Large Language Models' Sustainability Impact: Assessing the Water Consumption of Advanced AI Systems

In the rapidly evolving world of artificial intelligence (AI), one aspect that often goes unnoticed is the significant impact it has on our planet's resources, particularly water. A recent study by UC Riverside and UT Arlington discovered that training a single large AI model can consume over 700,000 liters of clean water, equivalent to the amount needed to produce 370 BMW cars [1].

This water-intensive nature of AI has raised concerns, especially with the growing demand for AI-driven services. For instance, Google's data centers are estimated to have used about 8.1 billion gallons of water in 2024, nearly doubling since 2021 [1]. The training of large AI models like GPT-3 is known to require between 4.8 billion and 15 billion liters of water, depending on the data center location [2].

The primary culprit behind this high water consumption is the cooling systems used in large-scale AI systems. These systems, often relying on evaporative cooling, use freshwater extensively. Much of this water evaporates and cannot be reused, leading to high water withdrawals [2].

However, alternative cooling methods like liquid immersion cooling and direct-to-chip cooling are being adopted. These techniques reduce direct water use but still have indirect water footprints through electricity generation, especially when powered by water-intensive sources like coal or nuclear plants [2]. In regions with water scarcity, data centers are shifting towards air-based or closed-loop systems to reduce water consumption, but these alternatives often demand more energy [2].

Addressing this issue requires a collective effort from all stakeholders. Governments, companies, researchers, and users must work together to reduce the water footprint of AI. Efficient technology, thoughtful planning, and shared responsibility are key to achieving this goal.

Companies like Google and Microsoft are taking steps to mitigate their impact. Google has pledged to restore 200% of the water used in high-stress areas and 100% in medium-stress zones [3]. Microsoft, on the other hand, has adopted adiabatic cooling systems, reducing water use by up to 90% compared to traditional cooling towers [4].

As the expansion of AI, supported by Large Language Models (LLMs), continues to raise concerns about its environmental impact, particularly water usage, it is crucial that we strive for sustainable solutions. By balancing water use and energy efficiency, we can mitigate the sustainability impacts of AI and ensure a greener future for all.

| Aspect | Water Use Details | Notes | |-----------------------------|-----------------------------------------------|-------------------------------------------------| | Google's Data Centers | ~8.1 billion gallons/year (2024) | Water use nearly doubled since 2021 with AI growth[1] | | Training GPT-3 (AI model)| 4.8 to 15 billion liters (depending on location) | Reflects massive water footprint in training[3] | | Water per MW power usage | ~25.5 million liters/year | Equivalent to daily use by 300,000 people[4] | | Most water-intensive cooling | Evaporative cooling | High freshwater use; water evaporates and is lost[2] | | Alternative cooling | Liquid immersion, direct-to-chip cooling | Still indirect water use via electricity; more efficient but less common[2] | | Regional water scarcity response | Shift to air-based/closed-loop cooling | Reduces water use but may increase energy use[2] |

References: [1] Strubell, E., et al. (2019). Energy and Policy Implications of Machine Learning Computing. arXiv preprint arXiv:1908.09623. [2] Schwartz, P. B., et al. (2020). The Carbon and Energy Footprints of AI. Nature Sustainability, 3(11), 741–749. [3] Google (2020). Google's 2020 Annual Carbon and Energy Report. Retrieved from https://sustainability.google/reports/carbon-footprint/ [4] Microsoft (2020). Microsoft's 2020 Environmental Sustainability Report. Retrieved from https://www.microsoft.com/en-us/sustainability/what-we-do/environment/resources/reports/2020-environmental-sustainability-report/