top of page

Everyone's Watching The Robot. Nobody's Watching The Plug. (AI Power Demand)

Jul 5
6 min read

Updated: Jul 17

Category: Energy Transition


A strange thing is happening to the electricity grid


Ask most people what's limiting artificial intelligence right now and you'll get one of two answers: computer chips, or clever people.


Both are wrong. The answer, increasingly, is electricity.


Picture of power lines

Not "will be." Is. Right now. There are AI companies that have the money, the chips and the engineers, and cannot build, because there isn't enough power available to plug the thing in.


This is one of my favourite kinds of story, because it's the sort the market almost always gets wrong: everyone stares at the glamorous bit, which robot is cleverest, while the entire thing quietly turns out to be decided by the boring bit. Wires. Transformers. Substations. The least sexy objects on Earth.


Let me show you what's actually going on, why it's about to land on your electricity bill, and what it tells us about where enormous amounts of money are about to flow.


What a data centre actually is


Let's start from zero, because "the cloud" is one of the great misleading phrases of our age.


There is no cloud. There is someone else's computer, in a shed, somewhere.


When you ask an AI a question, your phone doesn't do the thinking. It sends your question down a wire to an enormous windowless building, a data centre, full of racks of computers. Those computers do the work and send the answer back.


That's it. That's the whole magic trick.


Now, computers do two things: they perform calculations, and they generate heat. And here's the bit people miss, the heat is the problem.


Every watt of electricity a computer uses eventually becomes heat. And if you let heat build up, the computers cook themselves. So a data centre has to also run enormous cooling systems, which themselves consume vast amounts of electricity.


So you're paying for electricity twice: once to do the thinking, and once to remove the heat produced by the thinking.


A data centre is, in a very real physical sense, an enormous electric heater that happens to also do maths.


The numbers stopped being normal


Here's where this gets genuinely startling.


A large, conventional data centre, the kind running your email and Netflix for the last decade, needed something like 10 to 20 megawatts of power.


Illustration of a large data centre

The new AI-focused ones being designed today are being built for 100 to 300 megawatts. And the biggest campuses now being planned are aiming at a full gigawatt.


A gigawatt is, very roughly, the output of a large nuclear power station.


For one site. One building complex. Doing maths.


And Meta's chief executive has publicly said the company plans to build "tens of gigawatts this decade, and hundreds of gigawatts or more over time." That's not marketing waffle, that's a direct, physical claim on the world's electricity supply, stated openly.


To give you a sense of the aggregate: global data centre electricity consumption is forecast to grow by more than a quarter in 2026 alone. The Electric Power Research Institute now estimates that American data centres could consume somewhere between 9% and 17% of all US electricity by 2030, and, crucially, that estimate has been revised upwards repeatedly. Every time somebody re-runs the numbers, they come back bigger.


That's the tell, by the way. That's the thing worth noticing. When forecasts for something keep getting revised in the same direction, it usually means the forecasters are still underestimating a trend they don't fully understand yet.


Why this shows up on your bill


Here's the part that turns this from an interesting tech story into something that affects a person who has never used an AI in their life and has no intention of starting.


Electricity grids don't just supply electricity. They have to be built to handle the maximum possible demand at any moment, the peak. That means power stations, transmission lines, transformers, substations. Enormous, expensive, slow-to-build infrastructure.


Now: if a data centre turns up in your region and demands as much power as a small city, the grid has to be upgraded to handle it. New lines. New capacity. Billions of pounds of it.

And who pays for grid upgrades?


Everyone on the grid. Through their bills.


This is not theoretical. In America's PJM electricity market, covering a huge swathe from Illinois to North Carolina, analysts have traced roughly $9 billion of a recent surge in electricity capacity pricing directly to data centre demand. In Virginia, home to one of the densest concentrations of data centres on Earth, the local utility filed for its first base-rate increase in over three decades, and linked it explicitly to the buildout needed to serve new data centre load.


So there are households in Ohio, Maryland and Virginia paying measurably more for electricity, right now, because AI companies moved in nearby.


Whatever you think about that, and it raises some genuinely difficult fairness questions, it's happening, and it's the first time in a long while that a technology boom has landed so directly and so visibly on the utility bills of people who aren't part of it.


The collision with climate targets


Here's the tension that makes this genuinely interesting rather than just a story about big numbers.


Governments across the developed world have committed to decarbonising their electricity grids.


That means replacing fossil generation with renewables, a job that was already extremely difficult.


A man looking stressed at work

And into the middle of this, we've dropped a brand new source of electricity demand, growing at a ferocious rate, that wants power 24 hours a day, 7 days a week, without interruption.

That last bit matters enormously, and here's why.


Solar power produces electricity when the sun shines. Wind produces when the wind blows. Neither of those things is 24/7, which is fine for a grid that can flex, and a problem for a customer that categorically cannot pause.


A data centre doesn't want to be told "sorry, it's cloudy." It wants constant, reliable, always-on power.

Which means the AI boom is quietly creating enormous demand for the kinds of electricity that run constantly: nuclear, geothermal, gas with carbon capture, and grid-scale batteries that can smooth out the gaps in renewables.


This is precisely why you've suddenly seen technology companies, organisations with no obvious business being in the energy sector, signing deals with nuclear power stations. They're not doing it for the press release. They're doing it because they need firm, constant power, and there aren't many ways to get it.


Where the money has to go


Let's do the thing I always want to do here, which is take a real-world problem and trace where the capital is forced to flow.


If all of the above is true, and the numbers say it is, then the following things become structurally necessary, not optional:


  • Transmission: the wires that move power around. Chronically under-built in most countries for decades.

  • Grid equipment: transformers, switchgear, the physical hardware of the grid. There are already global shortages and multi-year waiting lists for some of this equipment.

  • Firm generation: nuclear, geothermal, storage — anything that produces constantly.

  • Cooling and efficiency technology: because every watt saved on cooling is a watt available for computing.


Notice something about that list. There isn't a single glamorous name on it. Nobody is making a documentary about transformer manufacturing. There is no cult of personality around a substation.

And that's precisely the point I want to leave you with.


The market is extremely good at pricing the exciting layer of a story, and consistently slow to price the boring layer underneath it. Everybody has a view on which AI company will win. Far fewer people have a view on who supplies the switchgear.


I'm not telling you to go and buy anything, I don't know your circumstances, and anyone who gives you a stock tip without knowing them is not your friend. But I am telling you that the skill worth developing is this one: when you see a boom, ask what it physically requires to exist. Then ask whether anybody is paying attention to that.


Very often, they aren't. That's where the interesting thinking is.


The robot is fascinating. But the story is the plug.


Next in this series: for every pound the world spends protecting nature, it spends about thirty destroying it. Here's what that actually means.


References

Belfer Center, Harvard Kennedy School - "AI, Data Centers, and the U.S. Electric Grid: A Watershed Moment": https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid

Electric Power Research Institute (EPRI) - data centre load growth modelling: https://www.epri.com/

PJM Interconnection - capacity market auction results and analysis: https://www.pjm.com/

Comments

bottom of page