A new industrial revolution is underway, with companies and countries competing for dominance in artificial intelligence. Rather than factories and coal, this race needs data centres and electricity. Data centres are the backbone of the internet, delivering everything from search results to video streaming, and now increasingly they crunch answers to questions posed to AI chat tools. Worldwide, data centres consume more power than the UK and this is set to more than double by 2030. The Government recently commissioned the AI Opportunities Action Plan which calls for drastic action to boost the UK’s AI and computing capabilities. But how will scaling up the number of servers and power-hungry computer chips impact electricity demand?
The Government agreed to take forward all recommendations from the Action Plan, including to expand the UK’s sovereign compute capacity by at least 20x by 2030. This requires rolling out infrastructure UK-wide and setting up AI Growth Zones with fast-track access to the power network and planning approvals. This is critical to the growth of the AI industry as it can take years for new grid connections to be approved. A new AI Energy Council will be appointed to assess energy demands and accelerate investment in clean energy for data centres.
Globally the world’s data centres consumed around 500 TWh of electricity in 2024, overtaking British electricity demand in 2021. Forecasts see continued growth at 10–20% per year until the end of this decade. Since the release of ChatGPT in 2022, investment in generative AI has caused a surge in data centre energy consumption. Global power demand from AI increased by three times between 2023 and 2024 and is forecast to overtake total demand in Britain by 2030.
Global electricity consumption from data centres and AI models, compared to Britain’s total electricity demand. Historical data and forecasts aggregated from BNEF, Goldman Sachs, McKinsey, IEA, and NESO.
The huge cost of training AI models has made headlines recently. OpenAI spent over $100 million training ChatGPT 4 and Elon Musk plans to spend $3–4 billion on training xAI. However, electricity forms only a small portion of this cost. Current models require 300–1300 MWh per training run, costing around $25,000–100,000. The next suite of models could feature 100 times more parameters, increasing costs, but new model designs may counter this. DeepSeek, a Chinese startup, reportedly trained a leading model for just $6 million.
The energy required to manufacture a car is small relative to the fuel consumed over their lifetime. The same is true of AI models, which are trained once, and then run millions of times to answer questions, generate videos and the like. It is their usage (known as inference) which will stretch power grids. Answering a generative AI prompt request consumes ten times more electricity than a standard internet search. ChatGPT cost around $0.36 per query in 2023, or $700,000 per day. The cost of inference is falling, but usage will rise as AI becomes cheaper, a phenomenon known as Jevon’s paradox. As people want fast responses, inference must happen close to users, so data centres will be needed across the UK, not concentrated overseas in countries with cheaper power.
European countries use 1–5% of their electricity powering data centres, and this will grow quickly as AI servers move onto home soil. Ireland already houses data centres for large technology companies like Apple and Google and uses nearly one-fifth of its electricity powering them. The Irish system operator has imposed a moratorium on new data centres in Dublin until 2028. With the UK’s compute capacity set to increase 20-fold by 2030, electricity demand will surge. Data centres require a 24/7 supply of electricity and so firm generation and spare grid capacity will be needed, and new capacity must be clean to ensure that AI does not drive up emissions.
Annual electricity consumption from data centres in Europe’s ten largest markets in 2022 (latest year available), and the share of each country’s total national demand.
