Energy demand to supply AI data centres in the UK could triple in the next five years, according to industry experts – placing huge strain on existing infrastructure
Newly published minutes from a meeting of the government’s AI Energy Council reveal some alarming details about the huge energy demands of data centres.
Photo by Michal Pech / Unsplash
While there’s a widespread enthusiasm – or hype – for the potential of AI to improve public services and save time and cost, there remains serious questions about the energy required to power AI, the pressures this will place on wider infrastructure and how much it will undermine efforts to decarbonise, achieve Net Zero et cetera.
The AI industry is notoriously guarded about such matters as how much energy is needed to power AI systems that, for example, answer questions, analyse data or generate images and video.
The AI Energy Council meeting held on 30 June, the minutes of which were published this month, boasted a presentation by Thomas Spencer, an analyst at the International Energy Agency (IEA). This included the prediction that global electricity demand for data centres will double by 2030 and nearly triple by 2035 – when it will be roughly equivalent to the total electricity demand of Japan.
The increase is not merely due to more people using AI and more AI systems being launched. Also critical to the increase, said the IEA presentation, is the use of more power-intensive, accelerated servers optimised for AI.
The IEA presentation was followed by one from Fintan Slye, CEO of the National Energy System Operator (NESO), and Sanjeet Sanghera, NESO’s Head of Strategic Futures. This focused on the situation in the UK, where they suggested that energy demand from data centres could triple by 2030, when it would comprise some 7% of all domestic demand.
This presentation also noted that grid capacity varies by region, with the south-east of England heavily congested and the north of Scotland more favourable to new grid connections. It concluded that, across industry, there is a vital need to better understand grid constraints, latency trade-offs and the challenges in siting new data centres.
In the discussion that followed, further challenges to investment in data centres were noted, such as high wholesale prices for electricity and the long lead teams for new grid connections.
The minutes of the June meeting were published on the .gov.uk website by the Department for Science, Innovation and Technology (DSIT) and Department for Energy Security & Net Zero (DESNZ). It’s a concise document, listing the Ministers, industry leaders and government officials who attended, and summarising the main points of discussion.
Others have attempted to calculate the energy demands of AI based on limited available figures. For example, in May, MIT Technology Review worked from Open AI’s claim in December 2024 that its popular ChatGPT received 1bn queries from users per day. This figure was applied to research company Epoch AI’s estimate that responding to each query requires some 1,080 joules of energy, or 0.3 watt-hours.
On that basis, a year’s worth of queries to ChatGPT would require more than 109 gigawatt-hours of electricity per year, which the authors of the piece say is equivalent to heating 10,400 US homes over the same period.
The authors also add that the Epoch AI figure is considered too low by some experts, while the total figure they give is only for query messages to ChatGPT and does not include the energy needed to generate images or video, let alone to serve the AI products of other companies.
At the end of the AI Energy Council meeting June, there was general agreement of the need for a credible and co-ordinated approach to forecasting and investment planning. That will surely require more openness from AI companies about the energy requirements of this technology, now and as it develops further.
In related news:
London study exposes big obstacles to implementing artificial intelligence in NHS
Leave a Reply