Data centres use 1.5% of the world's power, and 22% of Ireland's

Two numbers describe the AI power story, and they point in opposite directions. Globally, data centres consumed about 415 terawatt-hours in 2024, around 1.5% of world electricity¹. In Ireland, they consumed 22% of the country's metered electricity in the same year, a share that has more than quadrupled from 5% since 2015². Both numbers are correct. Which one matters depends entirely on where you are standing, and that gap between the calm average and the strained cluster is the actual subject of the data-centre energy debate.

The growth is real and steep: the IEA's base case has global data-centre demand more than doubling to about 945 TWh by 2030, adding slightly more than Japan's entire current consumption in six years, and reaching around 1,200 TWh by 2035¹. It is also, at world scale, not the biggest thing happening to electricity: data centres account for about one-tenth of global demand growth to 2030, behind industrial motors, air conditioning, and electric vehicles¹. The machines that cool bodies and move cars are adding more load than the machines that finish sentences.

The distinction that keeps both sentences true: in advanced economies, data centres are over 20% of all demand growth to 2030¹. Rich-world grids spent two decades with flat consumption, planned their capital cycles around it, and are now meeting their first large new industrial customer in a generation. The shock is not the absolute size of the load; it is the return of growth to systems that had stopped expecting any. A utility that has not built a greenfield gigawatt since the 1990s is being asked for several, on five-year notice, in specific counties. The 1.5% is also a share of energy, not of capacity: a campus draws its load flat around the clock, so its claim on peak-hour capacity and firm generation is larger than its annual share suggests.

Concentration is the story

Averages mislead because computing clusters harder than any previous industrial load. The United States held 45% of world data-centre consumption in 2024, China 25%, Europe 15%; inside the US, nearly half the capacity sits in five regional clusters¹. Northern Virginia, the largest of them, carries a data-centre load share closer to Ireland's than to America's. The grid experiences none of the global average and all of the local extreme: transformers, interconnection queues, and ratepayers live in the cluster, not in the mean.

The concentration also explains why the same industry can be praised and rationed in the same news cycle. A campus that is invisible in the national statistics is a fifth of the substation's load in its county, all of the marginal demand in its interconnection queue, and, during a heat wave, the largest flexible or inflexible customer the local operator has, depending on contracts written years earlier. The unit of analysis that matters is not the country but the feeder line.

Ireland is the longest-running natural experiment in what happens when a small grid hosts a global industry. Dublin's cable landings, tax regime, and cool climate drew the hyperscalers early; two decades on, the national statistics office publishes the sector's share of the country's electricity as a headline figure, the grid operator has effectively frozen new Dublin connections for years, and the 22% keeps compounding through moratoria because the connected sites keep filling². Ireland demonstrates the end state other clusters are heading toward: the debate stops being about carbon and starts being about who gets the next megawatt.

What actually powers it

The supply answer through 2035, on the IEA's arithmetic, is renewables first and firm power close behind: over 450 TWh of additional renewable generation for data centres, about 175 TWh of new natural gas, and roughly 175 TWh of nuclear¹. The nuclear line is the novel one. Hyperscalers have signed power purchase agreements to restart and life-extend reactors and seeded a queue of small modular reactor projects, because a training campus wants exactly what a reactor sells: gigawatt-scale, around-the-clock, carbon-accounted electricity in one contract. The AI build-out has done more for nuclear's order book in three years than two decades of climate policy managed.

Gas is the quiet second winner, especially in the American South, where speed beats everything: a combined-cycle plant arrives years before a reactor and connects where the pipelines already run. The emissions consequence depends on which grid the load lands in, which is why identical racks produce very different carbon in Virginia, Quebec, and Ireland, and why siting has become the industry's main climate instrument whatever its sustainability reports emphasise.

What the panic gets wrong, and right

The public debate runs ahead of the meter in both directions. Claims that AI will consume some double-digit share of world power by 2030 are not supported by the central estimates: even after doubling, data centres of every kind, AI and the ordinary internet together, sit near 3% of global consumption¹. In the other direction, dismissals that cite the internet's history of efficiency saving the day skip over the local arithmetic: efficiency improves globally while substations overload locally, and a neighbourhood whose power price rises to fund grid reinforcement is not comforted by the world average.

The real uncertainty is demand itself. The 945 TWh case assumes the AI build-out's announced campuses largely materialise and run; a capex slowdown, an efficiency jump in inference, or model architectures that need less compute per query would bend the curve down, and agentic workloads that run continuously would bend it up. Electricity forecasting for this sector inherits all the volatility of the industry it powers, and the grid's planning horizon, a decade, now has to absorb an industry whose planning horizon is a quarter.

The bottleneck is the wire, not the watt

Generation is the solvable half. The IEA's sharper warning is about delivery: without intervention, around 20% of planned data-centre projects risk delay for grid connection, transformers, and switchgear¹. Interconnection queues in the big US markets run to years; high-voltage equipment has global waiting lists; and the skilled trades that build substations are the same ones the chip fabs and the energy transition are bidding for. Electricity is becoming the binding constraint on AI capacity planning, which is a sentence nobody in software believed five years ago. The wire is the modal bottleneck, not the only one: in dry regions the binding constraint is cooling water, in dense ones permitting, and in isolated grids generation itself.

That scarcity is repricing behaviour up and down the stack. Developers now shop for grid headroom the way they once shopped for fibre routes, striking deals in Texas, the Gulf states, the Nordics, and the Middle East where power can be contracted at scale; utilities demand take-or-pay terms so ratepayers do not fund stranded substations; and the efficiency frontier inside the racks, better chips per token, better cooling, load-shifting between regions, is advancing because power now appears on the income statement, not just the sustainability report. Scarcity is doing what scarcity does: making the industry ingenious about the thing it wasted.

One more correction the meter makes to the discourse: the marginal query is cheap, and the marginal campus is not. A single AI request costs less energy than the kettle boiled while waiting for it, and repeating that comparison misses where the money and the megawatts actually collide, in the training clusters and always-on inference fleets whose demand is contracted years ahead regardless of any individual user's habits. Personal-consumption framing flatters everyone and decides nothing; the real decisions are made at grid level.

The politics arrive with the meter

Because the load clusters, the politics are municipal before they are national. County boards in Virginia now hear data-centre zoning cases the way they once heard highway fights; Dublin's connection moratorium is the ur-example of a rich place rationing its grid; and everywhere the industry lands, the same three-way negotiation repeats between a developer who wants speed, a utility who wants guaranteed offtake, and residents who want neither the substation nor the rate case. The industry's social licence, cheap to ignore during the fibre era when its buildings were invisible, now gets negotiated one interconnection at a time.

Host governments have also noticed the leverage. Power-rich states are converting electrons into industrial policy: Gulf sovereigns offer gigawatts as the anchor of AI partnerships, Nordic and Canadian grids market hydro surplus as a siting subsidy, and every one of these deals trades energy for a seat in the compute economy. Electricity used to be an input the tech industry abstracted away. It is now a currency countries negotiate in, which is the quiet geopolitical consequence of the curve this piece began with.

What to watch

Three gauges through 2030. Ireland's percentage, because it tests whether a democracy keeps hosting a load that takes a fifth of its power while employing very few people; the CSO now publishes the number annually². The IEA's 945 TWh base case against delivery, because the projection embeds efficiency gains that may or may not arrive on schedule¹. And the share of new campuses signing firm-power contracts, nuclear restarts, gas with capture, geothermal, because that number decides whether AI's power demand becomes the grid's biggest problem or the cleanest anchor tenant decarbonisation has yet found. The racks are indifferent; the contracts are not, and the contracts are being signed now, cluster by cluster, for the grid the 2030s will have to run on.