The Solutionist Identity

Power brokers

The first of our Solutionist series of articles describes the AI data-centre boom that is reshaping energy geography—and demanding new solutions

11 Jun 2025

Equinix ME1, an Equinix Melbourne Data Centre, showcases its distinctive angular architecture against a cloudy sky. The building features dark grey and black panels, with a glass entrance visible at the lower right.Photograph: Equinix

The world’s newest arms race isn’t about warheads or semiconductors, it’s about watts. As artificial intelligence (AI) explodes in capability and commercial value, the data centres powering it have become the key infrastructure of economic competition. These often dull-looking, windowless buildings hum with high-performance chips and are the digital economy’s most prized real estate. Their location is determined by myriad factors, including proximity to data activities, permitting and regulation. But the critical factor is the electricity they can access—how much, how clean and how constant.

Training and using large AI models is computationally and energy intensive. The models require high compute density, stable electricity voltages and continuous cooling, first to train, and then with different requirements, to run. Those intense needs rule out many legacy data centres and also often require more than traditional utilities are equipped to provide.

The combined demands of new, GPU-stocked AI facilities and existing facilities risks saturating power systems. Equinix, a global digital infrastructure company and big developer of data centres that works closely with Nvidia, the world’s leading AI chip designer, operates in 35 countries. About 100 of its 270 facilities use liquid cooling, enabling the most advanced AI. It says underinvestment in energy generation, transmission and, most crucially, in local distribution, is a bottleneck for AI almost everywhere. “Utilities have not historically had to invest ahead of demand and are often slow to upgrade,” notes Raouf Abdel, executive vice president of global operations at Equinix.

A query to ChatGPT uses around ten times the energy of a Google search, according to the International Energy Agency (IEA), which advises governments on energy policy. Demand from such energy-intensive AI models largely explains why the agency expects global data centre electricity use to double by 2030 on current trends. By 2035, it says, data centres could consume roughly five and a half times Britain’s annual electricity use—on a scenario of faster AI growth and fewer energy bottlenecks (see chart 1).

Those striking numbers must, however, be kept in perspective, says Fatih Birol, executive director of the IEA. After all, air conditioning and industry overall will account for a bigger increase in global electricity demand between now and 2030, than will AI, he notes (see chart 2). What truly distinguishes AI, Mr Birol explains, is the intensity of its energy demand and its concentration in certain places and moments. “The impacts of AI on power systems could be extremely pronounced at the local level,” he says.

Evidence of that local impact is visible across the globe. Ireland is discovering that data centres may account for a third of its total electricity demand by 2030. Such are the constraints on the Irish grid that Equinix has to deploy its own energy sources there, such as gas turbine generators. Large clusters of data centres in northern Virginia, popular for its proximity to Washington, DC, risk straining power systems. Geographies that possess in abundance what data centres need, meanwhile, are attracting many more.

Central and local governments now see accommodating data-centre expansion as critical to economic competitiveness. Gulf states advertise their cheap and abundant energy resources to draw huge data-centre investments. The United Arab Emirates has launched a $20bn project, backed by American and Japanese firms, to build the largest AI data centre outside America, due to begin operations in 2026.

A new data centre can be built in 18 months, but building new transmission lines can take four to eight years.

Fatih Birol, Executive director of the International Energy Agency

Photograph: Equinix

Vying for AI supremacy, America and China are similarly driving expansion. The US Department of Energy has begun fast-tracking federal permits for AI facilities. Chinese state-led efforts are integrating data-centre expansion with investment in regional grids.

Elsewhere, early prudence is giving way to a game of catch-up. Singapore and Amsterdam lifted data centre moratoria in 2022 and 2020, respectively, with the proviso of strict energy-efficiency standards. But persistent high rents and restricted capacity in both places still favor data-centre growth in nearby competitors like Malaysia or Paris. Such developments underline the AI data centre boom’s two huge challenges: first, how to secure enough energy at reasonable cost, and, second, how to ensure AI factories can be green while still matching the needs of the latest and best chip technology.

The biggest AI players of all–hyperscale tech companies–are simply securing electricity for themselves. They have carbon emissions targets in mind. Last year Microsoft signed a deal with Constellation Energy to revive the Three Mile Island, Pennsylvania nuclear power plant, to match the power its data centres use with carbon-free energy. Amazon has also struck nuclear energy deals, including its own investment in developing small modular reactors, and Meta plans to scale up geothermal power for its vast compute needs.

Having deep pockets means the hyperscalers can overcome obstacles like grid capacity constraints and delays connecting to the public grid. They are also vertically integrating: buying land, building their own transmission lines and co-locating generation assets like battery farms or even natural-gas turbines directly with their compute campuses.

By greening the electrons, we enable data centres’ carbon emissions to move towards zero.

Pankaj Sharma, Executive vice president of secure power, data centres & global services at Schneider Electric

The result is a two-tier system: on top sit hyperscalers with the capital and technical muscle to build energy-integrated campuses optimised for AI. Some large, global independent data centre operators, such as Equinix, can do the same. Then further down sit smaller legacy data-centre groups and facilities that run up against limits of trying to retrofit buildings with AI-compatible equipment. They can be stuck waiting years for grid connection permits.

The mismatch in timelines between building data centres and getting them connected to power, both in America and across Europe, is acute. “A new data centre can be built in 18 months,” says Mr Birol, “but building new transmission lines can take four to eight years.” Power density needs are rising all the time. Designs from Nvidia for AI server racks in 2027 will require 600kw, or 50 times the needs of server racks currently powering the internet.

Photograph: Equinix

The explosion of different needs is sparking innovation. Some data-centre operators are exploring microgrids and advanced software to modulate load dynamically in response to grid fluctuations. And innovators are selling grid tools to optimise energy use and ensure the reliability data centres depend on.

The challenge of powering AI in a sustainable fashion comes as the technology itself is evolving. The efficiency of algorithms and computational processing units is changing rapidly. Patterns of energy use are different to train AI models than to conduct inference (running them). It is fiendishly hard for data-centre operators to plan decades ahead for an unknown technological future, but they must, notes Equinix’s Mr Abdel.

When it comes to energy sources, the risk that environmental goals will be downgraded in the global race for AI leadership is a real one. National moratoria, now scrapped, were originally driven by legitimate concerns about meeting sustainability targets. For their part, Microsoft, Amazon and Meta invest in a variety of generation assets. They state that they have reached what they call 100% renewable targets, but have also purchased data-centre capacity that is powered by new gas plants.

Many new AI training campuses, such as Pennsylvania’s massive planned Homer City Energy Campus, are backed by fossil-fuel turbines. “The ambition is there—net zero, 24/7 clean power, but a lot of these campuses are still powered by gas,” says Giampiero Frisio, president of electrification at ABB, a Swiss-Swedish industrial-technology company. Energy infrastructure as well as generation is another key bottleneck. "By greening the electrons, we enable data centres’ carbon emissions to move towards zero." says Pankaj Sharma, executive vice president of secure power, data centres & global services at Schneider Electric, a French-based energy technology company.

Today’s decisions about how to power AI will shape the carbon profile of compute for decades to come, and are crucially important. If clean baseload power and essential grid upgrades are rolled out slowly, the IEA predicts that AI’s electricity demand will stay tethered to fossil fuels well into the 2030s (see chart 3). On the other hand, coordinated planning could steer AI infrastructure toward integration with low-carbon supply.

“Strong dialogue between governments, the energy sector, the tech industry and civil society is imperative to plan the future of AI data centres,” says the IEA’s Mr Birol. As electricity becomes the chokepoint of intelligence, today’s power brokers are positioned to become tomorrow’s AI kingmakers.

More from the series

Powering up innovation for AI data centres | The Solutionist
Supported by Schneider Electric

Powering up innovation for AI data centres | The Solutionist

The AI factory transforming your neighbourhood | The Solutionist
Supported by Schneider Electric

The AI factory transforming your neighbourhood | The Solutionist