AI and Energy: It Devours Electricity and Helps Economise It

The paradox is simple: the mind accused of devouring the planet's electricity may be the one that helps manage it. But slogans do not settle a paradox; figures do, along with the question of who stands behind them.

The first pan: accelerating demand

In its report "Energy and AI" (April 2025), the International Energy Agency estimates that data centres consumed about 415 terawatt-hours in 2024, around 1.5 per cent of the world's electricity, and that consumption grew by about 12 per cent a year over the previous five years. In its base case it reaches about 945 terawatt-hours in 2030, roughly 3 per cent. Accelerated servers, driven by AI, are growing at about 30 per cent a year and account for nearly half of the net increase.

The Agency itself states that uncertainty is considerable. In its 2035 scenarios, demand ranges from about 700 terawatt-hours, if adoption slows and supply is constrained, to more than 1,700 terawatt-hours if it accelerates. So do not build your judgement on a single figure.

The second pan: efficiency in cooling and in the grid

Cooling is among the largest sources of waste in a data centre. In 2016 DeepMind announced that neural networks trained on data from thousands of sensors in Google's facilities cut cooling energy by 40 per cent and overall overhead (the PUE measure) by 15 per cent. It is an example of AI reducing what it consumes, but the figures were announced by the company that ran the experiment at its own sites, so we should not generalise them to every facility.

At grid level, the Agency estimates that AI helps integrate renewable electricity and could unlock up to 175 gigawatts of extra transmission capacity in existing lines. In power plants, operating and maintenance savings could reach up to 110 billion US dollars a year by 2035. These are estimates in a "widespread adoption" case that assumes many obstacles are overcome, which the Agency lists: poor access to data, a skills gap, weak digital infrastructure and resistance to change.

Why a grid is harder than a single machine

Electricity is consumed at almost the same moment it is generated. When the sun and wind, which keep no timetable, entered the system, and homes became producers as well as consumers, the complexity of balancing from moment to moment multiplied. Hence the operator's need for tools that forecast output and loads and flag faults before they occur, tools that learn from data and need it clean and complete. A model is no more truthful than its measurements.

Limits not to be forgotten

• Security: the more digitally interconnected the grid, the wider the attack surface; see the article "AI and Cybersecurity".
• Fairness: the decision on which districts have their load reduced in a shortfall is a distributional decision involving justice or injustice, and should not be left to an unreviewed black box.
• The human in the loop: the machine handles precise calculation, while direction and accountability remain with people.

Our context

In January 2021 the Kingdom announced, through the president of the King Abdullah City for Atomic and Renewable Energy, a goal of generating half its electricity from renewables by 2030, with the rest from gas. As the share of sun and wind grows, so does the need for smarter forecasting, storage and management. Abundant sunshine is not enough on its own as a blessing; we give thanks to God for it by investing in people who understand both the language of energy and the language of the algorithm, so that we do not import the intelligence and then find ourselves unable to question it.

The simplest wisdom this question teaches is an old one: not to be wasteful with what God has bestowed on us. A single watt well used is worth more than many watts squandered.