AI is eating the power grid – and on both sides of the Atlantic, you’re the one picking up the bill
Last Thursday, the U.S. Department of Energy handed the operator of America’s largest power grid emergency authority to “direct backup generation resources to operate as a last resort” before rolling blackouts across 13 states and Washington, D.C. It was the latest in a string of such orders this year. In Europe, meanwhile, a new data centre in Frankfurt or Amsterdam can wait the better part of a decade just to be plugged in. This is what the AI boom looks like from the grid’s side, and it’s not pretty.
The scale of the problem
The numbers have stopped being abstract. According to the International Energy Agency, global data centre electricity use grew 17% in 2025, while consumption at AI-focused data centres surged 50%. The IEA expects total data centre demand to roughly double to around 945 TWh by 2030. Big Tech’s capital spending topped $400 billion last year and is expected to jump another 75% in 2026.
In North America, the grid reliability regulator NERC now forecasts summer peak demand growing by 224 GW over the next decade, 24% above last year’s peak, with new data centres accounting for most of it. It says 13 of its 23 assessment areas face elevated or high risk of supply shortfalls over the next five years.
The core problem is time. As the UN Economic Commission for Europe warned this month, a large data centre can be built and connected in two to five years, but the power lines and grid upgrades to feed it can take more than ten. Data centres are, in the UNECE’s words, “being built much faster than the power grids needed to supply them.”
America: the bill arrives
Nowhere is that clearer than in PJM, the grid serving 65 million people from Illinois to North Carolina. Its latest capacity auction, which sets what utilities pay to guarantee enough power plants will be available, cleared at the price cap of $329.17 per megawatt-day. Without that cap, PJM estimates the price would have hit $555. In the auction before, data centres accounted for 63% of the price increase, or $9.3 billion recovered from customers through their electricity rates.
Households are noticing. During debate on the bill below, lawmakers pointed to power prices rising 13% in Virginia, 16% in Illinois and 12% in Ohio over the past year, more than double the 6% national average. Last Wednesday, the House passed the Ratepayer Protection Act 417–3, a rare bipartisan push to make states consider rules ensuring data centres cover the grid upgrades and costs they cause. (Not everyone accepts the link: House Majority Leader Steve Scalise insisted data centres are “good neighbors and studies prove that they don’t add to the cost of the grid,” and a Rutgers analysis of New Jersey found data centres mostly aren’t raising bills there… yet.)
Texas has gone further. On August 3, Governor Greg Abbott froze new data center projects pending a statewide audit, after ERCOT found itself facing roughly 474 GW of connection requests, more than five times the state’s record peak demand. About 90% of that is data centres.
And the grid is getting jumpier. In May, NERC issued a rare Level 3 alert, its most urgent kind, after incidents in which 1,000 MW or more of computing load dropped off the system in seconds. “Rapid, major swings in load… can impact the [bulk power system’s] ability to maintain frequency, regulate transmission voltage, and otherwise maintain stability,” it warned.
Plan B: build your own power plant
Unable to get connected, the industry is increasingly going around the grid. More than 130 GW of generation has been proposed to power U.S. data centres directly, over 80% of it natural gas, according to RBC Capital Markets. That’s colliding with a global shortage of gas turbines, and the nuclear deals that grab headlines are mostly long-term contracts with existing plants. New small reactors won’t add meaningful capacity until the next decade, the Carnegie Endowment notes.
President Trump has embraced the do-it-yourself model, asking tech companies to build or secure their own power rather than lean on the grid. It’s the same administration that, at the UN today, promised to “encourage it, not rein it in.” Encouraging the demand while hoping someone else builds the supply is not an energy policy.
Communities are pushing back regardless. Local opposition blocked or delayed 45 U.S. data centre projects worth $68 billion in the second quarter alone, according to Data Center Watch.
Europe: the grid said no
If America’s problem is paying for the boom, Europe’s is plugging it in at all. In the continent’s five biggest data centre hubs, Frankfurt, London, Amsterdam, Paris and Dublin, new facilities wait seven to ten years for a grid connection, and up to 13 in the most congested areas. Dublin has a de facto moratorium until 2028. The Netherlands and Frankfurt have effectively closed to new connections until around 2030.
Ireland shows why. Data centres went from 5% of the country’s electricity demand in 2015 to 22% in 2024, and grid operator EirGrid expects 31% by 2034. Its regulator now requires new data centres to match 80% of their annual demand with new Irish renewable generation within six years of switching on.
Then there’s price. EU electricity for energy-intensive industry averaged more than twice U.S. levels in 2025, the IEA says, and it’s expected to stay that way this year. The UK is resorting to discounts of up to £24 per MWh from 2027 to lure data centres to its “AI Growth Zones” in Scotland and the north of England, where there’s spare wind power but not enough cable to move it.
Brussels wants it all anyway: its planned Cloud and AI Development Act aims to triple EU data centre capacity in five to seven years. Yesterday, the Commission proposed a rating scheme for data centres over 500 kW that would score their energy and water use and how flexibly they can adjust to grid conditions, with the first labels due in 2027. It’s sensible. It’s also a label, not a power line.
It doesn’t have to be a crisis
To be fair, there is a better version of this story. The IEA says energy use per AI task is falling at a rate “unprecedented in energy history.” U.S. solar generation is set to jump 17% this year, per federal forecasts. And data centres that can dial down when the grid is stressed could become a stabilising force rather than a liability. That’s exactly what Europe’s new rating scheme and America’s emergency backup orders are groping towards.
MadRobot’s take
The issue isn’t that AI uses a lot of electricity. It’s that the companies building it move on a two-year clock, the grid moves on a ten-year clock, and households are left covering the gap through their bills. Three things would help on both sides of the Atlantic: make data centres pay upfront for the grid capacity they need rather than spreading it across everyone’s rates; make flexibility a condition of connection, not an emergency afterthought; and treat transmission lines with the same urgency as the AI race itself.
Politicians love talking about winning the race to superintelligence. None of it runs without a plug.
