The Environmental Cost of Artificial Intelligence

AI models such as GPT-4, Gemini, Llama-3 and other large language models have been widely praised for their transformative capabilities and their potential to support climate action. Compared to traditional computing systems, these models have a smarter energy grid, improved climate prediction, and assist in disaster relief efforts, ultimately positioning AI as a powerful ally in the fight against climate change. However, behind their impressive capabilities lies a darker truth: these models are trained and deployed on vast warehouse‑scale data centers that consume immense amounts of electricity and rely heavily on water‑intensive cooling systems.

This article examines a paradox: AI is celebrated as a tool for advancing climate action, yet the infrastructure that powers it quietly drains a vital resource in our world. Water is consumed to keep servers cool, often in regions already facing water stress such as areas in North and East Africa. What are companies doing to address this imbalance, and what must we – as students, institutions, and responsible users of AI — do, expect and demand from the rapidly evolving industry?

Why AI’s Water Footprint is a Growing Concern

Source: Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models, (Li et al., 2023)

How Does AI Actually Work?

AI models such as GPT-4 must operate from data cooling centers, as their high-performance servers generate heat. As a result, these centers require intensive cooling. Cooling occurs in two stages: server‑level and facility‑level. Server‑level cooling uses air or liquid methods and does not typically consume water. Facility‑level cooling, however, often relies on water-intensive systems like cooling towers, where water is evaporated to reject the heat being generated. This process ultimately wastes freshwater, since water in cooling towers can only be recycled a limited number of times before it must be discharged and replaced.

Agriculture, for example, mainly relies on green water, which refers to rainwater stored in soil and available for plants. In contrast, AI’s water footprint depends on blue water, which is water drawn directly from rivers, lakes, or underground aquifers (or natural reservoirs holding groundwater). Blue water is far more limited and competes with human drinking water, sanitation, and industrial use. This distinction matters because blue water depletion directly impacts communities and ecosystems.

Every time a user enters a prompt into a large language model like ChatGPT, servers require cooling that ultimately consumes blue water through evaporation. At a large scale, these seemingly invisible interactions add up: a medium‑length email uses about 3 liters, while even one study estimates that generating a 10‑page report using GPT‑4 could indirectly consume up to 60 liters of freshwater (Shumba et al., 2024). This hidden footprint highlights why AI water use is an emerging environmental concern.

Why Is This a Climate Problem?

When taking into account the indirect and direct consumption of energy, it is estimated that 1 MWh (megawatt-hour) of energy — which is the amount of electricity used by running one million watts for one hour — consumption by a single data center requires about 7.1 m3 of water, depending on the climate and geographical location of the data center. Data centers that use water for cooling can evaporate between 1 and 9 liters of water per kWh (kilowatt-hour), depending on the cooling method, location, and climate. For example, Google’s global average is about 1 L/kWh, while Meta’s average is approximately 3.7 L/kWh (Li et al., 2023). To put this in perspective, an older dishwasher model can use up to around 38 litres of water per cycle, significantly more than modern, efficient dishwashers (Alliance for Water Efficiency, n.d.). This means that running a 3 to 4 kWh AI workload could evaporate about the same amount of water as a single household dishwasher cycle.

AI water usage spans three distinct scopes, which together form the collective water footprint: on-site water for data centre cooling, off-site water for electricity generation, and supply-chain water for hardware manufacturing. 

The first way that AI uses water is through on-site water for data centre cooling. In this scope, this is the data centre using the water itself directly mostly for evaporative cooling towers. For example, this is when water is evaporated to prevent the system from overheating. 

The second scope of AI’s water usage is through off-site water for electricity generation, which simply accounts for indirect water use at power plants that generate the electricity consumed by AI models. Thermal power plants, which includes coal, natural gas, and nuclear energy, require large volumes of water for steam cycles and cooling. This means that even if a data center does not use water locally, its electricity use drives the off-site water consumption. 

Finally, the third scope of AI’s water is the supply-chain water for hardware manufacturing. This includes the embedded water footprint of producing the servers and other data center infrastructure. To put simply, water use comes from making the hardware itself. These chip factories use large amounts of ultra-pure water to clean silicon wafers during production (Li et al., 2024). 

Water used for server cleaning is not a renewable system. At the molecular level, H20 molecules split apart when under immense heat. This stress to the ecosystem–defined as a condition in the environment that disrupts the ecological balance. In general, increased heat in the water reduces the dissolved oxygen available to aquatic life. The depleted oxygen increases the solubility of metals and other toxins in water and pollutants to aquatic organisms (Itua  et al., 2024).

There are more efficient, and greener ways to go about training large AI models–ways in which we will discuss in this article — to limit the amount of water being depleted, and not put marine life at risk. 

What’s Being Done Now: Industry and Policy Responses to AI’s Water Use

Building Greener Infrastructure  

Companies are accountable for the spatial-temporal diversity–also known as deciding the “when” or “where”–to train their AI models. These sites have unexpectedly significant effects on their water footprint. The underlying issue points to companies choosing the wrong when and where to build and operate their data centers for training. Centers that are located on-site are more efficient depending on their outside weather conditions, and off-site water efficiency changes depending on the variation of the grid’s energy fuel mixing to meet time-varying demands.

One particular strategy that companies may follow is attempting to “unfollow” the sun. Training models during peak hours of the day may result in overheating, and requiring more water for cooling (Li et al., 2023). However there is the challenge between “following” or “unfollowing” the sun, as “following” the sun is efficient for carbon efficiency. This is because the sun is required for solar power. Thus, it is encouraged for companies to find common ground between “following” and “unfollowing” the sun. 

Some companies are looking into incorporating models that use less energy. For example, in a recent study looking at data centers in African countries, they developed a model–Llama-3–and compared it with GPT-4. Shumba et al. compared the energy consumption of both models by making them do a couple of tasks: write a 10-page-report, and write a medium-length-email (Shumba et al., 2024). 

The Llama-3 outperformed GPT-4, as it used 52.25Wh (compared to 4.66 kWh, making GPT-4 use 90 times more energy). When it comes to writing a medium length email, Llama-3 used 10 Wh, and GPT-4 used 232 Wh–that’s 23 times more energy (Shumba et al., 2024)! 

What makes Llama-3 efficient is the fact that its smaller, more specialized architecture compared to massive models like GPT-4. Instead of relying on trillion of parameters that demand enormous amounts of electricity and cooling water to operate. Llama is trained to balance its performance to do less work. 

What we can learn from Llama-3 is that bigger is not always better. Models that are optimized for efficiency can accomplish many of the same user tasks while consuming significantly fewer resources. If companies prioritized right-sizing their models to fit specific applications rather than defaulting to the largest model available, then AI can scale more sustainably. This approach demonstrates that climate responsibility and technological innovation and expansions work hand in hand–and not in isolation (Sumba et al. 2024).

Research and Transparency Pushes

Surprisingly, AI’s water footprint stays ambiguous. Model cards routinely inform the public about the model’s carbon emissions, however they omit information on AI’s water consumption (Li et al., 2023). This lack of transparency is a growing concern, as it may prevent efforts to drive greener innovations. Going forward, companies must be transparent as to what their water consumption is.

What More Can Be Done: Strategies for a Sustainable AI Future

For the Tech Industry: Reimaginging Responsible AI 

The tech industry holds a great amount of responsibility for curbing AI’s water footprint. While some companies have begun to publicly report carbon emissions, water usage reporting is far less common (Li et al., 2023). The lack of transparency makes it difficult for the public and policymakers to understand the true environmental cost of adopting, training, and testing AI. Implementing mandatory water usage disclosures, analogous to carbon footprint reports, allows researchers and environmental groups to hold companies accountable and identify which practices (and where specifically) are most harmful.

Another essential step is model efficiency. Large models like GPT‑4 or Gemini demand massive computational resources that are directly linked to heavy water usage. By shifting to smaller, purpose-built AI models for general tasks, companies can serve most users without relying on water-intensive server centers. For example, instead of building models with billions of parameters for simple generation, lightweight AI tools can be used locally or in smaller clusters (Li et al., 2020). 

Large tech companies can report on their water usage–similar to reporting their carbon footprints–and shift towards more efficient AI (with smaller models for general use), and prompt decentralized AI, where local devices are run without depending on massive server farms. 

However, these changes will require institutional support and clear regulatory frameworks to ensure accountability. Just as governments have mandated emissions reporting and renewable energy targets, similar policies could compel AI companies to disclose their water footprints, adopt efficiency standards, and align with broader climate goals.

For Students and Institutions: Shifting the Culture of AI Use

While the tech industry leads on infrastructure, there also exists the demands that shape environmental outcomes. Academic institutions and individual users of AI can foster a culture of conscious AI use. Students and researchers should ask themselves questions like, “Where is this AI model being hosted?” or “How much water is being used to process my request/prompt?”

Limited and unnecessary AI usage is another impactful step. Everyday tasks such as brainstorming essay topics, or summarizing tech jargon, or generating social media posts, can often be completed without invoking a large AI model that consumes liters of water per prompt. In research, traditional methods like literature review or small-scale computations can reduce dependence on cloud-based AI. This particular use of AI mirrors the way that society learned to turn off the lights to save electricity. Small actions multiplied across millions of users can produce meaningful resource savings. 

Education plays a key role in shaping the future of climate-cautious use of AI. AI water waste is not just a tech problem, but also competes with agriculture, drinking water, and marine life. Blue water scarcity has already affected regions like Africa, the U.S. Southwest, and South Asia, where rivers and reservoirs are under immense pressure (Shumba et al., 2024; Siddik et al., 2021). Students should understand that technological innovation must be coupled with ecological awareness and responsibility.

Rethinking “Smart” Technology

To reiterate, AI models such as Gemini and ChatGPT have been praised for their innovation to revolutionize the world, and act as a greener solution to modern technologies. However behind the algorithms and closed doors, what most fail to see is the harsh impact AI has on the climate, particularly water efficiency. To deem AI as “responsible” or “smart” is an understatement, and companies are encouraged to take a second thought to reimagine their models. 

Current strategies implemented by companies can only do so much, and we demand greater efforts. On the other hand, we proposed strategies for tech companies, as well as students and institutions to make smarter choices when utilizing these AI models for personal or academic uses. 

The environment is at risk, and with technologies coming out that do not take into account the existing damage already done from decades of industrialization, the stakes are higher than ever. AI can  not be allowed to repeat the same mistakes that past industries have made, which is prioritizing speed and profiting over sustainability. 

So what–why should this topic, let alone, be debated on? Every innovation that we celebrate in AI comes with a hidden trade‑off. The water that cools servers and powers model training is the same water that fills our drinking glasses, irrigates our crops, and sustains aquatic life. Ignoring the footprints’ risks turning the “smart” technologies into the silent drivers of environmental harm.

As we continue to push the boundaries on what AI can achieve, it’s time to rethink what it means for technology to be “smart”. Models’ intelligence shouldn’t come at the cost of environmental integrity. Rather, these technologies must be redefined to include efficiency, transparency, and climate responsibility.