Plans for AI datacentres in the US, UK and mainland Europe include 181GW of proposed on-site gas-fired capacity, first reported by Theguardian. BloombergNEF estimates the planned rollout could emit about 345m metric tonnes of carbon dioxide equivalent a year in the US if it goes ahead, as grid-connection delays drive developers to seek power on site.
The proposed capacity is almost four times Great Britain’s total peak electricity demand, and 90% of planned on-site gas-fired power stations are in the US. BloombergNEF’s estimate for annual US emissions is equivalent to the greenhouse gas emissions from about 72.8m fossil fuel cars.
Mark Daly, BloombergNEF’s head of datacentres and emerging tech, said AI’s economics make speed to power one of the most important priorities in defining constraints on datacentre developers. He said the climate outcome also depends on how quickly grids connect facilities and whether power systems add cleaner generation alongside demand.
Developers have claimed their gas-fired power supplies would be temporary while they wait for grid connections. Daly said the economics could encourage developers to continue generating energy once the plants are online, creating a longer-term fossil-fuel lock-in risk. BloombergNEF says connecting new datacentres to the grid is the more sustainable option, and faster connections are crucial to limiting emissions and that risk.
Planned UK datacentres made more than 100 inquiries in 2024-5 to connect to gas as a result of delays securing access to the electricity grid, according to Future Energy Networks, the representative body for UK gas networks. Wapseys Wood in Buckinghamshire and Quest Park in Bedfordshire are seeking to build their own on-site gas-fired power stations after long waits for grid access.
On-site fossil-fuel generation is likely to be repeated at many more UK sites after chancellor John Healey gave datacentres the green light to build and run temporary fossil-fuel power stations while waiting for grid connections, as part of a so-called datacentre covenant.
BloombergNEF projected AI datacentre electricity demand would almost quadruple by 2035 to 2,354 terawatt-hours, equivalent to 7% of global electricity demand.