AI's data centres are outrunning the grid built to feed them

Climate & Energy 4 min 80 sources

The world's electricity demand is speeding up, and computing is a big reason why. New reports this week put hard numbers on the gap between what AI needs and what the grid can build — and on who pays.

Key takeaways

The IEA says global electricity demand is accelerating — up 3.6% this year — with data centres, cooling, EVs and industry all pulling it higher.
In the US, data centres could add roughly 125 gigawatts of load by 2030 while utilities plan only about 93, and the cost of covering the gap is starting to show up on ordinary bills.
Much of the surge comes from things getting cheaper and more efficient — computing, cooling — being used far more, not less.

The world’s appetite for electricity is speeding up, not levelling off. A new report from the International Energy Agency — the West’s energy watchdog — landed Thursday. It forecasts global demand growing 3.6% this year and 3.8% next, up from 3% in 2025 [5]. That lifts world consumption to a projected 30,700 terawatt-hours in 2027, from 28,600 in 2025 [5]. The IEA names the drivers plainly: industry, appliances, cooling, electric vehicles, and data centres [5].

Data centres are the loudest of those, and this week the numbers stacked up fast.

The gap between what AI wants and what the grid can build

Data centres — the warehouses of computers that run AI — are on track to consume far more power than utilities have planned for. A Kansas Health Institute report published Monday found US data centres used about 183 terawatt-hours in 2024, more than 4% of the country’s electricity [2]. It projects that demand rising 133% by 2030, to 426 terawatt-hours, citing figures from Lawrence Berkeley National Laboratory [2].

Bank of America analysts put the supply side in stark terms. They expect the US to need more than 230 gigawatts of new generating capacity over five years. Regulated utilities are set to add only about 93 — a shortfall above 100 gigawatts [1]. Data centres alone could add roughly 125 gigawatts of load, pushing overall demand growth to a 4.1% annual rate through 2030 [1]. Tellingly, utilities have revised their demand forecasts upward three years running, as AI power needs “materialized faster than expected” [1].

The mismatch is already reshaping how power gets built and bought. Large gas turbines are largely sold out through 2030 [1]. GE Vernova, a major turbine maker, said its gas-turbine order backlog climbed to 116 gigawatts, up from 100 in the first quarter [55]. On Wednesday, Georgia Power, a Southern Company subsidiary, signed a 25-year deal with OpenAI [72]. Its new facility is expected to need about 3,200 megawatts — the output of roughly three large nuclear reactors [72]. Under the deal OpenAI will pay the full cost of the infrastructure to serve it [72].

Who pays

That last detail matters, because the central fight is over who bears the cost. Take PJM, the grid region that runs the power market for 13 eastern states. Over its last four capacity auctions, data centres accounted for $29.4 billion — 46% — of the $63.6 billion in charges, the market’s independent monitor said [22]. Capacity charges are what the grid pays generators to promise power will be there on the worst day; those costs land on everyone’s bill. “It is really a paradigm shift,” the monitor said, warning that pretending otherwise “imposes costs on others” [22].

Households are noticing. One analysis cited this week estimates monthly utility bills could rise 15% to 40% by 2030 as demand climbs [6]. On Thursday, President Trump expanded a voluntary pledge meant to shield consumers from those costs [6]. Nearly 200 utilities, data-center developers and governors have signed it, covering 80% of US power [6]. Whether it delivers real savings while demand keeps growing is unclear [6]. Opposition is spreading regardless: data-centre protests have gone national as communities weigh water, land and power against the promised jobs [45].

There is an environmental catch too. The power feeding data centres today is 56% fossil fuels, 22% renewables and 21% nuclear, the Kansas report found [2]. The new load is “delaying coal plant closures” that states had planned [2]. Water is the quieter strain: US data centres used an estimated 17 billion gallons directly in 2023, a figure the report says could double or quadruple by 2028 [2].

The same story, beyond the server farm

Data centres get the headlines, but the IEA’s list is broader. Much of the demand surge comes from things getting cheaper and better, not worse. Air conditioning is the clearest case. This June was the second-warmest on record, and 2026 is likely to be the hottest ever [4]. Both are lifted by a strong El Niño — the periodic Pacific warming that pushes global temperatures up. As cooling gets more affordable, more of the world runs it more of the time; one analysis notes AC’s projected electricity growth actually outpaces AI’s [4]. In Europe, long resistant to air conditioning, the argument that it is now a necessity is gaining ground [73]. Electric vehicles and electrified industry add to the same upward pull [5].

The under-covered corner: building clean, moving it dirty

Meeting all this cleanly is harder than building the panels. A report from Global Energy Monitor this week found a gap in China’s build-out [43]. The country installs more wind and solar than the rest of the world combined — yet still sends mostly fossil power down its long-distance lines [43]. Wind and solar make up only about 20% of the electricity carried on China’s ultra-high-voltage network, little changed since 2021, while 42% is coal [43]. The clean power exists; the wires and timing to use it don’t yet. It is a useful corrective to the demand panic: the bottleneck in the transition is often not generation but everything around it.

02 · Lesson · why it matters

Why making something efficient can make us use more of it

When a thing gets cheaper to use, we rarely pocket the saving — we use more of the thing, sometimes far more than the saving ever covered.

A puzzle hiding in the numbers

Here is something odd about this week’s reports. The computer chips inside AI data centres get more efficient every year — each new generation does more calculation for the same watt of power. By the old logic, that should mean the same work for less electricity. Instead, US data-centre power use is set to more than double by 2030, and the world’s total electricity demand is speeding up, not slowing down.

Efficiency went up. Consumption went up with it. That is not a mistake in the forecasts. It is one of the most reliable patterns in how people use resources, and it has a name.

The saving gets spent

Call it the rebound effect. When something becomes more efficient or cheaper to run, the real cost of using it drops — and a lower price changes what people decide to do. They use more.

The saving per unit is real. It just gets eaten by the extra units, and sometimes more than eaten. A cheaper light bulb doesn’t mean a smaller electricity bill; it means more lamps, left on longer, in rooms nobody’s in. A more fuel-sipping engine doesn’t cut the fuel bill; it makes driving cheaper, so people drive further. Efficient chips don’t shrink the data centre; they make computing cheap enough that companies pour it into things they’d never have bothered with before.

This is why it isn’t the same as “a cleaner grid can still pollute more if it grows.” That story is arithmetic — a rate times a total. This one is about a decision. The price of doing a thing falls, and so we choose to do far more of it.

It’s the whole list, not just AI

The IEA didn’t only name data centres. It named cooling, appliances, electric vehicles, industry. Look closely and most of that list is the same shape.

Air conditioning is the clearest case. As it got cheaper and better, it stopped being a luxury and became a default. Whole regions became livable that weren’t before; cool air went from something you rationed to something you assume. Each unit of cooling costs less than it used to — so we cool more homes, more offices, more of the year. The efficiency was real. It didn’t lower the world’s cooling bill; it raised how much cooling the world buys.

Efficiency, again and again, doesn’t just let us do the same thing for less. It quietly lowers the bar for doing the thing at all.

The promise built into the word

There’s an arrangement hiding under all this, and it’s worth seeing plainly. “Efficient” carries a promise — that using less per unit means using less overall. That promise sits inside energy policy, inside the marketing of every greener gadget, inside our own sense of doing better by upgrading.

It is often not true, and it’s worth asking who the assumption serves. It lets a company sell more of an efficient product as if selling more were the environmental win. It lets all of us grow our consumption while feeling we’ve economised.

But hold the other half too: efficiency is not a con. A world running today’s AI on the inefficient chips of a decade ago would burn vastly more power for the same work — or simply couldn’t do it. The gains are genuine. What’s false is only the quiet assumption that the gain automatically becomes a reduction. It doesn’t. What we do with the saving decides that, and mostly we spend it.

You’re in this, not watching it

It’s tempting to read this as a story about tech giants and their server farms. It isn’t only that. The rebound runs through the ordinary choices cheapness makes easy.

The AI query you fire off because it’s free and instant. The extra degree of cooling because the unit sips power now. The second screen, the always-on device, the trip you take because the car’s cheap to run. None of these feels like a decision to use more energy. Each one is. And the cost doesn’t vanish — it travels. It lands in the capacity charges that lift everyone’s bill. It shows up in the coal plant kept open past its planned close, and in the strain on a grid already short of spare capacity. The person protesting a data centre in their county and the person asking a chatbot to summarise an email are at two ends of the same wire.

What efficiency was never going to do alone

None of this is an argument against getting more efficient. It’s an argument against expecting efficiency, by itself, to shrink our footprint. It rarely has.

The saving is real; where it goes is a choice. It’s made not once, in a policy or a purchase, but a billion times a day — by people who each see only their own small “well, it’s cheap now.” No single seat sees the sum of those choices. That’s the humbling part. The forecasts keep getting revised upward — not because anyone lied. It is genuinely hard to picture how much more of a thing the world will do once that thing gets easy. We are, all of us, worse at that than we think.

03 · Lab · your turn

The Efficiency Dial

Rehearse the rebound effect: dial up efficiency and watch whether cheaper use conserves power or, when demand responds hard, burns more of it.

04 · Hope · carry this

The same ingenuity that keeps making our machines do more with less is real progress, even when we spend the saving faster than we save it — and the moment we notice that habit is the moment we can start choosing what the saving buys.