AI’s energy-hungry data centres bring sustainability challenges

Data centres have expanded in line with the growing demand for AI. Now experts are calling for transparency regarding their environmental impact

Artificial intelligence is becoming all-pervasive and is set to grow massively. The market for AI technologies amounted to around $200bn in 2023 and is expected to reach more than $1.8tn by 2030. While others debate its benefits or its dangers, scientists and environmental campaigners want more attention focused on its energy footprint.

AI requires significantly more computing power than conventional internet use, which has led to a boom in the data centres that house the IT infrastructure. AI will be responsible for an 18.5 per cent annual growth in data centre storage capacity requirements between 2023 and 2027, predicts Jones Lang LaSalle, a UK-based real estate and investment management company.

“Almost every forecast [of AI growth] that we see shows the massive extrapolation of data requirements and comput[ing] power and therefore energy consumption,” Matthew Bell, global climate change and sustainability services leader at consultancy EY, tells Sustainable Views.

Data centres and data transition networks are responsible for 1 per cent of annual energy-related greenhouse gas emissions and this figure is set to increase, says the International Energy Agency.

Energy use by data centres increased between 20 and 40 per cent annually between 2010 and 2022, largely driven by growth in blockchain and AI, the IEA says. It predicts AI will have “significant implications for data centre energy use in upcoming years”, with the use of chatbots helping to accelerate growth in energy demand.

Training one large language model – the kind of AI model used by chatbot Chat GPT to generate text – over a six-month period has a carbon footprint of around 300,000kg of CO₂, a 2019 study by the University of Massachusetts Amherst found, a figure equal to 125 round-trip flights between New York and Beijing.

IEA analysis of data from Meta and Google shows that training these models accounts for 20 to 40 per cent of their overall energy use, with 60 to 70 per cent used for “inference” (the use of AI models) and up to 10 per cent for model development.

Impact on local grids

Grid capacity limits and national clean energy targets are leading some governments to restrict the construction of data centres.

Thanks to a corporation tax that is much lower than the EU average, and as home to the European operations of major companies such as Meta and Microsoft, Ireland has seen a rapid increase in data centres. Its Central Statistics Office says the percentage of metered electricity consumed by data centres rose from 5 per cent of the country’s total metered energy consumption in 2015 to 18 per cent in 2022.

Additionally, data centres’ total energy consumption increased by 400 per cent between the start of 2015 and the end of 2022, which the CSO says was “driven by a combination of existing data centres using more electricity and new data centres being added to the grid”.

The IEA predicts that data centres will account for 36 per cent of Ireland’s total electricity demand in 2026, which it attributes to “AI applications penetrating the market at a fast rate”.

In July 2022, an Irish government report said that in the short term there was “limited capacity for further data centre development” due to grid capacity constraints and the need to “rapidly decarbonise” under the country’s carbon budgets. While state-owned EirGrid put a moratorium on the development of new facilities in Dublin until 2028, applications for data centres elsewhere in the country are still being granted.

While Ireland has a lot of data centres relative to its size, Germany has the most in Europe. Its government introduced clean energy targets for operators as part of its energy efficiency law, which came into force in November 2023.

From 2024, data centres in Germany must cover at least 50 per cent of their energy use with renewable energy, and by 2027 this increases to 100 per cent. Under the regulation, they must also reach targets for power usage effectiveness, meaning the energy they consume for uses other than computing (such as lighting) must be limited.

Companies including Microsoft, Amazon and Open AI are exploring the use of on-site nuclear generation to power their centres, with OpenAI CEO Sam Altman investing $500mn into US nuclear fusion company Helion. But these developments have yet to result in generation or be deployed.

The problem of cooling

Besides their constraints on the local grid supply, data centres also require huge amounts of water to keep servers cool. According to a report published in March by the global Climate Action Against Disinformation coalition of NGOs, data centres are contributing to water scarcity.

The report authors cite a study by US university Virginia Tech, which found that at least one fifth of the data centres operating globally do so in areas with moderately to highly water-stressed watersheds, where demand in the region exceeds natural supply. Increased use of AI will result in a “vast increase” in water consumption for data centre cooling, the report adds.

Tech company Google’s global data centre fleet consumed about 4.3bn gallons of water in 2021 – equivalent to London’s total weekly water consumption, though Google says it is looking to use alternatives to freshwater for cooling, such as wastewater or sea water.

A Google spokesperson tells Sustainable Views the company considers the health of local watersheds when planning locations, and evaluates whether using a water source for cooling will impact future supply, potentially leading to water scarcity or drought.

Reducing our use of AI

Michael Khoo, climate disinformation programme director at non-profit Friends of the Earth and author of the CAAD report, says that the way that LLMs, such as online AI chatbots, are being used is driving up their energy use.

Rather than the unlimited public use of chatbots, whereby models are asked to write a “50,000 word Edgar Allan Poe poem,” Khoo advocates for the use of specialised models trained to carry out specific tasks, reducing the energy cost of training.

Training LLMs is “expensive, computationally, and carbon expensive” and having multiple models developed by different companies performing the same task is unsustainable, says Scott Hosking, a natural environment senior research fellow at the Alan Turing Institute, the UK’s national institute for data science and AI.

“When we’ve reached a model that does what we need it to do, then we need to be able to stop,” he tells Sustainable Views. While he agrees the selective use of AI models could help to reduce their footprint, this approach flies in the face of “a temptation in big tech to keep pushing the boundaries of what’s possible”, he says. Khoo also highlights the reluctance of governments to “stand in the way of innovation”.

The need for transparency

However, Khoo insists that regulation to increase transparency on how data centres use energy and water resources “is a basic ask”.

EY’s Bell agrees that transparency will become increasingly important for technology companies and points to the materiality reporting obligations of the Corporate Sustainability Reporting Directive. “If energy is material and […] it sits in your supply chain, you have to report on it,” he says.

He adds that data centre reporting will be covered under the EU’s recently agreed Corporate Sustainability Due Diligence Directive, which requires businesses to report on the environmental and human rights impact of their supply chain.

Besides the regulatory incentive, there is a “competitive advantage” to operators greening their data centres, Bell says, suggesting that as companies look to the impact of their supply chain, they will increasingly want to work with data centres that use renewable energy and consider their overall environmental impact.