AI models are out on an energy-intensive training session with no end in sight. The training takes place on the servers in the world’s data centres, which currently number just over 10,000. Especially the large language models and generative AI that creates images and videos consume huge amounts of electricity.

They are so voracious that the International Energy Agency (IEA) has estimated that the power they needed for computing increased a billion-fold from 2022 to 2024.

The entire global energy sector is now changing because the demand for electricity to run and cool servers is so high.

“What we see happening now in the development of artificial intelligence is truly extraordinary and perhaps a pivotal point in human history,” said Sebastien Gros.

He is a professor at the Department of Engineering Cybernetics at the Norwegian University of Science and Technology (NTNU) and head of the Norwegian Centre on AI-Decisions (AID). This is one of Norway’s six new centres for research on AI. Gros highlights the fact that developments in the AI universe are driven by just a few technology companies.

The scale, electricity consumption, investments and pace are formidable, with control concentrated in the hands of just a few private enterprises.

Large, undisclosed figures

We are not able to find out exactly how much electricity AI providers use.

“The companies that supply electricity to the data centres do not disclose these figures. The AI providers are commercial operators whose aim is to make money, and they have little interest in sharing this type of information. I don’t think anyone knows exactly, but the figures are clearly astronomical. Truly astronomical,” said Gros.

“Today we launched ChatGPT. Try talking with it here,” wrote Sam Altman, then CEO of OpenAI, on X on 30 November 2022.

Since then, ChatGPT, Claude, Meta AI, DeepSeek, CoPilot and others have been competing to provide us with everything from cookie recipes to solutions to the world’s biggest problems.

Every instruction we give them passes through the data centres’ servers. Thousands of calculations are needed to determine which words the models should respond with.

Regardless of whether it has become your office assistant, meal planner, or psychologist: every keystroke you make consumes electricity.

Altman has said that our polite but completely unnecessary ‘thank you’ or ‘please’ costs the company tens of millions of dollars in electricity each year.

Currently, data centres use approximately 400 terawatt-hours, or one and a half per cent of the global electricity consumption. The IEA estimates that this will double in the next five years, reaching a level comparable to the entire electricity consumption of Japan.

Some of the data centres built for companies like Microsoft, Google, Amazon and Meta consume more electricity than large cities like Pittsburgh and New Orleans. Already heavily burdened power grids are under even more strain, and Americans are starting to notice that their electricity bills are rising.

The world’s largest data centre to date is being currently built in Jamnagar, India. American Nvidia, which has become the world’s most valuable company by selling chips for AI development, is heavily involved. According to the IEA, the centre could end up using as much electricity as the 10 million people living in the area.

Of all the aspects of AI development, commercial language models have received the most attention so far. However, in the shadow of these popular models, entirely different, efficient tools are being developed.

They run locally, use far less electricity, and can help us with entirely different tasks, such as detecting diseases faster, optimizing the power grid and perhaps even tackling the climate crisis.

At the AID centre, NTNU and SINTEF researchers are working to integrate AI more closely with industrial players and government authorities.

The goal is to develop tools that can manage risk and make decisions in challenging situations. These might be as varied as determining when it is safe to discharge a patient from the hospital, or ensuring stable electricity supply and production that is optimally adapted to consumption.

“Energy for AI, and AI for Energy,” said Fatih Birol, Executive Director of the International Energy Agency, when presenting the Energy & AI report in April 2025.

So, what was his point? That AI can also be part of the solution. According to the report, if AI is used to operate power grids more efficiently, we might be able to save up to 175 gigawatts of transmission capacity.

To put this into perspective, that could cover the electricity needs of 175 cities the size of Oslo for one year.

The Norwegian Water Resources and Energy Directorate (NVE) has analyzed how the country's energy market will change and grow (in Norwegian) They estimate that AI and data centres will account for two per cent of electricity consumption in Norway by 2050.

“I’m not exactly blown away by this number. The trends in Norway may not be as dramatic as we thought,” said Magnus Korpås, an energy systems expert and professor at NTNU’s Department of Electric Energy.

“Based on NVE’s figures, it looks like computing power will consume similar amounts of electricity as transport and other major consumers of electricity, which will increase in the years to come. And in any case, two per cent is very little compared to the 15 per cent used by electric panel heaters to warm our homes,” Korpås said.

There remains a lot of uncertainty. How things will look in 2050 partly depends on the development of data centres. Currently, a little over 70 centres are registered in Norway.

“NVE considers two per cent to be a realistic level that developers can reach and that Norway is able to handle. Worldwide, however, the demand for power is inexhaustible. Norway may be very well suited to becoming a major exporter of clean computing power. But then again, no one has actually suggested this,” Korpås said.

Korpås doesn't think that making Norwegian electricity available for a moderate number of data centres poses a dilemma.

“But the big question is whether we should make the power system and Norway’s natural environment available for the inexhaustible global consumption of AI. Whether we want to become the world’s hub for computer power is a political question,” he said.

He adds that this could very quickly be the case if we do not establish regulations.

“Establishing a centre here is attractive. Electricity is cheap, we have a cool climate, and the market is endless.”

Google is developing a centre in Skien. Green Mountain and TikTok have established themselves in Hamar. In Arendal, Bifrost Edge AS wants to build a centre that will use as much electricity as just over 100,000 households per year. This summer, Kjell Inge Røkke and Aker launched the major Stargate Norway project in Narvik together with OpenAI.

Professor Sebastien Gros sees rational arguments for developing data centres in Norway, and especially in the north.

“Financially, it is very rational to build in Narvik, an area with large electricity surpluses and low prices. Politically, it also makes sense, because it provides the country with revenue. And environmentally, clean Norwegian hydropower is better than coal power in California or China. We need to consider the advantages and disadvantages and look at the bigger picture,” said Gros.

Tomas Moe Skjølsvold, professor of Science and Technology Studies at NTNU, thinks the discussion about AI, data centres and energy is too narrow.

“We often get blinded by content like cat videos, but we need to zoom out and ask the big questions,” said Skjølsvold. He is head of the Norwegian Centre for Energy Transition Strategies – NTRANS.

As electricity becomes an increasingly scarce resource, it will be difficult to implement new energy projects.

“We need to ask who is going to be given access to the power, and under what conditions. When data centres apply for huge amounts of energy, we need to impose stricter requirements than just technical assessments, an area where the centres often score highly. Many more criteria should be assessed, such as whether the projects will contribute to verifiable cuts in emissions or other measurable environmental benefits,” Skjølsvold said.

He also highlights the geopolitical perspective. The United States has been open about its plans to secure technological global dominance through AI strategies.

“At the same time, we see companies from different countries building and operating data centres for each other’s services, such as Israeli actors operating facilities for Chinese TikTok. These are not neutral processes,” Skjølsvold said.

Skjølsvold believes the real questions relate to who gains power, control over infrastructure and influence over tomorrow’s digital economy – and how Norwegian resources would be used to make this possible.

Skjølsvold feels Norwegian society should consider two important questions:

• Whose political goals and global interests are we supporting by making Norwegian electricity and natural advantages available to data centres?
• What consequences will this have for Norwegian energy security, emergency preparedness and value creation?

“We need frameworks for transparency, societal benefit, environmental and climate considerations, as well as for the geopolitical direction Norway is actually helping to support,” Skjølsvold said.