The IEA puts 424 of the next 530 TWh of data center load inside the US and China

Split the 530 TWh the projection adds between 2024 and 2030 ^[1]. Nearly 80% of it, about 424 TWh, sits in two countries ^[6][2]. The agency calls the American increase a 130% jump from 2024 ^[7]. That puts the 2024 US base near 185 TWh, about 44% of the world's 415 TWh ^[3]. China takes most of the remainder, roughly 184 TWh ^[4].

Getting from 415 to 945 in six years takes 14.7% a year ^[5]. The pace since 2019 has been about 12% ^[4]. Hold that and 2030 comes in at 819 TWh, some 126 TWh under the projection, so about a quarter of the projected increase depends on the rate picking up ^[6]. The IEA attributes the acceleration to AI workloads running on accelerated servers ^[5]. For scale, the report puts 945 TWh close to the combined annual electricity consumption of France and Germany, and at nearly 3% of all global electricity demand ^[3][2].

A data center takes 18 to 24 months to build. The transmission and generation behind it take five to ten years to permit and construct, between 2.5 and 6.7 times as long ^[12][7]. The 945 TWh is a consumption forecast, which means it assumes the power shows up; the agency's findings call for expedited permitting and more diverse generation ^[14]. The report as summarised by cryptobriefing does not compare chip lead times with grid lead times, so the case for power being the tighter input rests on those build schedules alone.

Money is moving faster than either. Data center spending was roughly $500 billion in 2024, nearly double the 2022 figure ^[9]. Keep that annual rate flat through 2030 and $3 trillion of spending sits against 530 TWh of new load, about $5.7 billion per additional TWh of annual consumption ^[8].

In my view the asset being priced here is an energised interconnection, because a developer holding one can start in 18 months and a developer without one is waiting on a five-to-ten-year queue ^[12]. The counter-thesis is in the projection itself: 15% a year is an IEA assumption, and if efficiency per accelerated server improves faster than workloads grow, the TWh curve flattens and hardware goes back to being the scarce input. Faster permitting at the supply end would get to the same place from the other direction. cryptobriefing argues that bitcoin miners competing for the same cheap power, interconnection and regulatory treatment may find themselves in a tighter market ^[15]; that argument is the publisher's, not the agency's.