Unexpected efficiency achievements by Chinese artificial intelligence (AI) company DeepSeek have cast a shadow over a bullish narrative on booming US electricity demand in the coming decade to power data centers running AI software.
Share prices for US independent power producers, natural gas producers and gas pipeline companies fell sharply at the beginning of the week as investors feared DeepSeek's achievement implied significantly less electricity might ultimately be needed to run and train AI models than has been expected. This greater efficiency "calls into question the significant electric demand projections for the US," as the investment case for independent power producers and most integrated utilities is "entirely dependent on data centers," US bank Jefferies said in a note to clients this week.
DeepSeek's apparent ability to achieve comparable results to some major US AI companies using far less computing power — and thus far less electricity — may also be bad news for what is widely expected to be the main fuel source to generate incremental power for AI this decade: natural gas. EQT, one of the largest US gas producers by volume, has called growing power demand from planned data centers the "cornerstone" to its "natural gas bull case." Large US gas pipeline companies like Williams, operator of the Transcontinental pipeline, have also touted recent forecasts showing surging demand for gas-fired power, as greater gas generation would require greater pipeline capacity to move those incremental volumes from wellhead to generator.
DeepSeek's achievement could even cast doubt on the investment case for nuclear power, which has been recast as something of a silver bullet for major technology companies looking to secure zero-emission electricity to enable their AI development efforts. While investors have generally assumed significant premiums for nuclear power, to the tune of more than $100/MWh, new demonstrated efficiencies might cause those assumptions to be questioned, Jefferies said. A loss in power demand for AI data centers may also undercut the investment case for next-generation small modular reactors (SMRs), into which tech companies like Google and Microsoft have poured substantial capital.
Revising the revisions
News of DeepSeek's efficiency achievements are a shock to prevailing expectations for surging US power demand in the coming decade, when those expectations have already been substantially revised over the past year, following decades of stagnant power demand.
US grid operator PJM, which serves 65mn customers and is the largest US electric grid, on 24 January released a report showing significant upward revisions in its peak seasonal power demand projections. Peak summer power demand in PJM's territory in the mid-Atlantic was projected to surge to 210GW in 2035 and 229GW in 2045, substantially steeper than PJM's load forecast just one year earlier, which showed peak summer power demand in PJM rising to 177GW in 2034 and 191GW in 2039. Consultancy firm McKinsey in November forecast US data center power demand to reach 606TWh by 2030, up from 147TWh in 2023. Under this scenario, data centers at the end of the decade would comprise 11.7pc of total US power demand.
If efficiency gains in AI reduce power demand as much as some investors fear, those big forecasts might require big revisions. But efficiency improvements can go two ways — they can reduce demand for fuel, or simply increase output. In the case of AI, more efficient operations could be exploited to accelerate the development of more powerful AI models — using the same amount of power that was previous expected, but to far greater effect. That latter explanation is why, "despite uncertainties," FactSet head of power markets Matthew Hoza tells Argus he remains "bullish" on power demand growth in the coming years.
"With AI's increasing integration into company tech stacks and its growing presence in daily life through AI agents, we anticipate continued growth in AI adoption and the resulting power needs," Hoza said.
