A sponsored briefing from Utility Dive, based on a Siemens webinar with NVIDIA and Dominion Energy, says AI factories are changing the scale and speed of electricity demand. Traditional data center racks used 15–25 kW, but current AI racks can require roughly 230 kW and future systems could approach 1 MW each. Loads are also clustering in areas like Northern Virginia, arriving before transmission, generation, and interconnection infrastructure can be permitted and built.
The report argues that deterministic load forecasts are no longer enough. AI facilities may be delayed, resized, relocated, or phased, while renewable output, electrification, and transmission constraints add uncertainty. Utilities should pair deterministic studies with probabilistic and scenario-based planning that weighs project probability, ramp rates, hourly and seasonal demand, storage performance, and permitting timelines. The goal is to find decisions that hold up across many possible futures.
Flexibility could help. Some AI training workloads can be shifted or throttled, though time-sensitive inference may have strict requirements. Even partial flexibility might ease stress on constrained periods, improve use of existing assets, and defer some investments, but utilities need dependable data on which loads can move and by how much. At the same time, AI-enabled planning tools can automate repetitive studies and expand analytical capacity, giving engineers more time for high-value decisions. The report emphasizes that physical buildout remains necessary, but smarter, faster planning and better use of existing capacity are equally important.