Delphi Digital: Electricity is a greater constraint on AI computing power deployment than chips

On October 29, the cryptocurrency research institution Delphi Digital released a report titled "The AI Boom Has a Power Problem," authored by Accelxr. The report states that deployable AI computing power depends on chip supply, as well as sites that are powered, cooled, and connected to the grid. Electricity is harder to supplement in the short term than chips, as chip production capacity can expand quarterly, while new power generation and transmission take years.

The report cites data indicating that the PJM capacity auction has cleared at the price cap for three consecutive times, with the last two still falling short of reliability targets by 6.6 to 6.8 GW, marking the first time such a gap has occurred in the history of this auction. The scale of large load queues in Texas ERCOT increased from 63 GW at the end of 2024 to 474 GW in June, with about 90% being data centers; the governor requested an audit of the data centers in the queue in August, and ERCOT subsequently suspended power approval for data centers and cryptocurrency miners of 75 MW and above, including 17 projects that had passed stability studies.

The report also states that U.S. data centers are expected to account for about 4.7% of the nation's electricity consumption in 2024, with LBNL predicting an increase to 9.5% to 15.3% by 2030, and EPRI estimating a maximum of 17%. In a 1 GW facility, chip depreciation is about $2.8 billion to $3.3 billion per year, equivalent to about $8 million per day. SemiAnalysis predicts that by 2028, the U.S. will have over 40 GW of data center capacity utilizing behind-the-meter power sources; the delivery cycle for large gas turbines is about five years, with GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries forming an oligopoly. Wood Mackenzie expects gas turbine prices to reach $600 per kilowatt by the end of 2027, nearly three times the level of 2019.