The findings were published on Wednesday as global leaders gathered in Turkey for the 31st UN climate summit, where increasing electrification, ideally from clean sources, is a central part of the discussions.
Developers of artificial intelligence datacentres plan to build on-site gas-fired power stations on a huge scale to meet their energy demand, research by BloombergNEF has found.
The climate stakes are large. The report said the outcome would depend not only on how much electricity the facilities used, but on how quickly grids could connect them and whether power systems added cleaner generation alongside that demand.
That proposed capacity is almost four times the total peak electricity demand of Great Britain.
By 2035, the electricity demand of AI datacentres was predicted to almost quadruple to 2,354 terawatt-hours, equivalent to 7% of global electricity demand, the report said.
AI datacentres use a massive amount of electricity to power servers, cool equipment and run facility systems.
Mark Daly, head of datacentres and emerging tech at BloombergNEF, said the extraordinary economics around AI were making speed to power one of the most important priorities for defining constraints on datacentre developers.
Daly said developers of future datacentres claimed their gas-fired supplies would be temporary while they waited for grid connections. Once the plants were online, he said, the economics could encourage them to keep generating, creating a longer-term fossil-fuel lock-in risk.
The datacentres are among the first to be fast-tracked through the government’s nationally significant infrastructure project regime, which allows developers to deliver major infrastructure rapidly and bypass locally elected planning authorities.
The report said ensuring new datacentres were connected to the grid was the more sustainable option.
Developers have framed on-site gas as a bridge around connection delays. The research instead treats the proposed plants as a climate risk that could persist well beyond the wait for the grid.