The Energy Challenges behind the AI Revolution Essay | Pierre Nedellec

AI has exploded in recent years, with ChatGPT igniting enthusiasm for building bigger and better AI models. This advancement relies on large amounts of computing power, which is harnessed by data centres, warehouses of computers which are brilliant for creating AI models, but have many undesirable characteristics, such as large power demand and carbon emissions. At the advent of the age of AI, it is imperative to progress responsibly, prioritising the need for better energy systems, the welfare of the population and net zero goals.

“Improvements of efficiency are no longer enough to meet the demand of data centres”

Data centres are huge consumers of power. They consume 1–2% of the world’s energy [1], and their consumption of energy is set to double before 2030 [2]. Data centres also have the fastest-growing energy demand of any industry, and will account for 10% of the growth in global energy demand leading to 2030 [2]. The energy cost of data centres will be huge, and the present power infrastructure is unsatisfactory. Virginia, US, the world’s leader in data centres [4], is engaging with this issue ahead of many others. The Dominion service provider — which has a near-monopoly in Viriginia — has promised to increase energy production by 150%, import 150% more energy from neighbouring states, and fund a 40% increase to their transmission infrastructure, all before 2030 [3]. This is an extremely ambitious goal and is testament to vast under-preparation for the surge in power demand from data centres. A study by Goldman Sachs shows that while the advances in efficiency in the years 2015–2019 kept the energy demand in data centres quite flat, improvements of efficiency are no longer enough to meet the demand of data centres.

Upgrading the grid is undoubtedly the solution, but this is far from cheap when considering power grids in the USA and Europe are roughly 40 years old. Investment in Europe’s grid, the oldest in the world, is expected to hit 1.6 trillion euros [1]. Nevertheless, governments should be ready to face steep costs, as resolving the energy crisis is fundamental in keeping the country running and allowing data centre construction for advancing AI.

Governments also have the responsibility of overseeing the responsible construction of data centres, which can be a burden on the local population. Data centres are an eyesore for residents, they are sometimes noisy, and they can pollute the air in the unlikely event that the on-site fossil fuel emergency power generators were turned on for an extended period of time. The danger with the widespread demand for building data centres is they will be built as fast as possible, without the proper consideration for the local population.

A power station in Wuhan, China.

Finally and most worryingly, the carbon emissions from data centres has risen in recent years, and threaten to double in the period 2022–2030 [1]. However, these are only rough estimations as most data centre companies do not report energy use and source. This makes it hard for analysts to assign a carbon cost to machine learning research, whereas transparency about the sourcing of energy would incentivise companies to power their data centres from renewables [5].

AI research has the potential to transform all fields of study, and has already been used to create innovative solutions to problems such as making more efficient batteries [6] or predicting the shape of proteins to research new ways to cure diseases [7]. AI can do the same for the energy sector and tackling climate change, which is why it’s imperative that AI research is allowed through the sustainable building of data centres.

Governments may be the decision-makers, but aren’t the only agents of change. As individuals, we have a responsibility to read about developments in data centre construction and how they will impact us and the environment. The quest for better AI models comes at a time where there is a worldwide power shortage [4]. There is an urgent need for innovative solutions to the energy problem, and it is vital that the next generation of researchers collaborate to find a solution that is cost-effective, permanent and is sustainable for the far future. Not only will this enable machine learning research to develop, but such a solution would have a deep societal impact for future generations.

[1] AI is poised to drive 160% increase in data center power demand (2024) Goldman Sachs. Available at: https://www.goldmansachs.com/insights/articles/AI-poised-to-drive-160-increase-in-power-demand (Accessed: 18 May 2025).  

[2] Energy and AI – Analysis - IEA (2025) IEA Energy and AI. Available at: https://www.iea.org/reports/energy-and-ai (Accessed: 18 May 2025). 

[3] WHRO | By Katherine Hafner (2024) ‘unprecedented’ energy demand from data centers poses big challenges for Virginia, commission says, VPM. Available at: https://www.vpm.org/news/2024-12-10/unprecedented-energy-demand-from-data-centers-poses-big-challenges-for-virginia-commission-says (Accessed: 18 May 2025). 

[4] Global Data Center Trends 2024 | CBRE (2024) Global Data Center Trends 2024 Limited Power Availability Drives Rental Rate Growth Worldwide. Available at: https://www.cbre.com/insights/reports/global-data-center-trends-2024 (Accessed: 18 May 2025). 

[5] Patterson, D. et al. (2021) Carbon emissions and large neural network training, arXiv.org. Available at: https://arxiv.org/abs/2104.10350 (Accessed: 18 May 2025). 

[6] Padavic-Callaghan, K. (2024) Ai comes up with battery design that uses 70 per cent less lithium, New Scientist. Available at: https://www.newscientist.com/article/2411374-ai-comes-up-with-battery-design-that-uses-70-per-cent-less-lithium/ (Accessed: 18 May 2025). 

[7] Database, A.P.S. (no date) Alphafold protein structure database, AlphaFold Protein Structure Database. Available at: https://alphafold.ebi.ac.uk/ (Accessed: 18 May 2025).