AI and energy consumption

I have already mentioned in these pages how excessive the fears about the energy consumption of artificial intelligence (AI) often are. Let us look at the issue more closely.

The main bogeyman is data centres. What are they? A data centre is a facility that houses computer systems used to process, store and distribute data and digital services. Today, about two thirds of the electricity consumed by the world’s data centres is not attributable to AI.

We all use data centres constantly, and we did so long before AI arrived. A huge share of our digital activity at work passes through these infrastructures – think of logistics and banking services – as does practically every aspect of our connected personal lives. We use data centres when we send messages on WhatsApp, browse and comment on social media, stream a film or a football match, book an appointment on our health record, read newspapers and books online, play games against a stranger on the other side of the world or listen to music on digital platforms.

Data centres also work massively for a vast range of other activities: from e-commerce to web searches, from cloud services to cryptocurrencies, from pornography (the biggest sites in the sector consistently rank among the most visited on the planet) to the activities of hackers who destabilise the gigantic digital infrastructure I have briefly described.

The International Energy Agency (IEA) estimates that data centres absorbed about 485 TWh in 2025, equal to 1.5% of the world’s electricity. About a third – so around 0.5% of the global total – is attributable to AI-focused data centres. For 2030 the IEA forecasts almost a doubling, to about 950 TWh, just under 3% of global electricity demand: not exactly an apocalypse. The consumption of AI-dedicated data centres is expected to triple by the end of the decade, becoming the main driver of this growth.

Beyond 2030, however, any forecast becomes highly uncertain, because at least two decisive variables remain undefined. The first is the actual size the AI sector will reach: after a few years of explosive growth, we do not know how many companies, applications and business models will manage to consolidate, and what demand for computing capacity will follow. The second variable is the pace at which the energy efficiency of hardware, software and infrastructure will improve. A look at the past can help here. About fifteen years ago, despite the explosive growth of internet traffic and the cloud, electricity consumption was far more contained than initially feared, thanks to very strong efficiency gains across the system and the shift towards larger data centres. It is by no means ruled out that, at least in part, history will repeat itself.

Our “energy fear” about data centres is largely borrowed from the United States, where the concerns are better founded. The American reality, however, is very different from the European one. The world’s main giants of the entire value chain are based in the US: OpenAI, Anthropic, xAI, Google, Meta, Amazon, Microsoft, Nvidia. The very rapid growth of physical infrastructure reflects these companies’ race not to lose ground in a market marked by fierce competition. Politics pushes in the same direction: Washington considers it strategic not to cede ground to China, whose AI industry is increasingly advanced.

The result is that in the United States data centres already absorb about 5% of national electricity demand, with enormous territorial differences: in some states, such as North Dakota, Nebraska and Iowa, the share exceeds 10%, while in Virginia, the country’s main hub, it reaches as much as 25%.

These numbers highlight the central point: data centre consumption is above all a problem of geographical concentration. At national level their weight remains modest, with the striking exception of Ireland, where geographical position and taxation have fostered a boom in data centres that now absorb over 20% of electricity demand. It is at local level that the issue can become significant. Where these infrastructures cluster, electricity demand can grow very rapidly, requiring new generation plants, grid upgrades and investment in connections. In Italy, the problem is concentrated mainly in the Milan area.

Meanwhile, in the United States there are growing calls for large data centre projects not merely to ask the grid for new capacity, but to contribute directly to the costs and, in some cases, to the availability of the energy needed to power them. This opens a new chapter: social acceptability. In several American counties, strong local opposition has already arisen against new data centres, and against the power lines and power stations built to supply them. It is a theme destined to enter election campaigns with force.

Before AI, concern about the energy consumption of data centres was almost absent from public debate. It is positive that attention is growing, but let us not lose sight of the loads that, in the longer run, will weigh far more on electricity production and infrastructure. To increase the efficiency of the system and reduce dependence on fossil fuels, the energy transition requires us to electrify transport, heating and industry as much as possible. In the meantime, we must also equip more and more buildings around the world – starting with central and northern Europe – with cooling systems, driven by ever-rising temperatures. This will mean enormous peaks in electricity demand, for a growing number of days every year. In other words, the risk in the current debate on AI is to overestimate what is concentrated, “fashionable” and visible, while ignoring the widespread consumption that will weigh far more.

In this climate, I occasionally hear people worried about the energy used by bots such as ChatGPT or Claude. By the numbers, though, this concern also looks excessive. According to the latest estimates, a typical text request consumes about 0.3 Wh, including processors and infrastructure. One hundred questions – a figure for a very intensive user – are therefore worth about 30 Wh. That is less than 5% of the consumption of a single 60 °C wash in a washing machine: very little, especially considering the enormous amount of information you can obtain from one hundred well-formulated questions.

AI consumption, of course, does not end there. Before use comes training, that is, the training of models on enormous amounts of data, which is also energy-hungry. But neither training nor text requests can explain the roughly 155 TWh consumed in 2025 by AI-focused data centres. Even assuming 10 billion questions a day at 1 Wh each, the total would be under 4 TWh a year.

Where does the rest go, then? We do not know: AI companies do not publish detailed data on where their consumption goes. Among the most energy-intensive activities are image and video generation, advanced reasoning systems and autonomous agents. But a probably very significant share comes from the pervasive and hidden use of AI inside the network itself: search engines, social media, advertising, translation, summaries, content moderation, online shopping, customer service and business automation. In other words: much of the AI work (and consumption) happens in the shadows, without us even noticing. So much for the guilt over our little questions to Claude.

In the end, all these considerations take us from numbers to more complex and uncomfortable questions, starting with individual responsibility. What is my personal contribution to the growth of data centre consumption? Is it really necessary to keep sending images and files on WhatsApp, or to stay for hours on social media scrolling through reels selected for us by AI?

If we have the clarity to ask ourselves these questions, we can take a further step and reflect more carefully on the powerful tool that has already entered our lives, often without our noticing. How much power are we handing to a few gigantic private companies, without particular controls? What does it mean to delegate more and more activities to machines with no consciousness, experience or feelings, which we (wrongly) consider “intelligent”? It is a question that Pope Leo XIV has posed forcefully in his first encyclical.

Ultimately, among the many open questions about AI, energy consumption has become the most discussed problem, without being the most relevant.

This article is also published on Sapere. illuminem Voices is a democratic space presenting the opinions of leading Sustainability Thought Leaders, their views do not necessarily represent those of illuminem.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.