AI Infrastructure Will Consume 2.27 Billion Cubic Meters of Water Annually by 2030

By 2030, global AI infrastructure could consume up to 600 billion gallons of water per year — approximately 2.27 billion cubic meters — according to calculations by analysts from Tom's Hardware publication, cited by 3DNews. The primary reason for the increase is not water itself as a separate resource, but the rapidly growing energy consumption of data centers: the more electricity AI accelerators require for training and running models, the more water is spent both on cooling the equipment itself and on generating the electricity that powers it.

Where Water Consumption Comes From

Data centers have two main water consumption channels, and both are directly linked to energy consumption. The first is direct cooling of server racks and AI accelerators: many large computing clusters use evaporative cooling towers, where a portion of water is permanently lost to the atmosphere as steam to dissipate excess heat from densely packed GPUs. The second, less obvious channel is electricity generation itself: thermal and nuclear power plants also consume significant volumes of water for turbine cooling, and increased data center energy consumption automatically increases the load on this part of the chain.

• Forecast — up to 600 billion gallons of water per year by 2030.
• Converted to metric system — approximately 2.27 billion cubic meters.
• Source of calculations — Tom's Hardware publication.
• Primary growth driver — increased energy consumption of AI data center equipment.
• Mechanism — water is spent simultaneously on cooling servers and producing electricity to power them.

Why Energy Consumption Growth Has Accelerated So Sharply

Over the past few years, the density of computing in data centers has increased several times over: new generations of AI accelerators consume significantly more energy per chip, and the quantity of such chips in individual clusters for training large models is in the tens of thousands. This represents a fundamentally different load profile compared to classical cloud data centers designed for web services and data storage — AI clusters operate at thermal capacity limits almost constantly, rather than in rare peak scenarios, which drives up both electricity bills and the associated water consumption.

How the Industry Responds to Growing Demand

The problem of data center water consumption has been discussed for several years by cloud infrastructure operators, especially in regions with fresh water scarcity, where construction of new clusters meets resistance from local communities and regulators. In response, major technology companies publicly announce transitions to more efficient cooling schemes — from direct liquid cooling of chips, which requires less water per unit of dissipated heat, to placing data centers in colder climates and using non-potable or recycled water instead of fresh water.

Against this backdrop, several major data center operators have publicly declared a "water positive" strategy — an intention by 2030 to return more water to the environment than they consume, through projects for water ecosystem restoration and improved cooling efficiency. Whether such commitments will be able to compensate for specifically AI-related consumption growth as described by Tom's Hardware remains an open question: most such programs were announced before the current explosive growth in computing capacity for large language models, and their initial calculations may not have accounted for current rates of construction of new clusters.

Another complication is that the growth in demand for AI computing outpaces the updating and expansion of energy and water infrastructure in the regions where data centers are being built. As a result, the load often falls precisely on the local water resources of specific territories rather than being evenly distributed across a country or worldwide, which makes such forecasts a subject for discussion not only among environmental scientists but also among local authorities regulating new cluster construction. The forecast of 600 billion gallons by 2030, in essence, marks the upper boundary of what the industry will need to handle if AI computing growth rates persist and cooling efficiency does not improve at comparable rates.

For the industry as a whole, such forecasts become additional motivation for more energy-efficient model architectures — the fewer computations required to obtain the same quality of answer, the lower the total load on the power system and, correspondingly, the associated water consumption.