AI’s Impact on Water Usage in Data Centres

February 27, 2024

Major global tech firms have significantly increased their water usage for cooling data centres, raising alarms about the environmental consequences of the surge in generative artificial intelligence. Companies like Microsoft, Google, and Meta have seen a spike in water use as the demand for their online services grows among millions of users.

The AI Impact on Water Resources

Scholars predict that the need for AI could escalate water withdrawals—taking water from underground or surface sources—to between 4.2 billion and 6.6 billion cubic metres by 2027. This volume is roughly half of the UK’s annual water consumption. A study highlighted in Nature this week by researchers from the University of California, Riverside, emphasises the urgent need to examine and tackle the hidden water footprint of AI models. This comes at a time when the crisis of freshwater scarcity is becoming more acute, exacerbated by prolonged droughts and the rapid deterioration of public water infrastructure.

The Race for AI and Water Use

This issue has escalated in the last year as top technology firms race to launch products powered by generative AI. These products are based on large language models capable of analysing and producing vast quantities of text, numbers, and other forms of data. Operating such models demands enormous computing resources, necessitating large server farms. These farms employ cooling systems that use chilled water to draw heat from the air, with some of the water evaporating during the process and a portion being recyclable.

Water in Energy and Fuel Production

Water consumption is a common requirement across various energy and fuel production methods. For instance, it’s used in extracting oil and gas, generating steam for thermal power plants, and it evaporates from reservoirs used in hydroelectric power generation. For the most recent reporting period in 2022, Microsoft, Google, and Meta saw significant increases in their water usage due to expanded data centre operations, with Microsoft’s water use up by 34%, Google’s by 22%, and Meta’s by 3%.

Corporate Water Replenishment Goals

These corporations aim to replenish more water into natural systems like aquifers than they use by 2030, through initiatives like funding the repair of inefficient irrigation systems or the restoration of wetlands. In the month leading up to OpenAI completing the training of its most sophisticated model yet, GPT-4, a data centre cluster in West Des Moines, Iowa, was reported to use 6% of the local water supply, sparking a lawsuit from the community.

The Call for Transparency and Efficiency

Shaolei Ren, an associate professor at UC Riverside, pointed out that generating between 10 to 50 replies using OpenAI’s ChatGPT, which operates on the earlier GPT-3 model, could consume the equivalent of a 500ml bottle of water, varying by the time and location of its use. Given GPT-4’s larger scale and increased power requirements, Ren indicated it would likely consume even more water, though specific details on its energy consumption remain undisclosed. There’s a call from researchers for greater openness from AI companies, including detailed reports on the water usage of different computing services, like search engines compared to AI applications.

Industry Responses and Future Directions

OpenAI, responding to inquiries, acknowledged the significant water use associated with training large models and highlighted ongoing efforts to enhance efficiency. The company also emphasised the potential of large language models in fostering scientific cooperation and the discovery of climate change solutions. Microsoft has stated that the electricity consumed by AI computing currently represents just a small portion of the overall energy usage of data centres, which in turn accounts for about 1% of the world’s electricity supply. The future impact of AI’s growth on the global effort to achieve net zero emissions will be influenced by various factors.

The Importance of Transparency and Environmental Impact Awareness

Kate Crawford, a research professor at USC Annenberg with expertise in the societal effects of AI, emphasised the need for greater clarity and increased reporting to accurately assess AI models’ environmental impact. She highlighted the urgency of this issue in light of the widespread and prolonged droughts affecting many regions worldwide, coupled with the scarcity of fresh drinking water. Crawford further cautioned against the indiscriminate use of generative AI tools without understanding their real environmental consequences, especially during an ongoing climate crisis.

Avery Fairbank Talent Insight

In the field of Data Centre expertise, operations skills are experiencing the most rapid growth, with an annual surge of 195%. Leading cities for data centre experts are New York, San Francisco, Washington, Dallas-Fort Worth Metroplex, and London. Dell, Cisco, IBM, Amazon Web Services, and Microsoft are the top employers with the highest number of data centre professionals.

Published on 27-02-2024