AI's Challenging Waters

An article from the Center for Secure Water (C4SW)

Written by Ana Pinheiro Privette

 

Fresh water is a precious and limited resource on our planet. Out of all the Earth's water, only 3% is fresh, and of that, a significant 2.5% is locked away in glaciers and polar ice caps making it inaccessible. This leaves just 0.5% of the planet’s water readily available to meet the needs of a growing global population alongside the rising demands from agriculture and industry. When the demand for freshwater surpasses the available supply needed to meet both human and ecological needs, we experience water stress. According to the UN-Water program, in 2021, approximately 10% of the global population - 720 million people - lived in countries with high and critical water stress levels. 

One additional factor intensifying the water crisis is the rapid growth of Artificial Intelligence (AI). The data centers required to run large AI models consume vast amounts of power. This significant energy consumption generates large amounts of heat, which in turn requires cooling systems to prevent server overheating. Most common cooling methods (e.g., cooling towers) depend on substantial quantities of clean, fresh water.

---

 

Data centers are growing like weeds, and they are thirsty

The rapid growth in the number of data centers—currently estimated at around 11,000 globally—reflects largely the exponential increase in computational demands for AI. The associated surge in resource consumption raises significant concerns about environmental sustainability, particularly regarding water usage.

Water consumption in data centers varies widely. For example, Google's hyperscale data centers, which support major services such as Gmail and Google Drive, averaged approximately 550,000 gallons (2.1 million liters) of water per day over the past year. In contrast, smaller data centers generally report much lower water usage, averaging about 18,000 gallons (68,100 liters) per day.  In the US, where the average per capita water withdrawal is 132 gallons per day, a large data center consumes water equivalent to that of 4200 persons. This makes data centers one of the top 10 of “water-consuming industrial or commercial industries” in the country. The U.S. is home to over 5,300 data centers, and by the end of 2021, around 20% of these centers were drawing water from moderately to highly stressed watersheds in the western U.S. This practice contributes to the growing problem of water scarcity in the region.

AI's water usage can be categorized into three main areas:

1. Cooling (scope-1): On average, depending on the weather conditions and operational settings, data centers can evaporate about 0.26 - 2.4 gallons (1 – 9 liters) per kWh of server energy for cooling purposes.
2. Water-Intensive Electricity Generation (scope-2): The electricity needed to power data centers often comes from thermoelectric or hydroelectric plants, which require significant amounts of water. The national weighted average for thermoelectric and hydroelectric water use is 2.0 gal (7.6 liters) of evaporated water per kWh of electricity consumed.
3. Water Consumption Associated with AI Supply Chains (scope-3). Producing a single microchip, for instance, requires 2.1 – 2.6 gallons (8-10 liters) of water to cool machinery and make sure wafer sheets are free of contaminants.

Difficult to track: water usage in data centers

Over the past decade, ESG (Environmental, Social, and Governance) disclosure requirements have increasingly focused on greenhouse gas emissions, highlighting the urgent need to address data center energy consumption. With increasing pressures for tech companies to reach their net zero goals, a large focus for data center efficiency has been on minimizing energy use and increasing reliance on renewable energy sources. The need to minimize water consumption has received considerably less attention or regulatory incentives.

Tracking water usage for data centers, along with its geographic and temporal distribution, is challenging due to insufficient reporting and transparency. AI model cards—fact sheets that detail how AI models are trained and their intended use—typically include information about the  scope-2 offsite carbon footprint related to energy consumption during model training. However, these cards generally provide little to no information about water usage.

Transparency leads to better resource management

There is, however, increasing pressure to fully understand the environmental impact of AI. Recently, Massachusetts Senator Edward Markey and other lawmakers introduced the Artificial Intelligence Environmental Impacts Act of 2024. This bill mandates that the National Institute for Standards and Technology (NIST) to develop standards for assessing AI’s environmental impact and to create a voluntary reporting framework for AI developers and operators. Similarly, the European Union's (EU) upcoming AI Act will require high-risk AI systems—including powerful models like those behind ChatGPT—to report their energy consumption, resource use, and other impacts throughout their lifecycle. This EU legislation is set to take effect next year.

Greater transparency will enhance resource management for tech companies and empower users to make informed decisions about their AI consumption. It can guide the strategic allocation of AI workloads by leveraging spatial and temporal flexibilities, allowing for more equitable distribution of AI’s environmental impact. For instance, by relocating AI training and deployment to regions with more abundant water resources, it is possible to balance the water footprint across various areas, thus preventing a disproportionate burden on regions already facing water scarcity and drought.

 

Innovating our way into cooler spaces

In the last 20 years, we have witnessed significant advances in data center efficiency. These advancements have reduced energy consumption, improved cooling efficacy, implemented water recycling and conservation practices, and enhanced the overall performance and reliability of data centers. Despite these technological advancements, AI's global annual water consumption is still projected to reach between 4.2 billion and 6.6 billion cubic meters by 2027 (4 to 6 times the annual water usage of a country like Denmark). This is particularly concerning in the context of global freshwater scarcity and the broader challenges posed by climate change.

Geographic location and resource availability are key considerations when designing environmentally sustainable data centers, but the optimization process is further complicated by a changing climate. As climate change intensifies weather extremes, making heat waves and droughts more frequent, data centers are increasingly at risk and are having to consider the business continuity implications of water availability.

As the race for resources increases, many tech companies are strategically positioning their data centers in developing regions, such as Latin America and arid areas of sub-Saharan Africa where real estate and energy are cheaper. This practice further strains local water resources, amplifying the water crisis and contributing to geographic inequality.

Setting environmental guardrails for AI

Addressing the water challenges associated with AI will require significant changes, including:

• Increased Transparency: Reporting the spatial and temporal water usage of AI workloads, makes It easier to better identify opportunities for efficiency and manage resources more effectively. Effective management relies on accurate measurement.
• Responsible AI Use: Balancing carbon and water efficiencies and adopting comprehensive approaches to sustainable AI will help minimize the technology’s environmental impact.
• Technological Innovation: Improving hardware and software, and adopting cutting-edge cooling technologies, can significantly reduce water consumption.
• Climate Change Awareness: Incorporating climate change considerations in data center design and operations may help with operational efficiencies, energy and water consumption, build infrastructure resilience, and meet sustainability commitments.

AI offers substantial potential for tackling critical water-related challenges, such as optimizing agricultural irrigation, improving wastewater treatment, and detecting harmful chemicals in drinking water. However, the full spectrum of risks, societal implications, and potential environmental impacts of AI remains largely unexplored. As this technology advances, it is crucial that we develop environmental safeguards while we address these uncertainties to ensure accountable development and deployment. This is not just the responsibility of tech companies, but that of our collective society as users of the technology.