Thirsty Intelligence: How AI Data Centers Are Draining Water Systems

01The Scale of AI's Water Footprint

When most people think about the environmental cost of artificial intelligence, they picture electricity. Server racks humming in climate-controlled warehouses, drawing megawatts from the grid. But there is a second resource that AI consumes in staggering quantities, one that is far less visible and far less discussed: water. Every time a large language model generates a response, somewhere a cooling tower is evaporating water into the atmosphere to keep the GPUs from overheating.

The numbers are genuinely difficult to wrap your head around. Researchers at the University of California Riverside estimated that training GPT-3 alone consumed roughly 700,000 liters of clean freshwater. That is enough to fill a decent-sized swimming pool, and GPT-3 was a relatively modest model by today's standards. The next generation of models, trained on far larger compute clusters, multiplied that footprint several times over. Inference, the phase where users actually query the model, adds a steady ongoing draw that compounds across billions of interactions.

What makes this particularly slippery is that the water does not come back in the same form. Evaporative cooling turns liquid water into vapor that disperses into the atmosphere. It is not returned to the local watershed in a usable state. When a data center reports its water consumption, it is describing a net withdrawal from a resource that communities, agriculture, and ecosystems also depend on. The scale of this withdrawal has grown quietly alongside the AI boom, and only recently has public attention caught up to the magnitude of what is happening.

02How Data Center Cooling Works

To understand why AI is so thirsty, you have to understand what happens inside a data center. The core problem is heat. Modern GPUs and TPUs, the chips that power AI training and inference, generate enormous amounts of waste heat when they operate at full utilization. A single high-end AI accelerator can draw over 700 watts and shed nearly all of that as thermal energy. Stack thousands of them in a facility and you have the thermal output of a small steel mill.

The primary mechanism for dealing with this heat is a cooling loop. Inside the facility, chips transfer their heat to a closed-loop liquid circuit, often through cold plates mounted directly on the processors or through immersion cooling where entire servers are submerged in a dielectric fluid. That internal loop carries heat to a heat exchanger, where it transfers to a second loop. The second loop is where the water story begins, because that is typically where the heat reaches the atmosphere.

The most common approach on the second loop is evaporative cooling. Warm water from the heat exchanger flows to a cooling tower, where it is sprayed over a fill material and exposed to moving air. A portion of the water evaporates, and evaporation carries heat away with it. The cooled water then returns to the heat exchanger to pick up more heat, and the cycle repeats. The physics are highly efficient, but the cost is water itself, consumed by the physics of phase change and carried away on the wind.

03Evaporative Cooling vs Alternatives

If evaporative cooling is so water-intensive, why not use something else? The answer is that alternatives exist, but each comes with tradeoffs. Air cooling, the simplest approach, uses large fans to blow air across server racks. It requires no water for the cooling process itself, but it is far less efficient at handling the extreme heat densities of modern AI workloads. A facility relying solely on air cooling would need dramatically more energy per unit of compute, and that energy itself has a water footprint at the power plant.

Dry cooling, which uses closed-loop air-to-refrigerant heat exchangers without evaporation, is another option. It eliminates on-site water consumption entirely. The penalty is efficiency. Dry coolers can only reject heat to the ambient air temperature, which means on hot days they struggle to maintain safe operating temperatures or must run fans at maximum power, consuming additional electricity. In many climates, the energy penalty of dry cooling is significant enough that operators choose evaporative systems instead.

A growing middle ground is direct-to-chip liquid cooling and immersion cooling, which capture heat more efficiently at the source and reduce the overall volume of heat that reaches the facility level. These approaches can cut total facility water consumption substantially, especially when paired with dry coolers or adiabatic systems that only use water during peak temperature periods. The technology is maturing rapidly, but retrofitting existing data centers is expensive and slow, and the bulk of today's AI compute still runs on infrastructure built around evaporative towers.

04The Numbers Behind the Claims

The public conversation about AI water use is full of contradictory claims. One widely shared statistic suggests a single ChatGPT conversation consumes about 500 milliliters of water, roughly the contents of a standard water bottle. That figure, derived from research by Shaolei Ren and colleagues at UC Riverside, is based on average cooling water intensity for inference on large language models. It is a useful benchmark, but it is also a lightning rod for misunderstanding because the actual per-query consumption varies enormously depending on the model, the data center, the climate, and the time of year.

Training is where the numbers get really dramatic. The same researchers estimated that GPT-3 consumed approximately 700,000 liters during training. GPT-4, with a substantially larger compute footprint, is estimated to have consumed somewhere around 2.5 million liters. If the trajectory continues, a frontier model trained in 2026 could require 5 million liters or more. These are not firm, audited numbers, and the companies involved have not published precise water-per-model figures. But they are grounded in published power consumption data and standard cooling efficiency assumptions, and they align with the broader pattern of exponential AI compute growth.

The complication is that these estimates rest on assumptions about where the training occurred and what cooling technology was in use. A model trained in a data center in Ireland, where cool maritime air reduces evaporative demand, will have a very different water footprint than the same model trained in Arizona during summer. Without granular, facility-level data from the companies involved, researchers must make reasonable but uncertain assumptions. This is why the headline figures should be treated as order-of-magnitude estimates rather than precise measurements, and why calls for mandatory water reporting from data center operators have grown louder.

05Geographic Water Stress and Data Center Siting

Where a data center is built matters as much as how it is cooled. A facility drawing two million gallons of water per day in a region with abundant rainfall and full reservoirs has a fundamentally different environmental impact than the same facility in a region experiencing chronic drought. This is the concept of water stress: the ratio of total freshwater withdrawal to available renewable freshwater resources. In highly stressed watersheds, any additional withdrawal competes directly with municipal supply, agriculture, and ecosystem needs.

This creates a tension that the AI industry has not fully resolved. Many of the most attractive locations for data centers, places with cheap land, cheap power, and favorable tax incentives, happen to be in arid regions. The American Southwest, particularly Arizona and New Mexico, has seen massive data center construction despite being among the most water-stressed areas in the country. Northern Virginia, the densest data center corridor in the world, draws water from the Potomac River watershed, which while not currently in crisis, is under increasing pressure from population growth and climate variability.

Siting decisions increasingly factor in water availability, but the economic incentives do not always align with environmental considerations. Local governments eager for data center investment and tax revenue may not enforce water withdrawal limits aggressively. And once a data center is built, it is extraordinarily difficult to relocate. The infrastructure represents billions in sunk capital, and the companies that own it will fight to keep operating regardless of changing water conditions. This is why forward-looking water stress assessment at the siting stage is so critical, and why advocates argue it should be a regulatory requirement, not a voluntary best practice.

06What Hyperscalers Are Doing About It

The major cloud and AI companies are not oblivious to the problem. Google, Microsoft, Amazon, and Meta have all published water stewardship commitments, and several have set targets to become "water positive" by the end of the decade, meaning they pledge to replenish more water than they consume. Google's water stewardship program includes investments in watershed restoration, reclaimed water systems, and cooling efficiency improvements. Microsoft has committed to replenish more water than it consumes by 2030 and has funded projects across multiple watersheds.

On the technology side, these companies are investing in advanced cooling architectures. Liquid cooling, once a niche approach used mainly in supercomputing, is becoming standard in new AI-focused facilities. Direct-to-chip cold plate systems capture heat at the source, reducing the volume of heat that reaches the cooling tower and thus the volume of water evaporated. Some facilities are experimenting with immersion cooling, where servers are submerged in specialized dielectric fluids that absorb heat far more efficiently than air, eliminating the need for evaporative towers entirely for that portion of the load.

Reclaimed water is another lever. Several hyperscaler facilities now use treated municipal wastewater for cooling, avoiding withdrawal from drinking water supplies. This approach is sound in principle but requires proximity to wastewater treatment plants and appropriate infrastructure to deliver the water to the data center. It also does not eliminate water consumption, since evaporative cooling still turns reclaimed water into vapor. It simply changes the source. True reduction in consumption requires either moving away from evaporative cooling entirely or dramatically improving the efficiency of the entire compute stack so that less heat is generated in the first place.

07The Path to Sustainable AI Compute

The fundamental challenge is that AI compute demand is growing faster than efficiency improvements can offset it. Every generation of chips is more energy-efficient per operation, but the number of operations being performed is increasing at an even faster rate. This is the Jevons paradox in action: efficiency gains make computation cheaper, which drives more computation, which increases total resource consumption. Water is caught in the same dynamic because it is coupled to energy through the cooling system.

Several pathways could break or soften this cycle. The most direct is a shift away from evaporative cooling toward dry or hybrid systems that minimize or eliminate on-site water consumption. This requires accepting higher energy costs and higher capital costs for cooling infrastructure, but as water scarcity intensifies, the economics may shift. Regulation will play a role here. If water consumption from data centers is priced to reflect its true scarcity value, or if withdrawal limits are enforced in stressed watersheds, operators will have a direct financial incentive to invest in water-free cooling.

Transparency is the other critical lever. Right now, the public understanding of AI water consumption is hampered by the lack of standardized, facility-level reporting. Companies disclose aggregate water figures in sustainability reports, but the granularity needed to assess local impact, water per query, water per training run, and water source breakdown is largely absent. Mandatory disclosure frameworks, similar to what exists for greenhouse gas emissions, would enable better research, better policy, and more informed consumer and enterprise decisions. The technology to build AI without draining watersheds exists. What remains uncertain is whether the industry, regulators, and the public will move fast enough to deploy it at the scale and speed the situation demands.