approximately 530 million gallons of potable water every year.
Google and Microsoft, meanwhile, have promised to become water positive by 2030. This means that they intend to return more water to the environment than they consume. Amazon has made the same commitment. The company claims that by the end of 2025, it had already achieved approximately three-quarters of that goal.
The Idea of Undersea Data Centers
To reduce pressure on potable-water supplies and save the enormous cost of cooling servers, AI companies are now looking toward the sea.
Seawater is freely available and naturally cold, making it an ideal solution for cooling servers.
Based on this idea, Microsoft installed an experimental data center called Project Natick beneath the sea near Scotland’s Orkney Islands in 2018. Approximately two years later, the company retrieved the facility and found that the underwater servers had failed eight times less frequently than an equal number of servers on land.
However, because of various practical difficulties involving maintenance and repairs, Microsoft ultimately decided not to pursue the project commercially. In 2024, Noel Walsh, head of the company’s cloud operations division, told the media directly that Microsoft was no longer building undersea data centers anywhere in the world.
Although Microsoft stopped its project, China has made considerable progress in this area. Chinese state media CCTV reported that the Chinese company Highlander had launched the world’s first commercial undersea data center off the coast of Hainan Island.
The facility weighs 1,300 tons and has been installed at a depth of 35 meters. Highlander claims that if 100 modules are installed there as planned, the facility will save approximately 122 million kilowatt-hours of electricity and 105,000 tons of potable water per year compared with a land-based data center of equivalent capacity.
How Can It Operate Without Potable Water?
On land, servers are cooled by evaporating water. The same process cannot be used beneath the sea.
Underwater, the servers are housed inside completely enclosed or sealed boxes known as modules. The cold saltwater surrounding the modules absorbs heat only through the outer shell of each box.
The process is similar to placing a cup of hot tea into a basin of ice-cold water. Because the seawater never comes into direct contact with the equipment, there is no risk of rusting. There is no evaporation in this system; only heat exchange takes place.
Furthermore, the enormous volume of seawater means that the small amount of water warmed by the facility has virtually no environmental impact.
Why Not Reuse the Same Water?
Many people may wonder why data centers do not simply circulate and reuse the same water repeatedly.
The main reason is water quality.
In evaporative cooling, when water evaporates, ordinary minerals dissolved in the water—such as iron, calcium, and salt—remain behind. As a result, the mineral concentration in the remaining water increases.
If this water is reused repeatedly, the minerals form a thick layer, known as scale, on the pipes and equipment inside the data center. The effect is similar to fat accumulating in a person’s blood vessels and obstructing blood flow. When scale builds up inside pipes, it restricts the flow of water and can seriously damage expensive equipment.
For this reason, data centers are often forced to discharge the old, mineral-rich water and replace it with new potable water.
What Are the Alternatives?
Two alternative methods currently being discussed are closed-loop cooling and immersion cooling.
Closed-loop cooling is somewhat similar to the way a car radiator keeps an engine cool. Water or another specialized liquid circulates through a sealed pipe and does not come into contact with the outside air. As a result, the liquid cannot evaporate and can be reused repeatedly.
In immersion cooling, the entire server is submerged in a special electrically nonconductive chemical liquid.
These methods can reduce water wastage to almost zero. However, they create a new challenge.
Because the water or liquid does not evaporate and cool naturally, data centers must use enormous electric fans or powerful air-conditioning systems—such as chillers or refrigeration systems—to cool the hot liquid inside the closed system. These systems require a great deal of electricity. In addition, installing the entire system is extremely expensive.
There is therefore a common saying in the blog of technology industry:
If you want to save electricity, you may use more water; if you want to save water, you may consume more electricity.
Creating an environmentally friendly and sustainable balance between these two resources is now one of the greatest challenges facing engineers.
