For tech entrepreneur Elon Musk, the answer to the rocketing energy needs of artificial intelligence (AI) data centres is to launch them into space, where they could tap limitless energy from the sun. But until that happens, the places on Earth where these number-crunching mega-hubs are located face big spikes in electricity demand to run them.
In the US, this has sparked fears of higher energy prices for consumers. To allay those concerns, President Donald Trump will reportedly convene big tech firms this week to sign a pledge to provide or pay for the extra energy supplies they will need as their AI data centres expand.
According to the International Energy Agency (IEA), data centres accounted for 1.5% of electricity demand worldwide in 2024 – a share set to rise to about 3% by 2030. Overall, data centre demand is expected to more than double to about 945 terawatt-hours (TWh) by then, which is slightly above the electricity consumption of Japan today.
AI data centres, where AI models are trained and deployed, put far more strain on power supplies than traditional data centres, which each use between 10 and 25 megawatts (MW). In comparison, demand from a “hyperscale” AI centre can exceed 100 MW at any given time, which if running at full capacity could consume as much electricity in a year as 100,000 households.
Data-centre electricity consumption in household electricity consumption equivalents (million households), 2024
We look at where this power might come from and whether, as some warn, AI is going to blow the world’s efforts to transition away from fossil fuels out of the water.
Why does AI need so much electricity?
AI data centres differ in how they use electric power. In a conventional data centre, data requests from businesses, individuals and other users come in a randomised way, translating into a steady load level on the servers, with relatively little fluctuation in demand.
But in an AI data centre, processors need to go through training or learning periods, using so-called “graphical processing units”. These are synchronised, being started up and switched off at the same time. This translates into “power bursts”, which last just a few seconds, but happen very frequently and concurrently, according to Gerhard Salge, chief technology officer at Hitachi Energy.
“That is a different challenge than just providing the power and the energy for the conventional data centres,” he told journalists at the International Renewable Energy Agency assembly in Abu Dhabi earlier this year.
Here, officials and business executives discussed how to meet those demand peaks, noting they cannot be dealt with just by installing huge batteries as those would wear out quickly.
Martin Pibworth, chief executive of SSE, a Scotland-based energy firm, said AI-led demand will put pressure on the power system, but “the problem we all have is no one really knows the pace and trajectory of that demand lift”. In the UK, the government’s Clean Power Plan will be needed to make sure electricity operators can meet demand from AI and other data centres as more come online, he added.
In the US, meanwhile, the Trump administration is eager to ensure that communities that are home to data centres, as well as the wider public, do not turn against the industry due to its perceived unfairly high use of energy and water.
Ahead of a meeting scheduled on March 4, where US tech titans are due to sign a pledge on powering their own data centres, White House spokesperson Taylor Rogers told CNBC: “Under this bold initiative, these massive companies will build, bring, or buy their own power supply for new AI data centres, ensuring that Americans’ electricity bills will not increase as demand grows.”
Will electricity for data centres and AI come from clean or dirty sources of energy?
The answer to this question is key to how countries tackle climate change, as it will affect their energy mix, how electricity is produced and distributed, and therefore the trajectory of their greenhouse gas emissions. Decisions made by governments and businesses will shape how the AI industry powers the technology on which it relies.
Under pro-fossil fuel Trump, the US has walked away from policy support for clean energy, meaning data centre operators can choose their energy sources freely. In January, data from Global Energy Monitor (GEM) showed the US now has the most gas-fired power capacity in development, surpassing China and accounting for nearly a quarter of the world’s total.
More than one-third of this capacity is set to directly power data centres on-site, in hotspots like Texas, and many more grid-connected gas-fired projects are planned to meet an expected increase in energy demand from AI, GEM said.
On the other hand, some tech companies – especially multinationals – have set goals to cut their emissions to net zero, and so are choosing to power their data centres with renewables, including in the US.
For example, French energy giant TotalEnergies recently signed two long-term Power Purchase Agreements (PPA) to deliver 1 gigawatt (GW) of solar capacity for Google’s data centres in Texas. This followed two other PPAs with Google for 1.2 GW secured by Clearway, a California-based renewables company 50%-owned by TotalEnergies.
Sources of global electricity generation for data centres – base case, 2020-2035
Some countries are also moving to ensure the power needed for AI and the data centre industry is produced using clean energy.
In Ireland, an effective ban on new data centre connections was lifted in December, provided at least 80% of the centres’ annual energy demand is met by new renewable electricity sources. The government also plans to build Green Energy Parks, where data centres can be located alongside renewables plants to avoid straining the national grid.
Salge of Hitachi Energy said that with big investors wanting to drive investment in AI data-crunching so fast, “there is no other power generation technology than variable renewables which you can build in such a timeline” of two to three years. “Anything else will be in the 2030s and later,” he added.
Some governments – such as Sweden’s centre-right coalition – have proposed nuclear as a clean energy solution for AI data centres, saying they could fuel a “renaissance”. But building nuclear power plants requires massive investment and long timelines, while new small-scale modular reactors are not yet commercially available.
How are power systems and regulators coping so far?
In a February report forecasting electricity demand out to 2030, the IEA said AI and data centres are contributing to generation growth in advanced economies, which is now accelerating again after 15 years of stagnation. However, it flagged bottlenecks in connecting new data centres, because grids are not being built or improved fast enough to keep up with rising power demand, forcing big customers to wait.
The report noted that at least 150 GW of queued data centre projects are estimated to be in the advanced stages, while one-fifth of the global data centre build-out is at risk of delay due to grid congestion.
Comment: Using energy-hungry AI to detect climate tipping points is a paradox
Planning, permitting and completing new grid infrastructure can take five to 15 years, whereas data centres need one to three years. Prices for key grid components have also nearly doubled over the past five years, the IEA noted.
The European Commission, meanwhile, aims to support those operators that can save on energy use. It plans to adopt a “Data Centre Energy Efficiency Package” in April that will contain an assessment of data submitted under a reporting scheme, introduce a rating scheme for data centres in the EU, and start work on minimum performance standards.
Can AI help to resolve the issue?
Experts say it’s important to look at both sides of the coin, pointing to ways in which AI can contribute to more effective power grid management and integration of renewables into national power supplies.
According to new analysis by energy think-tank Ember, AI applications such as short-term renewables forecasting, predictive maintenance, and real-time monitoring and adjustment of transmission line capacity can deliver operational improvements in power systems.
It estimates that AI could enable Southeast Asian nations, for example, to reduce their power sector costs by $45 billion-$67 billion through to 2035, depending on how much renewable energy they deploy. Potential AI-driven efficiency gains could cut emissions by 290 million to 386 million tonnes of CO2 over the next decade in ASEAN countries, it adds.
“While power-hungry AI might initially stress the power systems, with various powerful applications it has the potential to significantly accelerate the energy transition and offset consumed energy rapidly,” Ember data analyst Lam Pham said in a statement.
