← Back to Home

Global Water Consumption by Generative Artificial Intelligence

Technical Report on the Water Impact of Data Centers

Executive Summary

The rapid expansion of generative artificial intelligence carries a significant environmental cost—particularly in the form of water consumption by data centers operated by major tech companies. This report presents a comprehensive analysis of global water usage associated with generative AI, projecting exponential growth in demand for water resources through 2027.

1. Introduction

High-performance servers (GPUs and others) generate substantial heat during the training and inference of large AI models. Cooling these systems requires massive volumes of potable water. Additionally, the electricity generation needed to power these data centers—especially from thermoelectric power plants—leads to large-scale water withdrawal and evaporation.

This document consolidates the most reliable and up-to-date estimates of global water use associated with generative AI, focusing on major players such as Google, Microsoft, OpenAI, and other big tech firms. It also lays the groundwork for a real-time water consumption counter.

2. Analysis of Water Consumption by Major Companies

2.1 Google (Own Data Centers)
In 2022, Google's data centers consumed approximately 5 billion gallons of freshwater (~20 billion liters) for cooling—a 20% increase from 2021. In 2023, the volume rose by another 17%.

According to Google's Sustainability Report, its data centers withdrew around 29 billion liters of water in 2023 for cooling, with over 23 billion liters effectively consumed (evaporated) in the process. This is nearly 80% of the annual water consumption of a major beverage company¹.

2.2 Microsoft (Azure and OpenAI)
Microsoft's water usage surged alongside its AI operations. In 2022, the company reported a 34% increase in water use over 2021, reaching almost 1.7 billion gallons (~6.4 billion liters)³.

Researchers attribute this spike primarily to the rise in generative AI workloads and Microsoft's partnership with OpenAI⁶. Preliminary estimates suggest a further 22% increase in 2023, bringing total usage to about 7.8 billion liters.

2.3 OpenAI
Although OpenAI does not publish its own water usage (since it relies on Microsoft's Azure infrastructure), its impact is embedded within Microsoft's figures. The training of GPT-3, for instance, used around 700,000 liters of freshwater over just two weeks⁴. In less efficient data centers (e.g., in Asia), this figure could triple.

A single ChatGPT session with 10–50 prompts consumes about half a liter of water, accounting for electricity and cooling⁵⁶. In other words, each 20–50 exchanges with an AI model equates to approximately 500 mL of water—varying by server location and season³.

2.4 Other Tech Giants (Amazon, Meta, and Others)
Firms like Amazon (AWS) and Meta also operate large-scale data centers and report growing water usage driven by AI applications. AWS has committed to becoming "water positive" by 2030, aiming to replenish more water than it consumes⁷.

Despite limited transparency, journalistic investigations and industry reports highlight a clear trend: data centers are increasingly "thirsty," and the rise of generative AI is accelerating this trajectory⁷⁶.

3. Global Projections and Estimate Ranges

Academic experts project an exponential increase in global water demand tied to AI. U.S.-based research estimates that, by 2027, AI operations could account for 4.2 to 6.6 billion cubic meters of water withdrawal per year¹⁷.

That translates to 4.2–6.6 trillion liters annually—factoring in both direct usage in cooling systems and indirect consumption via electricity generation⁷.

3.1 Visualizing the Magnitude
One trillion liters equals roughly 400,000 Olympic-sized swimming pools. Thus, 4.2–6.6 trillion liters would represent half the UK's annual water consumption or approximately 1–2 trillion gallons⁸.

3.2 Methodological Considerations
These projections carry uncertainties. They assume continued rapid AI adoption and apply average coefficients of water use per kWh. Some reports count only directly consumed cooling water, while others include the water footprint of the energy supply chain.

In Google's case, 23 out of 29 billion liters withdrawn in 2023 were consumed—i.e., not returned to the environment². This reflects evaporation in chillers and cooling towers. As nearly 80% of Google's water use is potable¹, the local impact on water availability is significant.

4. Real-Time Water Counter Modeling

To visualize ongoing water consumption, the annual projections for 2027 were converted into time-based metrics.

4.1 Calculation Base
Projected annual usage in 2027: ~5 trillion liters
Per day: ~13.7 billion liters
Per hour: ~570 million liters
Per minute: ~9.5 million liters
Per second: ~158,000 liters

4.2 Interpretation
If AI water demand reaches 5 trillion liters/year, over 150,000 liters would be consumed per second—globally—just to sustain AI algorithms. That's equivalent to a large glass of water every millisecond.

4.3 Conservative Estimate (2024/2025)
Summing up recent annual figures:
Totaling over 40 billion liters/year among major players—equating to about 1.27 liters per second. However, this excludes smaller players and especially the water used in electricity generation.

5. Methodology and Data Sources

5.1 Core Assumptions
Market Concentration: Major cloud providers (Google, Microsoft/Azure, Amazon/AWS) account for the majority of water use for generative AI workloads⁶.

Water Use Components: Water Quality: Figures refer to potable or near-potable water, as clean water is essential to prevent corrosion and residue buildup⁵.

5.2 Handling Variations
Given disparities in methodologies, estimate ranges were used instead of fixed values. The 4.2–6.6 billion m³ projection for 2027⁷ is preferred for its plausibility. Corporate reports from Google and Microsoft provide consistent annual increases linked to AI³.

6. Conclusion

Generative AI is already consuming tens of billions of liters annually, and may soon require trillions worldwide. This is a continuous process: every prompt to ChatGPT, every generated image, represents a small "sip" of water on a distant server⁵.

From academic papers to journalistic reports, multiple sources confirm the same trend: the explosion of generative AI—chatbots, code assistants, image and video generators—is dramatically increasing water usage by big tech³.

While exact figures vary, the 4.2–6.6 trillion liter annual range by 2027 is both realistic and perhaps conservative, given accelerating AI growth in 2024–2025. Without changes, researchers warn, AI's "thirst" may double or quadruple the water consumption of data centers in the coming years, exacerbating shortages in already vulnerable regions¹.

References

  1. Li, P. et al. (2023). Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models – arxiv.org
  2. Google Inc. (2023). Environmental Report 2023: Sustainability and Data Centers – Corporate Report
  3. Associated Press (2022–2023). Microsoft Water Usage and AI Infrastructure Reports – apnews.com
  4. UC Riverside (2023). Water Consumption in AI Training: GPT-3 Case Study – news.ucr.edu
  5. Yale Environment 360 (2023). The Hidden Water Cost of Artificial Intelligence – e360.yale.edu
  6. Solveo Research (2023). AI Infrastructure and Water Consumption Analysis – solveo.co
  7. Trellis Sustainability Group (2023). Data Center Water Usage and Future Projections – trellis.net
  8. University of Miami Law Review (2023). Environmental Justice and AI Water Consumption – race-and-social-justice-review.law.miami.edu