Why do AI data centers need so much water?
In September 2026 Sam Altman, cofounder and CEO of OpenAI, pushed back against claims about the vast quantities of water consumed by artificial intelligence (AI) data centers. According to Altman, growing a single California almond requires roughly as much water as ChatGPT uses to process about 38,000 prompts. In a blog post published the previous year, he had put the water consumption of a single ChatGPT query at approximately 0.3 milliliter (0.01 fluid ounce), or about one-fifteenth of a teaspoon. However, researchers at the University of California, Riverside, who studied GPT-3 offered an estimate that is substantially higher than Altmanâs claim. In 2025 they suggested that GPT-3 consumes roughly 500 milliliters (17 fluid ounces) of water to process every 10 to 50 medium-length queries, depending on when and where it is run. And ChatGPT is just one of many large language models currently in use.
An AI data center is a specialized computing facility housing dense clusters of hardware for training and running AI models, workloads that typically require greater power and cooling capacity than those in conventional data centers. Externally, they resemble large industrial warehouses. Their interiors are packed with rows of servers in addition to associated hardware, backup generators, and cooling systems. As of 2026 these are some of the largest AI data centers in the world:
- SpaceXAIâs Colossus 2 in Memphis, Tennessee, has about 946 megawatts (MW) of information technology (IT) power capacity and a computing capacity of about 1,112,000 H100-equivalent graphics processing units (GPUs).
- Amazon Web Servicesâ Project Rainier in New Carlisle, Indiana, has about 910 MW of IT power capacity and a computing capacity of about 686,000 H100-equivalent GPUs.
The exact quantity of water consumed by AI data centers remains debated. Water consumption ultimately depends on factors such as a data centerâs location and the nature and scale of its operations. It is generally accepted that the water usage of most data centers is considerable. What makes AI data centers so thirsty?
In data centers a large amount of heat is generated as the servers operate continuously. Removing this heat is vital to prevent damage to the equipment. One common method of expelling this heat from the facilities is water-based cooling. In facilities equipped with evaporative cooling towers, some of the circulating water evaporates, carrying heat out of the data center into the atmosphere. Evaporation of water also helps to cool outside air passed through the system to carry away the heat. By one estimate about 80 percent of the water used in evaporative cooling is lost in the process. Despite this significant water footprint, most large, modern data centers use evaporative cooling because it is energy efficient.
" height="20" alt="HEW Logo">
Water is also used indirectly to support the operation of AI data centers. Large quantities of water are used by the power plants that generate the electricity needed to run these facilities. Water is also consumed at various stages of the AI supply chain, such as in the manufacture of microchips.
AI data centers typically use clean fresh water to prevent pipe blockage and bacterial growth. This water often comes from municipal supplies or groundwaterâthe same sources that provide drinking water to local communities. This has raised concerns about water depletion, as estimates suggest that a large data center can use up to 19 million liters (5 million gallons) of water per day, or about 6.8 billion liters (1.8 billion gallons) annuallyâcomparable to the water use of a town of 10,000 to 50,000 people.
- The average data centerâs water usage effectiveness (WUE)âa metric to evaluate how efficiently a data center uses water relative to its energy consumptionâis 1.9 liters (0.5 gallon) per kilowatt-hour (kWh); an ideal score is 0 liters per kWh.
- Roughly 40 percent of U.S. data centers are located in regions experiencing severe water stress.
- About 57 percent of the water used directly by data centers is potable, and 80â90 percent comes from âblueâ sources such as lakes, rivers, and aquifers.
- By 2023, Amazon Web Services was using purified wastewater in 20 data centers.
- Generating the electricity required for U.S. data centers consumed an indirect 799 billion liters (211 billion gallons) of water in 2023.
- The noise of multiple servers operating at the same time may exceed 96 decibels (dB), and data center generators may produce noise levels up to 105 dB, both of which exceed the permissible industrial noise level of 85 dB.
- Global data centers consumed an estimated 485 terawatt-hours (TWh) of energy in 2025, representing about 1.5 percent of global electricity generation, an amount equivalent to the annual electricity generation of Germany. AI-focused data centers consumed 155 TWhâroughly equal to 0.5 percent of total global electricity generation.
Water-saving alternatives
By 2028, AI data centers in the United States are projected to consume up to 121 billion liters (32 billion gallons) of water annually, which is roughly equivalent to the indoor water use of 360,000 households. Although this is small compared with the water use of sectors such as agriculture and electricity generation, the concentration of data centers in water-scarce regions may exacerbate existing water stress. Key players in the AI industry are therefore shifting to or experimenting with alternative technologies that reduce water use.
Several alternative cooling techniques consume less water than traditional open-loop evaporative cooling. For example, in closed-loop cooling systems, cold water is circulated through sealed pipes to absorb excess heat and is recirculated after cooling with little or no evaporative loss. Although such systems can reduce freshwater use by up to 70 percent compared with evaporative cooling, they are more expensive and complex to maintain. Newer techniques in which cooling occurs directly at the server or chip level include direct liquid cooling (also called direct-to-chip cooling) and immersion cooling. In direct liquid cooling, a specialized coolant is piped directly to components such as processors, whereas in immersion cooling, servers or other components are immersed in a nonconductive coolant. However, these systems still require a heat exchanger or another heat-rejection system to transfer the captured heat to the external environment.
The trusted destination for professionals, college students, and lifelong learners.
SUBSCRIBE
More conventional alternatives include air-conditioningâbased cooling (air cooling) and the use of cold outside air or water for cooling (free cooling). These approaches are particularly suitable for small-scale operations in colder climates or areas where water is scarce, although some may require greater energy expenditure to operate cooling equipment.
Using reclaimed (treated) wastewater to cool data centers is another option being explored by big tech companies, including Amazon, Meta, Google, and Microsoft. Some facilities are also experimenting with reusing water across multiple cooling cycles, reducing the need to add fresh water during each cycle. AI data centers can also indirectly reduce their water consumption by shifting to renewable energy sources that require little or no water for electricity generation.