Data centres ‘straining water resources’ as AI swells.

However, “pressure on existing limited water resources [from data centres] may have significantly strained the provision of adequate clean water,” she added.

“The establishment of AI and data centres in Nigeria can facilitate technological development, foster collaboration with global tech companies, and generally improve economic growth,” said Ajijola.

“However, it can exacerbate the strain on the already limited water supplies and potentially impact the people’s access to water.”

AI’s ‘water footprint’

Conventional data centres already use large amounts of water for their cooling processes. But these may need to increase their cooling capacity to support the vast computing resources and storage requirements needed to support new AI applications.

Kaveh Madani, director of the UN University Institute for Water, Environment and Health and ex-vice-president of the UN Environment Assembly Bureau, explained: “AI requires high-performance processes and that results in more electricity and water use in the case of AI data centres when compared with conventional data centers.”

Dedicated AI data centres are springing up, mostly in the US, but increasingly also in the global South, where water resources are stretched.

Stanford University’s AI Index Report 2023 revealed that the environmental impact of training AI models is unprecedented as the adoption of AI skyrockets. The proportion of companies using AI more than doubled between 2017 and 2022, according to the report.

Madani said AI experts were becoming increasingly aware of AI’s environmental footprint.

“The public and the decision-makers must know about such footprints to proactively develop […] footprint-reduction measures,” he told SciDev.Net.

Diagram showing: AI water consumption

AI systems require a lot of data to train and operate. This data is stored in data centres, which are large facilities that house computer systems and other equipment.

The data centres are the backbone of the digital world, providing the computing power and storage space that AI systems need to function. They power our online activities, such as social media, online shopping, streaming services, online gaming and video conferencing.

Even after adopting best practices, machine learning takes up around 10 to 15 per cent of the total data centre energy at Google, according to a paper by computer scientists at the US tech company.

Water consumption at the firm’s data centres increased more than 60 per cent in the last four years, from 3,412 million gallons in 2019 to 5,565 million gallons in 2022.

Shaolei Ren, associate professor in the department of electrical and computer engineering at the University of California, Riverside, says the carbon footprint of AI models is usually documented, but “the water footprint is missing”.

“Once we have the visibility of AI’s water footprint, we can leverage the scheduling flexibility of AI to reduce it,” he told SciDev.Net.

Inside a data centre.CC0 1.0 DEED

Inside a data centre. Source: pixhere (CC0 1.0 DEED)

Why so thirsty?

Data centres use water in two ways: direct and indirect. Data centres require vast amounts of electricity for their operations and electricity generation requires significant amounts of water. In addition to this indirect impact via electricity use, data centres require considerable amounts of water for cooling.

Power plants use water to generate electricity by heating water to create steam, which then turns a turbine to generate electricity. The water used in this process is first treated to remove any contaminants that could damage the equipment.

After the water is used to create steam, it may be evaporated, discharged as effluent, or recirculated.

Data centres also use water to cool the equipment by circulating water through it, which absorbs heat. The water is then cooled and recirculated.

The amount of water used varies depending on the type of power plant and from region to region, as the water consumption depends on the weather conditions and cooling technology used.

Computer scientists interviewed by SciDev.Net say these local conditions are often not factored into decisions on where to locate data centres or what cooling systems to use.

“We need to consider the local constraints and priorities before deciding the most suitable cooling method,” Ren added.

“We need to consider not only the direct water consumption but also the indirect water consumption for electricity generation to holistically reduce the total water footprint.”

According to a recent preprint (a draft research paper that has not been peer reviewed) published by Ren and his colleagues, training GPT-3 in Microsoft’s state-of-the-art US data centres can directly consume a total of 700,000 litres of clean freshwater. They say the language prediction model needs to “drink” a 500ml bottle of water to have a simple conversation of about 20- to 50 questions and answers per user.

If this training was done in Microsoft’s Asian data centres, it would consume three times as much, according to the paper. This is because the climate in Asia is warmer and drier, which means that more water is needed to cool the data centres.

“This is extremely concerning, as freshwater scarcity has become one of the most pressing challenges shared by all of us in the wake of the rapidly growing population, depleting water resources, and aging water infrastructure,” the paper says.

Sub-Saharan Africa is one of several regions facing a water crisis, with water demand expected to almost quadruple by 2030. Yet it is becoming a hub for data centre investments.

The demand for data centres on the continent has doubled since 2016, according to the International Telecommunications Union, and it is expected to attract US$5 billion of investment by 2026, potentially putting further strain on water resources.

Lack of regulation

“This sector is not properly regulated in both the global North and global South,” said Madani from UN University Institute for Water, adding that stronger regulations, monitoring and reporting are “essential”.

“In addition, service users must become aware of the impacts of their ‘AI consumption’ on water resources,” he said.