AI and Water Usage

 Artificial Intelligence (AI) refers to the simulation of human intelligence in machines that are programmed to think and act like humans. These machines can learn, reason, solve problems, and make decisions, often performing tasks that typically require human intelligence with the help of data centers which provide them the infrastructure and computational power to train and operate on. AI has turned into a business imperative. IBM reports that 82% of large enterprises have either deployed AI or are experimenting with doing so.

Generative AI models continue to grow larger not only in terms of unique innovation but also both users and creators. This means that with each passing day, there are more models being trained and more models being used. Gartner (an IT service manager) stated, "Clearly, the AI revolution is here to stay. By the end of this year, 90% of companies are expected to have adopted AI technologies. However, in their race to innovate, most have overlooked the significant environmental consequences that accompany these decisions."

Unchecked water usage by power-hungry data centers could exacerbate the already dire water crises worldwide. The carbon footprint of these AI models has rightfully received public and academic attention, but even less discussed is the water footprint of these models. Google, Microsoft, and Meta have all pledged to reach at least net-zero carbon emissions by 2030. Amazon has set its net-zero deadline for 2040. All four companies have also pledged to be water-positive by 2030, meaning they'd put more water back into the environment than they use. However, many scientists do not believe that this is possible under the current infrastructure investment plans.

Water covers 70% of the Earth, and it is our most essential ingredient to survive, whereas freshwater, which we need to drink and irrigate our farms only 3% of the world’s water, and over two-thirds of that is tucked away in frozen glaciers and unavailable for consumption. The United Nations Environmental Report states that nearly two-thirds of our world's population experiences severe water shortages for at least one month a year, and by 2030, this gap is predicted to become much worse, with almost half of the world's population facing severe water stress. The World Wildlife Federation projects that 66% of the global population is likely to face water scarcity by next year. When looking at the large numbers of water footprints due to AI models, we really begin to question if innovations like these actually do more harm than good.

• GPT-3 (OpenAI’s now outdated and surpassed model from 2020) training evaporated 700,000 liters of fresh water.
• Google’s data centers in 2023 withdrew 29 billion liters of fresh water for on-site cooling.
• The global AI demand in 2027 is projected to account for 4.2–6.6 trillion liters of water.
• GPT-3 training accounted for 5 people’s yearly water consumption (1823 days of an individual’s water consumption).
• Google’s data centers in 2023 accounted for 206,906 people’s yearly water consumption (75,520,833 days).
• The global AI demand in 2027 is projected to account for 30–47 million people’s yearly water consumption (about the population of Canada).

It might be difficult to conceptualize the impacts of these AI models. AI uses water because the data centers that power these models generate electricity, and that produces heat. In order to cool that heat, they use water. The massive water consumption by data centers stems from their need to maintain optimal temperatures for densely packed servers and computing hardware racks, as they generate immense amounts of heat. Overheating can cause system failures, data corruption, and costly downtime. Data centers also use water for humidification systems that maintain a specific humidity range to ensure the functionality of all the equipment in the building.

Nowadays due to escalating demands for generative AI products companies are having to spend a lot of water supply to cool down data centers. As the demand for water increases due to growing populations, climate change, or industrial needs data centers (which use a lot of water to cool their servers) might find it hard to get enough water to keep operating normally. In drought-prone regions, some facilities may even face tough choices between shutting down operations or overtaxing municipal water supplies that communities rely on.

The environmental water footprint of AI can be divided into three areas:

• involves direct water consumption at AI data centers, mainly for cooling systems.
• includes water used off-site to generate the electricity that powers these centers.
• Supply chain water is used in producing and transporting servers.

--image

In the figure above, the process of AI-related water usage is illustrated. The substantial heat produced during operation is managed by cooling towers that evaporate water to remove the excess heat. While some of the water can be reused, it can only go through a limited number of cycles before it becomes unusable. On average, about 80% of the water used is consumed, meaning it evaporates and is not returned to the source. Importantly, this water is typically clean and often potable, as using untreated water risks clogging pipes or encouraging bacterial growth.

Although not shown in the diagram, an alternative method used in milder climates involves air-based cooling systems that don’t rely on water evaporation. However, when air temperatures rise above 85°F (29°C) or become too dry, water is still introduced to help maintain effective cooling.

Training of these models takes the majority of these energy and cooling demands, but each individual model has its own associated water costs as well. GPT-3 is estimated to consume the equivalent of a 500ml bottle of water for every 10 to 50 medium-sized interactions. The exact amount depends heavily on geographic location. For instance, in Illinois or Iowa, about 33 responses are needed to use 500ml of water, whereas in hotter or drier regions like Arizona or Sweden, just 17 responses can consume the same amount. Interestingly, Ireland emerges as the most water-efficient, requiring around 70 responses to reach that same level of consumption.

The time of day also affects water efficiency. Cooling at night—when ambient temperatures are lower—is more water-efficient. However, this presents a trade-off: carbon emissions can be minimized during the day when solar energy is available. These competing demands highlight the urgent need for smarter, more sustainable cooling technologies that can optimize for both water usage and carbon reduction simultaneously.

Water is a precious, limited resource, and companies have a responsibility to adopt sustainable water management practices. Veolia Water Technologies has come up with a solution called smart water treatment that helped data centers drastically reduce water use and operational costs.

Two data centers partnered with Veolia to improve water efficiency:

• Illinois: By replacing its alkaline system with a sulfuric acid treatment and installing a TrueSense RSG controller, the center doubled cooling tower cycles, cut water use by 50%, and saved 12 million gallons and $150,000 annually.
• Virginia: A high-cost water softening system was replaced with a pH-controlled treatment using TrueSense, eliminating softeners and saving $114,000 per year.