Data centres set to consume a third of US commercial power by mid-century, EIA warns

Data centres set to consume a third of US commercial power by mid-century, EIA warns New EIA projections frame AI-driven electricity demand as the dominant long-term supply challenge facing the United States and allied energy markets. Data centre servers consumed an estimated 7% of all US commercial-sector electricity in 2025, according to the US Energy Information Administration's Annual Energy Outlook 2026. By 2050, that share is projected to rise to between 22% and 33% of commercial building electricity use, a shift that will force utilities, grid operators and fuel markets to rethink long-run capacity assumptions.¹ That matters because the range of outcomes is enormous. Server consumption alone is projected to reach between 446 billion kilowatthours and 818 billion kilowatthours by 2050, a spread that reflects genuine uncertainty about how quickly AI workloads scale and how efficiently the industry adapts. Neither end of that range is trivial for any fuel that feeds baseload generation.¹ The same AEO2026 recorded US total energy production hitting 107 quadrillion British thermal units in 2025, a 3.4% increase from the previous record set in 2024. Industrial gas consumption averaged a record 23.6 billion cubic feet per day last year, 1% more than the record 23.4 Bcf/d reached in 2023. Demand growth is not a future problem. It is already visible in the data, compounding.¹ The power mix shift is already apparent in Texas. In the grid operated by the Electric Reliability Council of Texas, solar generation is expected to reach 78 billion kilowatthours in 2026, compared with 60 billion kilowatthours for coal. Whether renewables alone can absorb the data centre load curve, with its flat, around-the-clock draw, is a harder question.¹ Canada has given one answer. Bruce Power signed a memorandum of understanding with SaskPower on 2026-04-16, under which the Ontario operator will share its experience in developing large-scale nuclear reactors with the Saskatchewan utility. Bruce Power runs some of the world's largest operating CANDU units, and the arrangement signals that Canadian provinces are looking to proven domestic expertise rather than waiting on small modular reactor timelines to address medium-term capacity needs.² Australia is having a parallel but more contested version of the same conversation. The country spans an area comparable to the continental United States but holds a population of roughly 26 million, concentrated on the eastern seaboard and in Western Australia. That geographic profile is cited by nuclear advocates as an argument for utility-scale reactor development. The country remains, for now, without a civilian nuclear programme.³ Australian energy users are raising concerns through different channels. At the Energy Users Association of Australia National Conference held on Wednesday (2026-05-27) in Brisbane, industry representatives discussed their current concerns about the National Electricity Market, including the pace of the coal-to-renewables transition and reliability during the changeover period. The forum suggested that large commercial and industrial users are not yet confident the transition timeline aligns with their load planning horizons.⁴ The EIA's AEO2026 is a reminder that long-run energy outlooks are only as useful as the demand assumptions underneath them. A more aggressive data centre build-out produces a radically different gas, nuclear and grid-storage requirement than the base case. The more immediate signal to watch is whether the Bruce Power-SaskPower memorandum of understanding converts into a formal development agreement, and at what cost. That number, when it emerges, will tell markets something real about whether new large-scale nuclear can price competitively against the alternatives that utilities are actually buying.^2,1