AI infrastructure will use 2.27 billion cubic meters of water a year by 2030

By 2030, global AI infrastructure could consume up to 600 billion gallons of water a year, or roughly 2.27 billion cubic meters, according to analysts at Tom's Hardware, whose data 3DNews cites. The main cause of the growth is not water as a standalone resource but the rapidly rising power consumption of data centers: the more electricity AI accelerators need to train and run models, the more water goes both to cooling the hardware itself and to generating the electricity that powers it.

Where the water goes

Data centers have two main channels of water use, and both are directly tied to power consumption. The first is direct cooling of server racks and AI accelerators: many large compute clusters use evaporative cooling towers, where part of the water is lost irretrievably to the atmosphere as vapor to carry excess heat away from densely packed GPUs. The second, less obvious channel is electricity generation itself: thermal and nuclear power plants also use significant volumes of water to cool turbines, and rising data center power demand automatically adds load to this part of the chain as well.

• The forecast: up to 600 billion gallons of water a year by 2030.
• In metric terms: about 2.27 billion cubic meters.
• Source of the calculations: Tom's Hardware.
• Main growth driver: rising power consumption of AI data center hardware.
• Mechanism: water is used both to cool servers and to produce the electricity they run on.

Why power consumption has accelerated so sharply

Over the past few years, compute density in data centers has multiplied: new generations of AI accelerators draw noticeably more power per chip, and some clusters used to train large models contain tens of thousands of such chips. This is a fundamentally different load profile from classic cloud data centers built for web services and data storage. AI clusters run at the limit of their thermal capacity almost constantly, not in occasional peak scenarios, which pushes up both electricity bills and the water use tied to them.

How the industry is responding to the growing load

Cloud infrastructure operators have discussed data center water use for several years, especially in regions short of fresh water, where new clusters meet resistance from local communities and regulators. In response, large tech companies publicly say they are moving to more efficient cooling schemes: from direct-to-chip liquid cooling, which needs less water per unit of heat removed, to siting data centers in colder climates and using non-potable or recycled water instead of fresh water.

Against this backdrop, several large data center operators have publicly announced a "water positive" strategy: a goal of returning more water to the environment than they consume by 2030, through water ecosystem restoration projects and improved cooling efficiency. Whether such commitments can offset the AI-specific growth in consumption that Tom's Hardware describes is an open question. Most of these programs were announced before the current explosive growth in compute capacity for large language models, and their original calculations may not have accounted for the present pace of new cluster construction.

Another complication is that demand for AI compute is outpacing upgrades and expansion of the energy and water infrastructure in the regions where data centers are built. As a result, the burden often falls on the local water resources of specific areas rather than being spread evenly across a country or the world. That makes such forecasts a subject of debate not only among environmentalists but also among local authorities who regulate construction of new clusters. The 600 billion gallon forecast for 2030 essentially marks the upper bound of what the industry will have to handle if AI compute keeps growing at the current pace and cooling efficiency does not improve at a comparable rate.

For the industry as a whole, such forecasts add another argument for more energy-efficient model architectures: the less compute needed to reach the same answer quality, the lower the overall load on the power system and, with it, on the associated water consumption.