Is artificial intelligence's water consumption threatening our water security?

Research shows that ChatGPT queries use approximately 10 times more electricity than a Google search. The total water consumption of large technology companies using artificial intelligence is increasing every year. The increasing water and energy consumption of artificial intelligence is causing concern. These developments necessitate questioning the relationship between artificial intelligence data centres' water consumption and our water security.

Large data collection and processing systems that power artificial intelligence often rely on clean water sources for both energy supply and cooling systems. This clean water is primarily used to cool servers and maintain a stable ambient humidity. According to the OECD, by 2027, the annual water consumption of artificial intelligence systems will range between 4.2 and 6.6 billion cubic metres. This water consumption will be close to Turkey's annual net total drinking and domestic water consumption.

The energy and water needs of large data collection centres that feed artificial intelligence are increasing. This will directly and indirectly affect the water consumption of data collection centres located in regions affected by climate change, thereby increasing water scarcity.

Developments in artificial intelligence will also lead to an increase in the relationship between water for energy and energy for water. We will need more water for energy and more energy for clean water.

Currently, there is no study determining the water footprint of data centres in our country, nor is there any public data on the water consumption of data centres.

Large companies' ‘water positive’ goal is not realistic

Some technology companies are developing new technologies to reduce this consumption. Meta is promoting ‘dry cooling’ technology, while companies like NVIDIA and Alibaba are transitioning to direct liquid cooling methods.

Microsoft and Amazon Web Services have announced that they will be “water positive” by 2030, meaning they will return more water to nature than they consume. Apple aims to certify its data centres with the Water Management Alliance's certifications.

High-quality water used in evaporative cooling and humidification systems evaporates without being recovered after use. According to research, the evaporation rate of water reaches up to 74% at Google and 60% at Microsoft. This means that billions of litres of water are permanently extracted from nature each year.

To meet their commitments to become water positive within five years, large technology companies need to transition to more efficient operations, cooling, and energy systems. They also need to adopt new-generation technologies that reduce the rate of water evaporation without recovery. Additionally, they must reduce their indirect water footprint in energy consumption by transitioning to renewable energy sources.

Although today's highly advanced artificial intelligence models consume a lot of energy, many research studies are being conducted to make them more efficient. Technological developments and optimisation models in these studies may enable artificial intelligence to operate with less energy. For artificial intelligence to operate with less energy, there is a need for optimisation in the areas of hardware, software, and algorithms, as well as a more efficient energy infrastructure. If these studies are carried out together, energy efficiency can be increased.

Artificial intelligence that consumes less water and energy is possible

Using less energy and less water in artificial intelligence data centres are emerging as complementary processes. The energy efficiency of artificial intelligence data centres is very important. An energy-efficient artificial intelligence data centre will also reduce water consumption, especially since water is used in electricity production. Data centres powered by renewable energy (wind and solar) will consume less water and have a lower carbon footprint.

The amount of water used in cooling systems, i.e. the cooling water footprint of artificial intelligence, is being reduced with new generation systems. New-generation data centre cooling systems also use a two-stage liquid cooling system. In this system, the liquid absorbs heat, then evaporates, condenses, and re-enters the cycle. These systems require fewer pumps and provide high efficiency for high-density servers.

In addition, rapidly developing quantum computers are cooled using cooling gases such as helium and closed-loop systems instead of water. Quantum computers can perform much more work with much less energy than classical systems for certain tasks. In terms of water consumption, they are said to be much more advantageous than classical data centres.

Cooling alternatives for artificial intelligence systems

Artificial intelligence applications, cloud systems, and big data centres cause serious water consumption due to the cooling technologies used to reduce the heat of server systems. The electrical energy used in data centres enables servers to operate at high performance; however, this process generates significant heat.

Wet, dry, evaporative, and hybrid systems are used to cool these systems. In addition to wet cooling, dry cooling is also possible in data centres. However, dry cooling uses ambient air for cooling, so its efficiency decreases in hot climates. Larger fans have to work harder for cooling. Therefore, they consume more energy. Wet cooling and humidification offer higher efficiency. Wet cooling systems can be 20% to 30% more efficient than dry cooling in terms of energy consumption. Hybrid systems that use both air and liquid cooling systems can automatically switch between each other depending on the ambient temperature, season, and workload of the data centre.

For this reason, they are considered a more efficient cooling system for businesses.

Does Artificial Intelligence Have a High Water Footprint?

According to research conducted at the University of California Riverside (UC Riverside) campus, the massive data centres required to run artificial intelligence systems consume a significant amount of water, even for simple queries. According to the research results, an average of half a litre of water evaporates for every 5-50 ChatGPT queries. This water is considered to be completely consumed due to the cooling systems used to maintain the temperature of the systems at a certain level. In other words, there is no significant water return in the total water used.

If we only consider the water used directly for cooling in the data centre (excluding water used for electricity production), this figure is close to 0.5 litres per 300 queries. In practice, this means that ChatGPT's direct water consumption is relatively low but could increase as usage grows.

Li, Ren et al. (2023) and data from the U.S. Census Bureau have revealed the following results when comparing AI's water usage with other uses.

• A hamburger requires approximately 660 gallons of water, which is equivalent to approximately 200,000 ChatGPT queries.
• One hour of television viewing in an American household consumes approximately 4 gallons of water, equivalent to 1,200 ChatGPT queries.
• Leaky pipes in the United States waste over 10.5 billion gallons of water daily, while ChatGPT's estimated global daily water consumption is approximately 3.5 billion gallons.

Energy consumption follows a similar pattern. A Google search consumes approximately 0.3 watt-hours, while a ChatGPT query consumes approximately 3 watt-hours, which is 10 times more energy. However, watching a video for an hour consumes significantly more energy than ChatGPT.

Research by Shaolei Ren of the University of California, Riverside, along with statements from Microsoft and Google, shows that while artificial intelligence has a significant and measurable water and energy footprint, its environmental impact is relatively small compared to meat production, inefficient infrastructure, and digital entertainment. The primary challenge regarding water security lies not only in artificial intelligence itself but in water management across industries.