The reason for that massive disparity, unlike in the United States, where data centres are being opposed by public campaigns and political opposition, is the pipeline to get things built. Engineering is the hold-up.
“The first challenge is simply connecting the data centre,” says Keith Bell, professor of electronic and electrical engineering at the University of Strathclyde in Glasgow. Depending on a facility’s size and the voltage at which it wants to connect, the first question is whether the local substation has the physical space for the switchgear, including circuit breakers and disconnectors, and transformers needed to step between voltage levels.
But it’s not just about individual projects. The UK has an issue where the wider network may struggle to deliver the power needed. “It’s like the layers of an onion as you go out,” Bell says. A data centre drawing a few hundred megawatts might connect to the distribution network at 132kV, but that substation may need reinforcing where it meets the transmission network at 275kV or 400kV, he adds.
Even where the space exists on the network, the massive demands from data centres require kit that can sometimes be a challenge to make compatible. “If you need a big transformer that will connect between 132kV and 400kV, or between 33kV and 132kV, these typically take a couple of years or maybe three years to procure, to manufacture and install,” says Bell. In the south-east of England – where most data centres have clustered to keep latency low for their customers – land is at a premium and substations are hard to extend. Compact gas-insulated alternatives could be deployed, says Bell, but rely on sulphur hexafluoride, an electrical insulator that is also a greenhouse gas tens of thousands of times more potent than CO2. Because of that, it can be tricky to gain approval.
Faced with multi-year waits, some developers, particularly in the US, have turned to building their own on-site gas turbines. But Bell is unconvinced that’s the solution. “These are expensive and take time to procure as well,” he says. “It’s also not really a solution from the perspective of emissions.”
There may be a middle path for engineering. Stephen Somerville, managing director of the power consulting business at Morson Praxis, points out that what takes the longest time is the industry’s insistence on a firm, full-capacity grid supply. “This takes time that they don’t have,” he says. “Instead we could follow a design approach where we use a captive hybrid power plant, with a mix of battery storage, gas engines and a lower-rated second grid supply, mixed in with some demand-side reduction.”
That could work and ensure the system keeps going despite potential disruptions. “The battery storage would have islanding capability, so in the event of an upstream grid failure there is a load-shedding system,” Somerville explains. “The BESS [battery energy storage system] carries the load while the gas engines spin up, and the system switches over to the lower-capacity external supply.”
Taking that approach would also solve a less-discussed engineering headache when it comes to AI data centres: they produce pulsed loads as they switch on and off their models for training and inference (the answering of queries). They can move “from 100% to zero in like a second, and then some seconds later going back up to 100%,” he says. At the scale of a 100MW-plus facility, that can make grid voltage flicker for neighbouring users – something that needs to be avoided.
However, where data centres are built matters as much as how. Because Britain’s grid hasn't kept pace with the wind farms built in Scotland and offshore, the NESO routinely pays generators in the north to switch off while firing up gas plants in the south. “These constraint costs are significant these days – a couple of billion pounds a year, maybe going up,” says Bell. The solution is simple: “Build the new demand next to where the generation is.”
The government’s Delivering AI Growth Zones policy paper, published in November, recognises that. It offers electricity discounts of up to £24 per MW/h in Scotland, £16 in Cumbria and £14 in the north-east for a 500MW data centre.
But any new supply needs to be flexible to meet the changeable demand of data centres. AI training runs, unlike cloud services, unfold over weeks and could potentially run when clean power is plentiful. “Flexible data centre loads would be part of that bigger mix,” says Bell. “That would be really useful at a system scale.”
Want the best engineering stories delivered straight to your inbox? The Professional Engineering newsletter gives you vital updates on the most cutting-edge engineering and exciting new job opportunities. To sign up, click here.
Content published by Professional Engineering does not necessarily represent the views of the Institution of Mechanical Engineers.