You ask ChatGPT a question. It answers in seconds. What you don't see is the water that evaporated in a cooling tower hundreds of miles away—water that could have supplied a household for a day. Every AI interaction, from a simple search summary to a multi-hour training run, carries a hidden water cost that is not billed to you but is paid by local communities and ecosystems. This article breaks down exactly where that water goes, how much is consumed per task, and what developers, data center operators, and policymakers can actually do about it. No hand-waving, no scare tactics—just the real numbers, the real trade-offs, and the real options for reducing the footprint without turning off the servers.
Data centers that host AI workloads generate enormous amounts of heat. Processors running at full capacity for hours—or weeks—can reach temperatures that cause immediate hardware failure if not cooled. The most common method is evaporative cooling: water is pumped over cooling coils or directly into the air stream, and as it evaporates, it carries heat away. That water does not return to the system. It is lost to the atmosphere. This is not recycled; it is consumed.
Direct water use is the water that evaporates on site. Indirect use includes water consumed to generate the electricity that powers the facility. A coal or nuclear plant uses substantially more water per kilowatt-hour than a solar or wind farm. So the water footprint of an AI query depends not just on the server's efficiency, but on the local energy mix. A data center in a region with hydroelectric power has a very different indirect footprint than one powered by coal.
There is also water used in the manufacturing of the hardware—the chips, the servers, the cooling systems themselves. This embedded water is part of the upfront cost and is more difficult to attribute per query. But for a full lifecycle assessment, it matters. A single high-end GPU can require several thousand liters of water during fabrication. That cost is amortized over its lifetime, but it is real.
Estimates vary by data center design, local climate, and workload type, but some peer-reviewed and industry data points are available. A 2023 study from researchers at the University of California, Riverside, estimated that training GPT-3 (a 175-billion-parameter model) consumed approximately 700,000 liters of water—enough to fill a small reservoir. That is the training event alone, not the daily inference load. Inference—the act of using the model to answer a query—is much smaller per request but happens billions of times. The same study calculated that a conversation of 20–50 questions with ChatGPT consumes roughly the equivalent of a 500ml bottle of water.
To put that in perspective: if 100 million users each have a single 10-question session per week, that is roughly 500 million liters of water per week. Over a year, that approaches the annual water consumption of a small city. Of course, not all queries are equal: a simple translation uses less compute than generating a long document or an image. Image generation models like Stable Diffusion or DALL-E can be three to five times more water-intensive per output than a typical text response, because the image generation requires more passes through the neural network.
The water footprint also varies by season. In cooler months, data centers can use outside air for cooling, significantly reducing water consumption. In hot, humid climates, evaporative cooling is less efficient, and more water is needed to achieve the same temperature drop. A facility in Arizona will use more water per compute unit than one in Oregon, all else being equal.
Data center location is not just a matter of land cost and connectivity. It is a matter of water stress. Many hyperscale data centers are built in regions with cheap electricity and business-friendly tax policies—regions that are often already water-scarce. For example, data centers in the southwestern United States, including parts of Arizona and Nevada, draw from the Colorado River, a source that is already overallocated. In the Netherlands, data centers have been temporarily banned from expanding due to strain on the local water supply. In Chile, communities have protested data center construction in the Atacama Desert, one of the driest places on Earth, over concerns about groundwater depletion.
The inequity is sharp: a user in a water-rich region queries an AI model running in a water-scarce region, and the environmental cost is borne entirely by the local community. The water that evaporates is not just a resource—it is a resource that could have gone to agriculture, drinking water, or ecosystem maintenance. The AI industry currently has no mechanism to compensate those communities for the water consumed on their behalf.
Not all cooling is equal. Traditional chilled-water systems use a cooling tower and evaporate water. Direct-to-chip liquid cooling circulates a non-evaporative fluid, but the heat must still be rejected somewhere, often through a dry cooler that uses less water but more energy. Immersion cooling—submerging servers in dielectric fluid—eliminates evaporative loss entirely, but the fluid itself requires manufacturing and eventual disposal. There is no free lunch, but the water intensity of each approach varies by at least a factor of three. The industry is slowly moving toward closed-loop systems, but retrofitting existing facilities is expensive and slow.
Many companies report water usage as a single number—liters per megawatt-hour of electricity consumed. This metric is misleading for two reasons. First, it does not distinguish between water withdrawn and water consumed. Water withdrawn includes water returned to the source (e.g., from a once-through cooling system that discharges warm water back to a river), while water consumed is the portion evaporated or otherwise lost. Only the consumed portion represents a true depletion. Second, the per-MWh metric says nothing about the intensity of the workload. A data center running AI models at 90% utilization consumes more water per megawatt-hour of compute than one running idle servers, because the cooling system has to work harder. The only meaningful metric is water consumed per useful AI output (e.g., per query or per training epoch).
Another common mistake is conflating water use with water cost. Water in some regions is artificially cheap, so the financial incentive to conserve is weak. A data center may pay pennies per thousand liters, while the true social and environmental cost is orders of magnitude higher. Companies that claim to be water-neutral by purchasing offsets—such as funding water restoration projects elsewhere—are not actually reducing consumption. They are shifting the burden, often to a different watershed.
If you are a developer, data scientist, or IT decision-maker, you have more leverage than you think. The largest water savings come not from changing the cooling technology, but from reducing unnecessary compute. Here are concrete actions:
Not all water-saving measures are simple. For example, raising the data center temperature setpoint by a few degrees reduces evaporative cooling needs, but it also increases fan energy and may reduce hardware lifespan. The net environmental impact depends on the local energy mix. In a coal-powered region, the increased electricity consumption from fans may cause more water to be consumed at the power plant than was saved at the data center. You need to model the total system, not just the facility.
Another edge case: in some regions, data centers use treated wastewater for cooling. This seems like a win—it uses water that would otherwise be discharged—but it can have unintended consequences. The wastewater discharge permits may be tied to environmental flow requirements in local rivers. Diverting that water to a data center can disrupt those flows. The solution is not simply to use reclaimed water, but to ensure the allocation does not harm the receiving ecosystem.
Small, frequently used models can have a higher total water impact than a single large model if they are poorly optimized. For instance, if a company deploys a separate small model for each of 50 microservices, and each model is hosted independently on underutilized GPUs, the inefficiency can dwarf the per-query water footprint of a single large model shared across services. Aggregation matters.
The biggest blind spot in the industry today is the hidden water cost of model experimentation. Many teams train dozens or hundreds of versions of a model before settling on a final architecture. Each failed training run consumes water that no user ever benefits from. Some teams track GPU hours but not the water that went with them. Implementing a simple pre-training validation pipeline that eliminates clearly bad architectures early can reduce water waste by 40% or more.
At the policy level, there is an emerging debate about whether data centers should be classified as industrial water users or as public utilities. In some jurisdictions, they are currently categorized as commercial, which exempts them from the stricter reporting requirements of industrial users. This classification is outdated—a hyperscale data center can consume as much water as a medium-sized factory. Advocacy for reclassification is a concrete action that engineers and sustainability professionals can support within their organizations.
The water footprint of AI is not an inherent property of intelligence. It is a design choice, and like all design choices, it can be changed. The first step is seeing the water that is currently invisible. The next is deciding what you are willing to do about it. Every query you optimize, every model you shrink, every training run you cancel—that is water you leave in the ground for someone else. That is the real cost, and the real opportunity.
Browse the latest reads across all four sections — published daily.
← Back to BestLifePulse