Datacentres can in theory reduce bills.

They're a large, mostly well-behaved grid load which lets you spread fixed costs over more units of electricity.

The problems start when they are being built by well connected entities that can leverage their power to simply demand that other people pay more to give them a discount.

Very similar to the shopping around factories do. Usually a factory brings jobs and other good stuff, but you don't get the factory unless you do X, Y and Z in competition with all the other possible locations. That can lead to a winners curse where you're effectively subsidizing the factory and losing out overall if you "win" the auction.

It's true that all else being equal new power demand lowers the fixed cost per unit of marginal energy generated. But if the data centre requires new grid and storage capacity to handle its load and those fixed costs are charged to the entire electricity consumer base, there is a trade-off.

Also, all else is not equal because the UK is already grasping towards a stretch target of a >95% low-carbon power system by 2030 which has over-loaded supply chains, sacrificed generation cost to speed, and created a queue to connect to the grid that outpaces the annual rate of connections by over an order of magnitude. Adding huge blocks of new incremental demand to the system is quite a challenge, in this context.

> They're a large, mostly well-behaved grid load which lets you spread fixed costs over more units of electricity.

Is that true of AI datacenters? I recently read an article that talked about one that dynamically would spin up gas turbine generators based on compute load.