Two planned AI campuses would each draw a power station. The new homes wait behind them.
The deal on the economy desk this morning — SpaceX buying the maker of Cursor, partly on rent paid by a rival, Anthropic — looks like a story about money. Underneath, it is a story about electricity. Training and running these models means drawing vast, sustained power from somewhere physical: the Memphis complex Anthropic rents runs to more than 300 megawatts. The race to dominate AI is, at bottom, a race for places to plug in.
Britain is learning what that costs its grid. The National Energy System Operator says about 140 proposed data centres are seeking roughly 50 gigawatts of grid connections — more than the 45GW the entire country drew at its winter peak on 11 February. The connection queue swelled from around 41GW in late 2024 to roughly 125GW by mid-2025. Data centres took about 2 per cent of Britain's electricity in 2024; official scenarios have that climbing toward 8 to 16 per cent within a decade.
The collision is already physical, and it lands on housing. In west London, where a cluster of data centres takes nearly a fifth of the area's power, the grid is so congested that some finished housing developments have been warned they may wait until 2037 to be fully connected — the same year National Grid hopes to have added 7GW to the area. Two proposed campuses, at Elsham in Lincolnshire and Cambois in Northumberland, are each sized close to a gigawatt: roughly one nuclear reactor's output, each.
There is also a cost that never shows up in the wholesale price. Cornwall Insight expects non-commodity charges — the cost of building and reinforcing the network — to make up close to 60 per cent of a typical business electricity bill this year. So the grid build-out can push bills up even as the price of gas softens. The AI boom is weightless in the brochure and very heavy on the wire.
It need not only be a burden. National Grid trials have cut data-centre demand by a third within seconds without spoiling the computation, and a centre that powers down at the hardest hours could help pay for a clean grid rather than strain it. The industry's own rule is the "three Ps": power, planning and ping. Britain has the demand and the engineers; what it lacks is speed — government calls this "strategic demand" and promises AI Growth Zones while real projects wait years for a connection.
This is the part of automation no one photographs. The work may be done by software, but the residue is intensely material — turbines, substations, a fifteen-year connection queue, an estate left dark — and it settles on British ground whether or not a single one of these companies is British. We are hosting the engine of a race we do not own.
The work AI removes leaves a very physical residue — power draw, grid queues, homes deferred — on British soil, regardless of who owns the AI. Hosting the infrastructure of the automation without owning the automation is its own kind of exposure, and its own version of the question of who holds the value.
- Electric Insights Q1 2026 - NESO: 140 data centre proposals seek ~50GW (= GB peak); 15-year waits; flexible demand; three Ps
- E&T (Mar 2026) - Ofgem queue 41GW (Nov 24) -> 125GW (Jun 25); GB peak 45GW on 11 Feb 2026; 140 schemes could need 50GW
- EnergyCosts (May 2026) - NESO: DCs ~2.4GW / 7.6TWh in 2024 (~2% of GB), modelled 20-41TWh by 2035; Cornwall Insight non-commodity ~60% of business bills 2026; AI Growth Zones
- TechRadar - west London DCs delaying housing grid connections to as late as 2037; 29 DCs ~18% of London power; National Grid +7GW to West London by 2037
- Harper (Apr 2026) - Elsham (Lincs) and Cambois (Northumberland) campuses each ~1GW (~a nuclear unit); 8-10% of demand by 2030
- Computer Weekly (May 2026) - UKPN half-hourly data, 80 DCs in region; Ofgem Demand Connections Reform deadline 13 Mar 2026