April 14th,  2025

The amount of electricity needed to power the world’s data centres is projected to double by 2030, with artificial intelligence (AI) playing a major role in this dramatic rise. According to a new report from the International Energy Agency (IEA), data centres could consume as much as 945 terawatt-hours (TWh) per year by the end of the decade — roughly three times the UK’s total annual electricity usage.

Why the surge in demand?

The explosion in AI development — particularly the rise of large language models like GPT-4 — is driving the rapid growth in computing power and energy needs. Training advanced models can take weeks and consume vast amounts of electricity. For example, training GPT-4 is estimated to have used around 42 gigawatt-hours (GWh) of electricity, equivalent to the daily power consumption of over 28,000 households in developed countries.

But it’s not just training models that’s energy intensive. Running AI models (a process called “inference”) also requires significant power. Generating a single six-second AI-generated video clip, for instance, uses eight times more electricity than charging a phone — or nearly double the power needed to charge a laptop.

Concentrated pressure on power grids

The demand won’t be evenly spread. It’s expected to be concentrated around major tech and population hubs — especially in the US, Japan, and Southeast Asia — placing huge pressure on energy grids, utilities, and existing infrastructure.

In the US alone, the IEA estimates that data centres will soon consume more electricity than the entire energy-intensive manufacturing sector, including industries like aluminium, steel, and cement.

The paradox of AI and energy

Despite its growing energy footprint, AI also presents a potential solution. The IEA highlights how AI could help engineer smarter, more energy-efficient data centres, improve grid management, and speed up the rollout of cleaner energy sources.

Two key shifts have driven the AI boom:

The cost of compute has fallen by 99% since 2006.

The amount of compute used to train today’s leading AI models has increased by 350,000 times in the last decade.

Can supply keep up?

There are growing concerns about whether power supply can scale quickly enough. In the US, many utility companies have already reported being overwhelmed by energy requests from data companies — with some demands exceeding their current peak supply.

The risk isn’t just local. Global energy supply chains could be disrupted by geopolitical factors, such as high tariffs on China introduced by the US government, which may impact access to raw materials for clean energy infrastructure like solar panels and batteries.

While some, including the US president, are advocating for coal to meet the rising demand, the reality is that coal plants are costly and slow to build, potentially leaving supply lagging far behind demand.

The bottom line

As AI continues to reshape industries and everyday life, it’s also reshaping our global energy needs. Whether the world can balance this demand with sustainability goals, infrastructure capacity, and political realities is a question we’ll be facing head-on in the coming years.