A couple of streets from where I live, there’s apparently plenty of spare capacity at the local exchange. That didn’t stop my broadband install taking months to actually happen, because it turns out having capacity somewhere nearby and having a working connection to your actual house are two completely different problems. Multiply that everyday annoyance by about a billion, attach it to a building full of scorching-hot computer chips, and you’ve got the real story behind AI’s power problem.
Type a prompt into a chatbot and an answer appears seconds later, as if from nowhere. There is no “nowhere,” though. Behind that little text box sit thousands of specialised chips, running hot inside vast data centres, and those chips are extraordinarily thirsty for electricity.
Here’s the part that surprises people: the biggest problem facing the next wave of AI isn’t generating enough power. It’s physically getting that power to the exact patch of land where a tech giant wants to build a server farm — which means transmission lines, substations, transformers, thick copper cable, planning permission, and a lot of very unglamorous construction.
So will AI actually run out of electricity? Not exactly. We can always build another power station. The real crunch is that software moves at the speed of the internet, while the physical network meant to power it moves at the speed of steel, copper, and red tape.
What Is the Grid Bottleneck, Exactly?
Picture electricity generation as a huge reservoir. Building a new power station is like pouring more water into that lake. But if the pipe carrying it to the city is too narrow, all that extra water is doing nothing useful at all.
The grid works the same way. Power has to travel down high-voltage transmission lines, step down through local substations, and pass through enormous transformers before it ever reaches a data centre. If a single one of those links is maxed out, a billion-pound facility sits there in the dark and waits, even if there’s ample electricity being generated fifty miles up the road. The bottleneck usually isn’t watts. It’s wires.
Why Are AI Data Centres So Hard to Power?
Ordinary data centres already use a fair amount of energy, but AI pushes that into an entirely different league. AI relies heavily on GPUs — graphics processing units, chips that are brilliant at running thousands of complex calculations at once, which is essentially how AI “learns.” Pack hundreds of thousands of them into one building and the heat they throw off is genuinely staggering. Powering the chips and running the industrial cooling needed to stop them melting means a single modern AI campus can draw as much continuous electricity as a medium-sized city.
Developers can no longer just look for cheap land. They have to ask whether the local power lines can actually carry that much current, whether the nearby substation has any spare capacity, and how many years it might take to get a grid connection approved in the first place.
The Waiting Game
A tech company can order the servers, put up the building, and have the software running within months. Electricity infrastructure runs on a completely different clock.
According to the International Energy Agency, getting new high-voltage transmission lines permitted and built in advanced economies typically takes four to eight years. Ordering a large power transformer isn’t much quicker — the IEA’s own industry survey found that lead times for transformers and cables have very nearly doubled since 2021, with big transformers now taking up to four years to secure and some specialised cabling running past five. In the world’s largest data-centre market, Northern Virginia, the average wait for a grid connection is now around seven years. Across Europe’s biggest data-centre hubs — Frankfurt, London, Amsterdam, Paris and Dublin — queues of seven to ten years are becoming normal.
You can have all the money, the cleverest software, and the fastest chips on the planet, and still be stuck waiting on a piece of unglamorous electrical hardware that takes years to arrive.
Big Tech Is Taking Matters Into Its Own Hands
Faced with all this, tech companies have stopped behaving like ordinary electricity customers and started becoming energy players in their own right.
Microsoft and the nuclear revival. Microsoft made headlines by backing the restart of a dormant reactor at Three Mile Island in Pennsylvania, signing a 20-year deal to buy the electricity it produces, largely so its data centres have a dedicated, reliable supply that isn’t fighting for scraps on a crowded local grid.
Google and the mini-reactors. Google has taken a related route, backing next-generation Small Modular Reactors — factory-built, smaller nuclear units that can, in principle, be dropped in close to where the power is actually needed, sidestepping long-distance grid bottlenecks almost entirely.
How Do You Actually Fix the Grid?
There’s no single fix, but the energy sector is working the problem from a few different angles at once.
Smarter power lines. Rather than only building new pylons, grid operators are increasingly using software — sometimes AI itself — to squeeze more capacity out of lines that already exist. One technique, dynamic line rating, tracks real-time weather so lines can safely carry more power on cold, windy days when the cables run cooler.
Better locations. Instead of forcing the grid to bend around the data centre, developers are increasingly building where the power infrastructure already exists — next to older power stations that already have thick, high-capacity connections in place, saving years of waiting.
Big battery parks. Large-scale battery storage can soak up cheap, abundant renewable power during quiet periods and release it back during peak demand, smoothing out the load rather than requiring new capacity to be built for the worst moments alone.
Could More Efficient Chips Just Solve This?
It seems logical enough: build chips that use half the electricity, and the problem shrinks with it. Unfortunately, that’s not really how it tends to play out. Economists call this the rebound effect — make a technology cheaper and more efficient to run, and people generally respond by using a lot more of it, not less. A chatbot might use noticeably less power per question than it did a year ago, but if ten times as many people are now asking it questions, total electricity use still climbs regardless.
Frequently Asked Questions
Will AI actually cause the world to run out of electricity?
No. A genuine global blackout caused purely by AI is highly unlikely. The real issue is regional, not global — specific areas with a high concentration of data centres will simply find themselves unable to connect new ones to an already-stretched local grid.
Why are tech companies so fixated on nuclear power?
AI data centres need to run 24/7. Solar and wind matter a great deal, but the wind doesn’t always blow and the sun does set. Nuclear supplies large amounts of steady, round-the-clock power without the carbon emissions of a coal or gas plant.
Is the grid really slowing AI down?
Yes, measurably. The IEA estimates that grid limitations and connection delays put roughly 20% of planned global data-centre capacity at risk of delay between now and 2030.
The AI Race Is Now an Engineering Race
For years, the limits on artificial intelligence were defined by code, algorithms and data. Those digital limits are fading, only to be replaced by the far more stubborn physical realities of infrastructure. AI cannot run on venture capital and hype alone — it needs electricity every single second of its existence, arriving safely through real copper cable and real steel transformers.
The next big winners of the AI boom might not turn out to be software developers at all. They might just be electrical engineers.