Artificial Intelligence (AI) is rapidly transforming our world, offering innovative solutions and improvements in various sectors. However, behind this technological revolution lies an often overlooked problem: the water footprint. While attention is rightly focused on the impact of carbon emissions, it is essential to analyze and understand the water consumption associated with AI to ensure sustainable development. This article explores the water footprint of AI, based on scientific data and concrete projections, and offers food for thought for companies and professionals interested in sustainability.
The Global Water Crisis: An Inescapable Context
Before delving into the water footprint of AI, it is essential to recognize the severity of the global water crisis. Freshwater resources are limited and under increasing pressure due to population growth, industrialization, and climate change. Statements from authoritative figures and international organizations emphasize the urgency of this problem:
- Facebook (now Meta): “Water is a finite resource, and every drop counts.”
- Google: “Fresh, clean water is one of the most precious resources on Earth… We are now taking urgent action to support water security and healthy ecosystems.”
- UN Secretary-General António Guterres: “Water is a human right and the common denominator of development for shaping a better future. But water is in serious trouble.”
These statements are not mere slogans but reflect a growing awareness of the need to act quickly to protect water resources.
AI and Water Consumption: A Problem Under the Radar
While the carbon footprint of AI has become a topic of public debate, its water consumption often goes unnoticed. Data centers, which are fundamental for the training and operation of AI models, are large consumers of energy and, consequently, also of water. This water consumption occurs both directly, for cooling the servers, and indirectly, through the production of the necessary electricity. The expansion of AI is leading to an unprecedented increase in water consumption by data centers.
Concrete Data: The Water Consumption of AI
To fully understand the scale of the problem, it is helpful to analyze some data and projections:
- Google consumed 23 billion liters of water to cool its data centers in 2023, of which approximately 80% was drinking water.
- Water consumption in Google data centers increased by 20% from 2021 to 2022 and by 17% from 2022 to 2023.
- Microsoft has recorded similar increases: approximately 34% from 2021 to 2022 and 22% from 2022 to 2023.
- Projections indicate that water consumption by US data centers could double or quadruple by 2028, reaching between 150 and 280 billion liters.
- It is projected that global AI demand will lead to a water withdrawal of between 4.2 and 6.6 billion cubic meters in 2027, an amount equivalent to the annual consumption of half of the United Kingdom or 4-6 times Denmark.
These numbers highlight a worrying trend: the increased demand for AI translates into increased water consumption, with significant implications for sustainability.
Withdrawal vs. Consumption: A Fundamental Distinction
It is essential to distinguish between “water withdrawal” and “water consumption” to assess the real impact of AI on water resources:
- Water Withdrawal: refers to freshwater taken from surface or underground sources for agricultural, industrial, or civil uses. Withdrawal indicates competition for water resources.
- Water Consumption: is defined as “water withdrawal minus water discharge” and indicates the amount of water that is evaporated, transpired, incorporated into products, or removed from the aquatic environment. Water consumption reflects the impact on water availability downstream and is fundamental to assess scarcity at the watershed level.
Both aspects are important for understanding the environmental impact of AI.
The Three “Scopes” of AI Water Consumption
The water consumption of AI can be analyzed in three main scopes:
- Scope-1: the water used directly in data centers for cooling servers, often through evaporative cooling towers. This scope includes the water that evaporates during the cooling process.
- Scope-2: the water consumed indirectly for the production of electricity that powers data centers. Thermal power plants, in particular, are large consumers of water for energy production.
- Scope-3: the water used in the supply chain, including the production of chips and servers. This scope, often underestimated, can be significant but more difficult to quantify.
It is crucial to consider all three scopes to have a complete picture of the water impact of AI. Most of the water footprint of AI consists of “blue water,” taken from rivers, lakes, and aquifers, directly accessible for human use.
Methodology for Calculating the Water Footprint of AI
To quantify the water impact of AI, it is necessary to calculate its water footprint, which includes two main components:
- Operational Footprint: includes both the water used directly on-site for server cooling (Scope-1) and that related to the production of the necessary electricity (Scope-2).
- Embedded Footprint: considers the water used for the production of servers and chips, amortized over the duration of their life cycle.
- Total Footprint: is given by the sum of the operational and embedded footprints.
Water use efficiency (WUE) varies over time and space, depending on climatic conditions and the energy sources used.
Case Study: GPT-3, an Illuminating Example
The GPT-3 language model, developed by OpenAI and used in many AI applications, offers a concrete example of the water consumption associated with AI.
- Training: The complete training of GPT-3 in a Microsoft data center in the United States can consume 5.4 million liters of water, of which 700,000 liters directly on-site (Scope-1).
- Inference: For every 10-50 medium-length responses, GPT-3 consumes the equivalent of a 500ml bottle of water.
This case study highlights how even a single AI operation, such as a simple conversation, can have a significant impact on water consumption.
Recommendations for Sustainable AI
To reduce the water footprint of AI, it is necessary to adopt a holistic approach that considers all aspects of its life cycle. Here are some recommendations:
- Transparency and Reporting: It is essential to include the water footprint (Scope-1, Scope-2, and Scope-3) in model cards and cloud dashboards, similarly to what is done for carbon emissions. This would increase awareness of the water impact and encourage more sustainable practices.
- Consider “When” and “Where”: Plan the training and inference of AI models at times and locations with greater water efficiency. Water Usage Effectiveness (WUE) varies depending on local climatic conditions and the energy mix used for electricity production. Scheduling training at night or in data centers with better WUE can reduce water consumption.
- Balance Carbon and Water: Optimize both energy and water efficiency, adopting a holistic approach. “Following the sun” to reduce carbon emissions by harnessing solar energy can lead to higher water consumption during the hottest hours of the day. It is therefore necessary to find a balance between the two aspects.
- Research and Innovation: Invest in more efficient cooling technologies, such as dry coolers that do not use water, and in methodologies to reduce water use in the production of servers and chips.
- Standardization: Include water consumption as a fundamental metric in international standards for sustainable AI.
Projections for Global AI Water Consumption in 2027
Based on energy consumption estimates of between 85 and 134 TWh, it is projected that global AI will consume between 4.2 and 6.6 billion cubic meters of water (withdrawal) and between 0.38 and 0.6 billion cubic meters of water (consumption) in 2027. These figures are comparable to the water withdrawal of 4-6 Denmarks or half of the United Kingdom.
Conclusions
The water footprint of AI is a critical issue that requires the attention of all stakeholders: companies, researchers, governments, and end-users. A holistic approach is needed that takes into account both energy and water efficiency, with greater transparency and awareness. The water footprint of AI can no longer be ignored.
Next Steps
To effectively address the water impact of AI, it is necessary to:
- Promote research and data collection on the water footprint of AI, particularly for Scope-3.
- Incentivize the adoption of sustainable practices by companies that develop and use AI.
- Raise public awareness about the importance of conscious use of AI, also considering the consumption of water resources.
AI has enormous potential to improve our society, but it is essential that its development is sustainable and does not compromise the availability of precious resources such as water. Collaboration and innovation are essential to achieve this goal.
I hope this article provides useful insights and stimulates debate on the sustainability of AI. If you have any questions or further comments, please feel free to share them.