AI
The fundamental problem lies in a severe timeline mismatch between how quickly a data centre can be constructed and how long it takes to build grid infrastructure.
“ To have a proper conversation, we needed to talk to all of the people in this ecosystem, including the data centre and utility industries,” Anuja says.“ What we quickly figured out was that the challenge was a mismatch in timing. Data centres needed power within about a two-tothree-year window for the growth they were anticipating. Those 20 years have been 5MW to about 100MW in scale. A data centre that is 100 MW can’ t get interconnected in a two-year window.”
As facilities scale up to support massive AI models, the power requirements look more like municipal infrastructure.“ When data centres go from being a 100 or 200MW load to a GW-scale load, we are now talking about a load that is the size of a small city,” Anuja points out.“ There is no way the grid can accommodate the size of a small city getting interconnected in two years.”
Today, the standard interconnection timeline for utilities serving data centres has stretched to between seven and ten years. Through DC Flex, which boasts 70 participating members, operators are exploring how onsite generation can keep facilities running when grid capacity is tight.
“ EPRI recognises the cost – including emissions – when it comes to technology,” Anuja says.
94 August 2026