The Thirst for AI Will Make All Species Thirsty: The Hidden Environmental Cost of Artificial Intelligence
The Digital Revolution Has a Physical Thirst
Artificial intelligence is often portrayed as a technology that exists beyond the physical world. It lives in the cloud, answers questions through a screen, generates images from a few words, writes computer code in seconds, and performs calculations that once required teams of specialists. Its power seems almost weightless. A person types a prompt, a machine produces an answer, and the transaction appears to leave no visible mark on the planet.
But artificial intelligence does not exist in a cloud. It exists in buildings filled with servers, processors, cooling equipment, electrical infrastructure, and industrial supply chains. These systems consume electricity, require minerals and manufactured components, generate heat, and depend on water.
Water is one of the least visible resources behind the AI revolution. While public debate frequently focuses on the electricity required to train increasingly powerful models and the carbon emissions associated with their operation, the water footprint of AI remains comparatively difficult to see. Yet water is essential at several stages of the technology’s life cycle, from manufacturing advanced semiconductors to cooling data centres and generating the electricity that powers them.
This creates a profound environmental contradiction. Artificial intelligence is increasingly promoted as a tool for solving humanity’s greatest challenges, including climate change, agricultural inefficiency, disease, and biodiversity loss. At the same time, the infrastructure supporting its expansion can intensify pressure on the very natural systems that humanity depends upon.
The central concern is captured by the title of this article: the thirst for artificial intelligence could become a source of thirst for people, wildlife, and entire ecosystems.
This does not mean that every AI query directly deprives an animal of drinking water or that every data centre necessarily causes a water crisis. The reality is more complicated. Environmental consequences depend on where computing infrastructure is built, how it is cooled, how its electricity is generated, how much water is consumed rather than merely withdrawn, and whether local ecosystems can withstand the additional demand.
Nevertheless, research has established that AI’s water footprint is real, multifaceted, and capable of growing substantially as computing infrastructure expands. Understanding that footprint is essential if technological progress is to remain compatible with ecological survival.
The Hidden Infrastructure Behind Artificial Intelligence
Every interaction with a large AI system requires physical computation. When a user asks a question, the request travels through a network to computing infrastructure where specialised processors perform mathematical operations. These processors generate heat, and the surrounding systems must remove that heat to maintain reliable performance.
The same process takes place on a much larger scale when companies train advanced AI models. Training involves processing enormous quantities of data, adjusting billions or trillions of numerical parameters, and repeating computational operations across large clusters of machines. After training, the model must continue operating to answer questions, generate content, analyse documents, and serve users around the world.
The facilities that support this work are called data centres. Some resemble enormous industrial warehouses containing racks of computing equipment. Others are highly specialised facilities designed for dense clusters of AI accelerators, with advanced electrical distribution and liquid-cooling systems.
Their environmental footprint extends beyond the electricity consumed by their servers. Facilities may require water for cooling, while power stations supplying their electricity may also consume water. Semiconductor fabrication plants need exceptionally pure water to manufacture advanced chips. Mining, refining, and producing the materials used in computing equipment create additional environmental demands.
Consequently, AI’s water footprint has several interconnected components: direct water consumption at data centres, indirect water consumption associated with electricity generation, and the water embedded in the manufacturing of semiconductors, servers, and other equipment.
A complete assessment must consider these components together rather than examining only the water visible inside a data centre.
The digital economy may appear detached from nature, but its infrastructure is firmly embedded within the physical economy. It requires energy, land, materials, transportation, industrial manufacturing, and water. Its environmental consequences therefore cannot be understood by examining software alone.
Why Artificial Intelligence Needs Water
Modern processors perform billions of operations per second. These operations generate heat, and the more densely computing equipment is packed together, the more challenging heat removal becomes.
Traditional data centres have often relied on air-conditioning systems that circulate cool air through server rooms. However, high-performance AI processors can generate far more heat per unit of floor space than many conventional computing systems. This has increased interest in liquid cooling, advanced heat exchangers, and other methods capable of managing high thermal loads.
Water is particularly effective at transferring heat. Depending on the design, a data centre may use water in cooling towers, evaporative cooling equipment, chilled-water systems, or closed-loop liquid-cooling arrangements.
These technologies differ significantly in their environmental effects.
In an evaporative cooling system, water absorbs heat and some of it evaporates into the atmosphere. The evaporated water is no longer immediately available to the local river, reservoir, or groundwater system from which it was obtained. Additional water may be required to replace losses and remove accumulated minerals from the cooling system.
A closed-loop cooling system, by contrast, circulates liquid through equipment and can substantially reduce ongoing water requirements, although heat must still be rejected somewhere. Depending on the design, the final heat-rejection stage may itself consume water or require additional electricity.
Air-cooled systems can reduce direct water consumption, but they may increase electricity demand under some operating conditions. Dry cooling can also become less effective in very hot weather, precisely when cooling demand is highest.
There is therefore no universally perfect cooling technology. Reducing water consumption requires a broader assessment of electricity use, local climate, equipment performance, and the availability of water.
The central issue is not simply whether a data centre uses water. It is whether its water demand is appropriate for the place in which it operates, whether that demand can be reduced, and whether its benefits justify the pressure placed on local resources.
Understanding Water Withdrawal and Water Consumption
One of the most important distinctions in understanding AI’s environmental impact is the difference between water withdrawal and water consumption.
Water withdrawal refers to water taken from a source, such as a river, lake, reservoir, or underground aquifer. Some of that water may subsequently be returned to the environment.
Water consumption refers to the portion that is not returned to the same water system in the same usable condition and relevant time frame. Evaporation is a prominent example.
This distinction matters because a facility that withdraws a large volume of water and returns most of it may have a different impact from one that consumes a smaller volume through evaporation. The temperature, quality, timing, and location of returned water also matter.
For example, water discharged at a higher temperature can affect aquatic organisms even if the water volume is eventually returned. Industrial discharge can also affect water quality if it contains contaminants or elevated concentrations of dissolved substances.
Conversely, water consumed through evaporation may be especially consequential in a watershed where rivers, wetlands, and groundwater reserves are already under pressure.
A data centre’s annual water consumption figure therefore cannot, by itself, reveal the full ecological significance of its operations. It must be interpreted alongside the source of the water, the local water balance, seasonal conditions, competing demands, and the environmental requirements of the watershed.
This is why water efficiency should not be reduced to a single corporate sustainability statistic. A facility can become more efficient per unit of computing while increasing its total consumption if its computing activity grows sufficiently quickly.
The relevant question is not only how efficiently AI uses water. It is how much water the entire AI industry consumes, where that water comes from, and what would otherwise depend upon it.
The Scale of the Problem: When Small Demands Become a Global Burden
One of the greatest challenges in evaluating AI’s environmental footprint is the difference between the resource consumption of an individual interaction and the cumulative impact of billions of interactions.
A single AI response may require relatively little water compared with many ordinary human activities. But the industry does not operate at the scale of a single question. It operates through vast computing networks serving individuals, businesses, governments, research institutions, and automated software systems around the clock.
The total environmental burden depends on the number of interactions, the computational intensity of each task, the efficiency of the hardware, and the water intensity of the infrastructure supplying the electricity and cooling.
A 2025 study published in Nature Sustainability examined the potential environmental consequences of expanding AI server infrastructure in the United States. Its scenarios estimated annual water footprints of approximately 731 million to 1.125 billion cubic metres by 2030, alongside additional annual carbon emissions of 24 million to 44 million tonnes of carbon dioxide equivalent.
These are model-based projections, not measurements of water already consumed or predictions that one particular outcome is inevitable. They concern AI server deployment in the United States and depend on assumptions about growth, infrastructure, energy supply, and efficiency.
Nevertheless, the magnitude is significant. One billion cubic metres of water equals one trillion litres. The estimates illustrate how an industry built around digital services can develop a water footprint comparable to that of major physical industries.
The same study found that direct water consumption at data centres was only part of the overall picture. Indirect water use associated with electricity generation represented a substantial share of the estimated total.
This distinction changes how the environmental problem should be understood. A company may report low water consumption at its facilities while relying on electricity whose generation requires substantial amounts of water elsewhere.
The water burden does not disappear merely because it occurs outside the company’s property boundary.
The industry’s expansion also creates a rebound effect. As AI systems become faster and cheaper to operate, they become attractive for an increasing number of applications. Businesses begin integrating them into routine communications, customer service, software development, search, advertising, data analysis, and automated decision-making.
Efficiency gains may reduce the resources required for each task, yet total consumption can still rise if the number and complexity of tasks increase faster than efficiency improves.
This is the central arithmetic of AI’s environmental challenge: a smaller footprint per query does not guarantee a smaller footprint for the industry.
Training Is Only the Beginning: The Environmental Cost of Everyday AI
Public discussion often concentrates on the enormous computing resources required to train a new AI model. Training can be an intensive process, involving large clusters of specialised processors operating for extended periods.
However, a model’s environmental footprint does not end when training is complete.
Once deployed, it may answer millions or billions of requests. Every response requires inference, the computational process through which the trained model generates an output.
Inference can occur continuously across multiple data centres as services respond to users around the world. The cumulative water and electricity requirements of these operations can become substantial even when each individual request is relatively inexpensive.
Research by Pengfei Li, Shaolei Ren, and colleagues, published in Communications of the ACM in 2025, examined the water footprint of AI systems across training, operation, and supporting infrastructure. Their analysis used GPT-3 as an example and estimated a total water footprint of approximately 5.4 million litres for training the model in Microsoft’s US data centres, including around 700,000 litres of direct on-site water consumption.
The distinction between direct and total consumption is important. The larger estimate includes water associated with electricity generation, not merely water used inside the data centre.
The researchers also demonstrated why universal claims about the water cost of a single AI prompt should be treated cautiously. Water consumption varies according to the model, prompt length, computing hardware, data-centre efficiency, electricity supply, cooling technology, weather, and time of operation.
A short text response is not environmentally equivalent to generating a long analysis, processing a large document, producing an image, or creating a video. More computationally intensive tasks may require substantially more energy and, depending on the infrastructure, more water.
Consequently, statements such as “every AI question consumes a fixed amount of water” can be misleading. They conceal the enormous variation between workloads and facilities.
The more useful question is how much water a particular service consumes across a clearly defined accounting boundary, and how that consumption changes as usage grows.
The future of AI will be determined not only by the size of the models that companies train, but also by how frequently those models are used, what tasks they perform, and whether those tasks deliver enough social value to justify their environmental cost.
From Data Centres to Rivers: How AI Can Affect Freshwater Ecosystems
Water is not simply an industrial input. It is the foundation of freshwater ecosystems.
Rivers, streams, lakes, wetlands, floodplains, and underground aquifers support extraordinarily diverse communities of life. They provide habitat for fish, amphibians, aquatic insects, molluscs, crustaceans, microorganisms, and countless other organisms. They also sustain terrestrial wildlife by supplying drinking water, food, shelter, breeding grounds, and migration corridors.
When a large industrial facility draws water from a stressed watershed, the consequences depend on local hydrological conditions. If additional demand reduces water available to other users or ecosystems, it can contribute to ecological pressure.
This does not mean that every litre consumed by a data centre produces a measurable decline in wildlife populations. Rather, it means that water demand must be evaluated within the environmental conditions of the source watershed.
A river with abundant seasonal flows may accommodate additional withdrawals more easily than a river already struggling with drought. An aquifer that is naturally replenished rapidly may respond differently from one that takes decades or centuries to recover.
The ecological consequences can emerge through several pathways.
Reduced river flows can shrink sensitive habitats, disconnect shallow stretches, isolate pools, and expose aquatic organisms to higher temperatures and lower oxygen availability. The significance depends on the scale and timing of additional demand. A withdrawal during a period of abundant rainfall may have a different effect from the same withdrawal during a severe drought.
Wetlands act as natural water filters, flood buffers, carbon stores, and breeding grounds for wildlife. Their ecological condition depends on the quantity, timing, and quality of water entering them. If groundwater extraction or surface-water diversion contributes to declining water levels, wetlands can lose important functions. Plants adapted to particular flooding cycles may decline, breeding habitat may disappear, and animals that depend on these environments may be forced to relocate.
Groundwater depletion presents another concern. Aquifers supply drinking water, irrigation, industrial processes, and the base flow that sustains some rivers during dry weather. When groundwater is extracted faster than it is replenished, water levels fall. Wells may need to be drilled deeper, pumping costs may rise, and smaller users may struggle to obtain reliable supplies.
Groundwater depletion can also reduce the flow feeding wetlands and streams. The impact of a data centre therefore depends not only on how much water it uses, but also on whether it relies on a renewable source, a stressed aquifer, treated wastewater, or another supply with different ecological consequences.
Industrial facilities can also affect water quality through wastewater discharge, cooling-system chemicals, concentrated mineral residues, and thermal pollution, depending on their design and operating practices. Poorly managed discharges can damage aquatic habitats. Higher water temperatures may reduce dissolved oxygen and increase stress on temperature-sensitive organisms.
Modern facilities can manage these risks through treatment, monitoring, and regulatory compliance. But effective management requires transparency and enforcement rather than an assumption that technological sophistication automatically guarantees environmental safety.
The broader lesson is that water has ecological functions that cannot always be replaced by money, infrastructure, or promises to restore resources elsewhere.
A healthy river is not merely a pipeline for human consumption. It is a living system with its own water requirements.
The Species That Cannot Compete With Artificial Intelligence
The phrase “all species thirsty” should not be interpreted as a literal prediction that every species will experience water scarcity because of AI. It is a warning about a deeper imbalance in the way modern economies allocate natural resources.
Human institutions can assign an economic value to water consumed by an industrial facility. Companies can negotiate supply agreements, invest in treatment systems, and pay for infrastructure. Households and farmers can sometimes purchase alternative supplies, although many cannot easily afford to do so.
Wildlife has no comparable purchasing power.
A frog cannot negotiate with an industrial operator for a higher groundwater level. A fish cannot pay for the restoration of its breeding habitat. A migrating elephant cannot file a commercial claim when a traditional watering point becomes unreliable.
Their survival depends on the continued functioning of ecological systems that often have no direct representation in commercial decisions.
This is one reason water allocation requires environmental safeguards. Without them, the economic benefits of a new industrial facility can be measured and negotiated more easily than the diffuse ecological costs borne by species that cannot participate in the decision.
The consequences may be indirect rather than immediate. Reduced water availability can alter vegetation, shrink seasonal wetlands, affect the insects that serve as food for birds and fish, or reduce the reliability of wildlife watering sites.
Such effects can cascade through food webs. When one part of an ecosystem changes, the consequences may extend to predators, prey, pollinators, and the communities that depend on them.
However, these outcomes should be evaluated through site-specific ecological evidence. It would be scientifically unjustified to attribute a particular species decline to AI infrastructure without establishing a causal connection.
The valid concern is that industrial water consumption can add pressure to already vulnerable ecosystems, and the AI boom creates a new and potentially expanding source of industrial demand.
In an era of biodiversity loss, the burden of proof should not rest entirely on demonstrating damage after it has occurred. Planning should also establish whether proposed infrastructure can operate without undermining critical ecological water requirements.
The Human Cost: When Digital Convenience Meets Water Insecurity
The environmental consequences of AI water demand are also social consequences.
Water scarcity is rarely distributed equally. Wealthier users and large industrial organisations may have access to storage, alternative supplies, treatment systems, and the financial resources needed to respond to shortages. Poorer households and small farmers may have fewer options.
Where water resources are limited, additional industrial demand can intensify competition among residential users, agriculture, manufacturing, and ecological needs. Whether this happens in practice depends on local supply arrangements, water rights, infrastructure, regulation, and the relative scale of demand.
For communities dependent on groundwater, declining water levels can mean deeper wells, higher electricity costs for pumping, and greater uncertainty about future supplies. For farmers, water insecurity can affect planting decisions, crop yields, and household incomes. For urban residents, it can mean more expensive supply, unreliable delivery, or greater dependence on tanker water.
These consequences are not inevitable outcomes of every data-centre project. They are risks that become more significant when facilities are located in water-stressed regions without adequate planning, monitoring, and enforceable limits.
The issue is especially important because the economic benefits and environmental costs may be separated geographically. A data centre can serve customers across continents while its cooling demand is concentrated in a single watershed.
The people and ecosystems closest to the facility may therefore experience its resource demands even when most of its commercial benefits flow elsewhere.
This raises a fundamental question of environmental justice: who receives the benefits of AI, and who bears the costs of the infrastructure that makes those benefits possible?
A responsible approach to technological development must consider both sides of that equation.
India: The AI Boom Meets a Water-Stressed Nation
India provides a particularly important example of why the relationship between AI and water deserves urgent attention.
The country supports a large population with a relatively small share of global freshwater resources. Its water systems are shaped by monsoon variability, uneven rainfall, groundwater dependence, rising urban demand, agricultural consumption, and the growing effects of extreme heat.
At the same time, India is investing heavily in digital infrastructure. Data centres are expanding to support cloud computing, digital services, financial systems, streaming platforms, and increasingly sophisticated AI workloads.
The expansion of India’s data-centre industry creates an important planning challenge. Industry projections indicate substantial growth in computing capacity over the coming years, with corresponding implications for electricity and water demand. These figures describe the broader data-centre sector, not AI alone, and estimates vary according to assumptions about expansion and resource intensity.
The geographical distribution of this infrastructure is especially important. Major economic and technology centres, including Mumbai, Chennai, Hyderabad, Bengaluru, and the Delhi National Capital Region, are important locations for data-centre development. Some of these regions already face substantial water-management challenges.
The locations attractive to technology companies are not necessarily the locations with the greatest surplus of freshwater.
Land availability, fibre-optic connectivity, reliable electricity, proximity to customers, tax incentives, and access to skilled workers all influence where data centres are built. Water availability must be treated as an equally serious consideration rather than an issue to be addressed after construction.
In water-stressed regions, a new industrial facility can add demand to systems already serving households, farms, hospitals, and other essential users.
The actual effect depends on how the facility obtains water. A centre using appropriately treated wastewater under a robust reuse system presents a different risk from one relying heavily on stressed groundwater. But simply describing a data centre as technologically advanced does not establish that its water consumption is sustainable.
The relevant questions concern actual withdrawals, consumption, source reliability, seasonal demand, and the ecological condition of the surrounding watershed.
India’s water challenge is also complicated by the relationship between surface water and groundwater. A facility drawing groundwater may affect resources that sustain local wells and, in some locations, rivers or wetlands. The degree of impact depends on aquifer characteristics and the cumulative withdrawals of all users.
The consequences can be particularly serious during prolonged dry periods, when rainfall replenishment is limited and competing demand rises.
India has opportunities to develop a more sustainable digital economy from the outset. New facilities can be designed around treated wastewater, efficient heat rejection, closed-loop cooling where appropriate, real-time water monitoring, and careful siting based on watershed conditions.
Planning should also consider the reliability of alternative water supplies. Treated wastewater is not automatically available in the quantity or quality required for industrial cooling. Reuse systems need adequate treatment, distribution infrastructure, and safeguards to ensure that diverting wastewater does not undermine existing beneficial uses.
The goal should be to reduce dependence on freshwater, especially in areas where human and ecological needs are already under pressure.
India’s digital ambitions need not conflict with water security. But achieving both requires governments and companies to recognise that a data centre is not merely an information-technology asset. It is also a major physical facility embedded in a local environmental system.
The Energy–Water Nexus: Why Solving One Problem Can Worsen Another
Water and electricity are deeply interconnected.
Electricity generation can require water for steam production, cooling, and other industrial processes. The amount consumed varies significantly across technologies, plant designs, cooling systems, and operating conditions.
Thermal power plants using coal, gas, or nuclear energy may consume water for cooling. Hydropower changes river flows and reservoir conditions, while evaporation from reservoirs can be substantial in some circumstances. Wind and solar photovoltaic electricity generally require little operational water compared with many thermal generation systems, although their manufacturing and supply chains still have environmental footprints.
Data centres increase electricity demand, and the water consequences of that demand depend partly on the electricity mix supplying them.
This means that a facility can have low direct water consumption while still being associated with substantial indirect water consumption through its power supply.
It also means that water-saving measures cannot be evaluated in isolation.
For example, switching from evaporative cooling to a dry-cooling system may reduce direct water consumption but increase electricity use under certain conditions. If the additional electricity comes from a water-intensive power system, some of the apparent water savings may be offset elsewhere.
The reverse can also occur. Improving server efficiency can reduce both electricity demand and the cooling load, potentially lowering water consumption across multiple stages of the system.
The most effective solutions therefore combine computing efficiency, appropriate cooling technology, low-water electricity generation, and environmentally responsible siting.
Why Carbon Neutrality Is Not Enough
The technology industry has made substantial commitments to renewable energy and emissions reduction. These initiatives can be important, but carbon performance and water performance are not interchangeable.
A power source with relatively low greenhouse-gas emissions can still have meaningful water impacts depending on its technology and location. Conversely, a system with low direct water use may have a higher carbon footprint if it requires more electricity from fossil-fuel generation.
A data centre can therefore improve one environmental indicator while worsening another.
Sustainability claims should account for these trade-offs instead of presenting one favourable metric as proof of overall environmental responsibility.
The objective must be to reduce the combined effects of electricity consumption, greenhouse-gas emissions, water use, pollution, land disturbance, and material extraction.
The Manufacturing Footprint: Water Before the First Prompt
The water footprint of AI begins long before a model answers its first question.
Advanced semiconductors are manufactured through complex industrial processes involving repeated cleaning, rinsing, etching, deposition, and other stages that require highly controlled environments.
Many of these processes rely on ultrapure water. Even trace contaminants can affect semiconductor quality, so manufacturing facilities need sophisticated water-treatment and purification systems.
Producing ultrapure water requires energy and equipment, and the manufacturing process creates wastewater that must be managed. Some water can be recovered and reused, but recovery depends on the contaminants involved, treatment technology, and operational requirements.
The broader hardware supply chain adds further environmental demands. Servers, accelerators, circuit boards, cooling equipment, power systems, and networking hardware all require raw materials and industrial processing.
Mining and refining minerals can affect local water availability and quality, particularly where operations generate polluted runoff or require substantial quantities of water. Manufacturing facilities also consume resources when producing and replacing equipment.
These impacts are not unique to AI. They are shared with other forms of computing and advanced electronics. However, the rapid growth of AI-specific computing capacity can accelerate demand for high-performance chips and supporting equipment.
The environmental accounting problem becomes more serious when companies measure only the water consumed during a model’s final training run or the daily operation of a data centre while excluding the manufacturing and development processes that made the system possible.
A comprehensive life-cycle assessment should include the environmental cost of hardware production, model development, training, inference, maintenance, and eventual equipment replacement.
This matters because computing infrastructure is not static. Hardware becomes outdated, more powerful models are developed, and companies continually expand capacity to meet new demand.
If the industry evaluates sustainability only at the level of the operational facility, a significant portion of its environmental footprint can remain invisible.
The Transparency Problem: Why We Still Do Not Know Enough
One of the greatest obstacles to managing AI’s water footprint is the limited comparability of available information.
Different companies may report different categories of water use, use different measurement boundaries, or publish data at different levels of detail. Some figures describe direct facility consumption, while others incorporate water used in electricity generation or broader supply chains.
Even when figures are accurate within their stated boundaries, comparing them without accounting for these differences can lead to misleading conclusions.
A company operating in a cool climate with an efficient cooling system may report low direct water consumption. Another operating in a hot region may require a different cooling strategy. A third may use electricity generated through a water-intensive process.
Without information about location, cooling design, electricity sourcing, workload, and accounting methodology, it is difficult to determine which facility has the lower overall impact.
Research on data-centre workloads has shown that water consumption per workload can vary enormously across different operating conditions and system configurations. Factors such as server efficiency, electricity-grid water intensity, utilisation, cooling technology, climate, and infrastructure efficiency all influence the result.
That finding undermines the idea that one universal water figure can adequately describe every AI interaction.
Transparency should therefore extend beyond a single annual total. Companies should disclose direct water withdrawal and consumption, separated by source; water used indirectly through electricity generation, with accounting assumptions clearly stated; facility locations and the water-stress conditions of their watersheds; seasonal consumption patterns; cooling technologies and their measured water and energy performance; water reused, recycled, and returned to the environment; and the water footprint of hardware manufacturing and other major supply-chain stages where reasonably measurable.
They should also report progress against targets, including evidence of whether claimed restoration benefits occur in the affected watershed.
Such reporting would help regulators, researchers, investors, communities, and customers distinguish meaningful environmental improvements from broad sustainability claims.
Transparency alone will not conserve water. But without credible information, it is difficult to establish accountability, enforce limits, or determine whether the industry’s environmental performance is actually improving.
Can Artificial Intelligence Help Solve the Water Crisis?
There is an important counterargument to the concern that AI will worsen water scarcity: the technology can also help societies manage water more effectively.
AI can analyse weather forecasts, satellite observations, soil moisture, river flows, groundwater measurements, and water-quality data. When supported by reliable data and appropriate models, these capabilities can improve decision-making in agriculture, utilities, and environmental management.
In agriculture, AI-assisted systems can help estimate crop water requirements, identify irrigation inefficiencies, and predict when fields are likely to experience water stress. These tools may help farmers apply water more precisely, although results depend on sensor quality, local conditions, and whether farmers can act on the recommendations.
In cities, machine-learning systems can help detect leaks, forecast demand, optimise pumping schedules, and identify unusual patterns in water distribution networks.
In rivers and wetlands, AI can support the analysis of satellite images, track changes in land cover, and help identify potential pollution events or ecological changes that warrant investigation.
Weather and hydrological models can also benefit from machine-learning methods, potentially improving some forecasts of rainfall, flooding, drought, and water availability.
These are meaningful opportunities. In regions where water is poorly monitored or infrastructure is inefficient, better information can support more effective resource management.
However, the existence of a useful application does not automatically mean that its environmental benefits exceed its costs.
An AI system designed to save water must be evaluated against the water, electricity, materials, and infrastructure required to build and operate it. Its effectiveness must also be measured against a realistic alternative.
A highly computationally intensive system that produces marginal improvements may not offer the same overall benefit as a simpler sensor network, a better-maintained irrigation system, or a straightforward conservation programme.
Similarly, more efficient irrigation can sometimes encourage farmers to expand irrigated acreage or grow more water-intensive crops. If efficiency gains lead to greater total consumption, the expected water savings may not materialise at the basin level.
This is an example of the rebound effect: improving efficiency can reduce the cost of a resource-intensive activity and encourage more of it.
AI can become a valuable instrument for water conservation, but its contribution must be demonstrated through measurable outcomes rather than assumed from its technological sophistication.
The correct question is not whether AI can help save water. It is whether a particular AI application produces a net environmental benefit, for whom, and under what conditions.
Building an AI Industry That Respects Water
The environmental cost of AI is not inevitable at its current or projected scale. Data centres can be designed more efficiently, electricity can be sourced more responsibly, hardware can be used more effectively, and infrastructure can be located where its resource requirements are compatible with local ecological conditions.
The challenge is to implement these solutions before expanding infrastructure locks communities into long-term patterns of water demand.
First, companies should evaluate the condition of local watersheds, groundwater reserves, seasonal river flows, drought projections, and competing water demands before choosing locations for new facilities. Environmental assessments should examine cumulative demand from existing and proposed facilities rather than evaluating each project independently. Several facilities that appear manageable individually may collectively place excessive pressure on the same water system.
Second, where technically and economically feasible, data centres should use appropriately treated wastewater or other non-potable supplies instead of drinking-quality freshwater. This approach can reduce competition with households and other users, especially in water-stressed regions. However, alternative supplies must be assessed carefully because treatment and distribution require infrastructure, and diverting wastewater can affect rivers or ecosystems that depend on existing discharge flows.
Third, companies should invest in cooling systems suited to their operating conditions. Closed-loop cooling, direct-to-chip liquid cooling, immersion cooling, dry cooling, and hybrid systems offer different opportunities to reduce water use or improve heat management. Their suitability depends on processor density, climate, facility design, electricity costs, maintenance requirements, and the quality of the available water supply.
Companies should compare technologies using measured performance rather than relying on marketing claims. Both water consumption and electricity consumption must be considered.
Fourth, the industry should reduce unnecessary computation. More efficient processors, improved model architectures, quantisation, specialised smaller models, caching, better scheduling, and higher server utilisation can reduce the resources needed to perform a task.
Not every problem requires the largest available model. A smaller system may be adequate for routine classification, summarisation, information extraction, or other relatively narrow tasks. Routing work to an appropriately sized model can reduce unnecessary computation without sacrificing the quality required for the application.
Fifth, AI companies should account for the water implications of their electricity supply, not simply its carbon intensity. Expanding wind and solar photovoltaic generation can help reduce reliance on electricity sources with substantial operational water requirements. Storage, transmission improvements, and better coordination between electricity demand and renewable generation can also help.
Finally, governments should establish transparent water-use reporting requirements, assess industrial demand in the context of local water stress, regulate groundwater extraction, and enforce discharge standards. New facilities should demonstrate that their water requirements can be met without undermining essential household supplies or ecological water needs.
Where water resources are severely constrained, regulators may need to impose restrictions, require alternative supplies, delay projects, or reject proposals that cannot demonstrate acceptable environmental performance.
These decisions should be based on hydrological evidence, cumulative impacts, and enforceable safeguards rather than promises that future technology will eventually resolve present-day constraints.
What Does Water Positivity Really Mean?
Several major technology companies have announced ambitions to become water-positive, generally meaning that they intend to replenish or restore more water than they consume under a defined accounting framework.
Such commitments can support useful projects, including wetland restoration, improved irrigation efficiency, watershed rehabilitation, and investment in community water infrastructure.
But the concept raises an important question: can water consumed in one place be meaningfully compensated by restoring water somewhere else?
The answer depends on the location, timing, quality, and ecological function of the water involved.
A restoration project in a relatively water-abundant region may not compensate for groundwater depletion in a severely stressed basin. Water restored during a wet season may not offset consumption during a critical dry period. A project that improves water availability for one group may also affect other users or ecosystems.
Claims of replenishment therefore need to be evaluated according to where the benefit occurs, whether it is additional to what would have happened anyway, how long it lasts, and whether it can be independently verified.
There is also a fundamental difference between reducing consumption and compensating for consumption.
Reducing water demand prevents some of the original pressure from occurring. Restoration attempts to improve water conditions through other interventions. Both can be valuable, but they are not interchangeable.
A credible water-positive strategy should begin with reducing avoidable consumption, avoiding highly stressed watersheds where possible, protecting existing freshwater ecosystems, and improving the efficiency of essential operations.
Restoration should complement these actions, not serve as permission to expand consumption without limit.
Ultimately, the most meaningful measure of success is not the number of litres a company claims to replenish. It is whether the watersheds affected by its operations are becoming more secure and ecologically resilient.
The Ethical Question: How Much Environmental Cost Is Worth One More AI Capability?
The AI industry often frames technological progress as an almost automatic good. More capable models, faster processing, larger data centres, and broader deployment are presented as evidence of innovation.
But innovation is not the same as necessity, and technical possibility does not automatically establish social value.
Some AI applications may improve medical research, accessibility, scientific discovery, disaster preparedness, or water management. Others may provide relatively minor convenience, automate repetitive content production, or consume substantial computing resources for benefits that are difficult to demonstrate.
These applications should not necessarily be treated alike when evaluating their environmental consequences.
The question is not whether society should abandon artificial intelligence. It is whether the resources required by each application are justified by its actual value, particularly when those resources are limited and unevenly distributed.
Water is indispensable for human survival and ecological functioning. AI can be extraordinarily useful, but it is not an independent biological necessity.
This does not mean that every household or industrial water use must be judged by a single universal hierarchy. Communities need water for sanitation, agriculture, livelihoods, public services, and many other purposes. The allocation of limited supplies requires evidence, democratic accountability, and respect for environmental limits.
However, where an AI facility competes with essential household supplies or threatens critical ecological flows, its economic benefits should not automatically outweigh those concerns.
Technological progress should be evaluated not only by what machines can accomplish, but also by what society sacrifices to make those accomplishments possible.
A civilisation that develops increasingly intelligent systems while degrading the natural systems on which it depends risks confusing computational power with genuine progress.
The Role of Individuals and Businesses
The environmental burden of AI is primarily shaped by infrastructure decisions, industrial design, energy systems, regulation, and the scale of commercial deployment. Individual users should not be made to feel personally responsible for an industry-wide resource challenge.
Nevertheless, individual choices can influence demand at the margins.
Users can avoid unnecessary computationally intensive tasks, choose simpler tools when they are adequate, and use AI where it provides a meaningful benefit rather than generating outputs simply because generation is easy.
Businesses can evaluate whether a large model is necessary for a particular task and request credible environmental information from their technology providers.
Institutions procuring AI services can include energy and water performance in their purchasing criteria. Researchers can develop more efficient models and more consistent measurement methods. Investors can ask companies to disclose environmental risks and demonstrate progress against measurable targets.
The greatest influence, however, lies with the organisations and institutions that shape the system.
AI companies can choose more sustainable locations, disclose resource use, develop efficient infrastructure, and adopt stronger safeguards. Governments can establish clear standards and ensure that the costs of industrial expansion are not unfairly transferred to communities and ecosystems.
Consumers can encourage these changes, but the responsibility for managing large-scale industrial demand must remain proportionate to the power of the actors creating that demand.
A Better Definition of AI Progress
For decades, technological progress has often been measured through speed, capacity, accuracy, cost, and convenience.
These remain important metrics, but they are incomplete when the technology depends on finite resources and can impose environmental costs on people and ecosystems.
A more comprehensive definition of progress would ask several additional questions. Does the technology solve a meaningful problem? Does it use electricity and water efficiently? Does it protect the resources needed by surrounding communities? Does it minimise harm to biodiversity? Are its environmental costs transparently measured? Can its benefits be achieved through a less resource-intensive alternative?
Such questions do not weaken innovation. They help direct innovation toward outcomes that are valuable and sustainable.
The future may bring more efficient processors, improved cooling systems, smaller specialised models, better electricity infrastructure, and stronger water-management technologies. These advances could significantly reduce AI’s environmental footprint.
But technical improvements alone are not enough. They must be accompanied by responsible siting, credible measurement, effective regulation, and decisions about which forms of computing deserve priority.
The goal should not simply be to make every AI system more efficient while allowing its total resource consumption to grow indefinitely.
The goal should be to ensure that the benefits of AI increase without placing unacceptable pressure on the natural systems that sustain life.
Intelligence Must Learn to Respect the Limits of Nature
Artificial intelligence represents one of the most consequential technological developments of our time. It has the potential to accelerate scientific discovery, improve public services, support medical research, strengthen environmental monitoring, and help societies respond to complex challenges.
Yet its digital appearance can obscure the physical resources required to sustain it.
Behind every large computing system lies an infrastructure of electricity generation, industrial manufacturing, heat management, and material supply chains. Water runs through that infrastructure in ways that are often invisible to the people using the technology.
Research has demonstrated that the water footprint of AI can extend well beyond the cooling equipment inside data centres. It includes indirect water use associated with electricity generation and the manufacturing processes that produce advanced computing hardware. As AI infrastructure expands, the cumulative demand can become substantial.
The consequences are not uniform. A data centre in a water-abundant region using efficient equipment and reclaimed water presents a different environmental challenge from one drawing heavily on a stressed aquifer. A computationally intensive application that delivers substantial public benefits may also deserve a different assessment from one that provides little demonstrable value.
Scientific accuracy requires recognising these differences. AI does not automatically cause water scarcity, and the environmental effects of individual facilities must be established through evidence. But uncertainty about the exact footprint is not a reason to ignore the problem. It is a reason to measure it more carefully.
The most important concern is that the world’s freshwater resources are already under pressure from climate change, pollution, urbanisation, agriculture, and industrial development. The expansion of AI adds another demand that must be managed within these existing constraints.
For humans, water insecurity can mean unreliable drinking supplies, rising costs, agricultural losses, and declining public health. For wildlife, it can mean shrinking habitats, disrupted breeding cycles, and reduced access to essential water sources. For ecosystems, it can mean diminished flows, degraded wetlands, and greater difficulty recovering from drought.
These effects should not be exaggerated or attributed to AI without evidence. But neither should the ecological needs of rivers, wetlands, aquifers, and wildlife be overlooked simply because industrial computing offers substantial economic opportunities.
The challenge is not to choose between intelligence and nature. It is to develop intelligence in ways that respect nature’s limits.
That requires a new standard for technological progress: not merely more capable models, but more responsible infrastructure; not merely lower emissions, but lower overall environmental pressure; not merely corporate promises, but transparent measurement and enforceable accountability.
It requires companies to reduce avoidable water consumption, governments to protect essential water supplies and ecological flows, researchers to improve the efficiency of computing, and communities to have a meaningful voice in decisions affecting their resources.
Above all, it requires recognising that water is not an expendable input into the digital economy. It is the foundation of life.
Artificial intelligence may become more powerful, more accessible, and more deeply integrated into everyday existence. But its future should not be measured solely by the number of tasks it can perform or the speed at which it performs them.
It should also be measured by whether the rivers continue to flow, the wetlands continue to thrive, the groundwater remains available, and people and wildlife retain access to the water they need.
The real test of human intelligence is not whether we can build machines that appear to think. It is whether we can build a technological civilisation wise enough to protect the conditions that make life possible.
If the thirst for artificial intelligence is allowed to grow without ecological limits, the costs may reach far beyond the data centre. But if intelligence is guided by responsibility, efficiency, and respect for the natural world, AI can develop without making the survival of other species an acceptable price of progress.
The future of artificial intelligence must therefore be about more than computing power. It must also be about water security, biodiversity, environmental justice, and the shared future of life on Earth.
