Sun, Oct 05, 2025 page7

• OpenAI’s power grab is tone deaf and infeasible

  • By Liam Denning / Bloomberg Opinion
  • • 
    • 
    • 

  OpenAI CEO Sam Altman might be a visionary, but he could use some help reading the room.

  Altman has made yet more headlines for reportedly telling employees he targets 250 gigawatts (GW) of datacenter capacity by 2033. In power terms, that is like building a not-small country, equivalent to roughly one-third of peak demand on the entire US grid. This is the same US grid already morphing into a political third rail amid fast-rising bills and warnings of blackouts, with proliferating datacenters linked to both. A bit more nuance and creativity, plus a dash of realism, is in order.

  Building dispatchable generation — non-renewable and, in practice, natural gas-fired plants — to power Altlandia in under a decade is infeasible.

  Looking back over the past 15 years, the fastest, sustained level of gas-fired deployment has been 12.5GW per year, Sector and Sovereign Research LLC Utilities and Renewable Energy Research cohead Hugh Wynne said. If we somehow achieved double that pace, immediately, it would result in only 200GW over eight years. Even that overstates things, since turbines, unlike datacenters, do not run flat out (so you need more than 1GW of generating capacity for a given gigawatt of datacenter demand).

  Now consider the cost — US$450 billion for the turbines at today’s upwardly mobile prices — and the vast quantity of gas required. Plus all those emissions the hyperscalers claim to care about as well as the fact that competitors also want dozens of gigawatts each to feed their own artificial intelligence (AI) ambitions.

  Altman’s mooted moonshot forms part of a broader AI arms race. Even a quarter of those 250GW would be bigger than California’s all-time peak demand for electricity.

  Datacenters, including for AI, needed about 30GW last year and the consensus of a very broad range of estimates is that this might reach 176GW by 2035, CreditSights analyst Andy DeVries said. This comports with a 121GW backlog of datacenter capacity that is already planned or under construction, as reported by the utilities he tracks. However, those utilities also report expressions of interest from datacenter developers that add up to another 601GW.

  There was a lot of double-counting and wishful thinking in that figure. Utilities, which earn a regulated return on what they build, are motivated to turn as much of that as possible into new power plants, transformers and wires.

  Wynne said that as it is, utilities’ planned capital expenditure for this year through 2027 is US$532 billion, up 91 percent from the prior three years.

  These costs ultimately flow into the bills businesses and households pay. Regulators are touting new tariff structures to ensure datacenters pay their own way.

  However, that is hard to do in practice, given the complexity of ratemaking on grids, and this collision between monopoly utilities serving the public and impatient, deep-pocketed technology giants creates potential for households to end up subsidizing the AI frenzy, Harvard Law School fellow Eliza Martin and university Electricity Law Initiative director Ari Peskoe said in a recent analysis. Indeed, the about 600GW of aspirational capacity, along with Altman’s outsized target, are dangling carrots for utilities to get creative to win business.

  One prominent risk is that forecasts turn out to be too ambitious and we overbuild power infrastructure, either because AI’s capabilities are overhyped or efficiency gains reduce energy needs. This is what happened in the first decade of the 2000s, when we built more than 200GW of gas-fired plants just in time for power demand to flatline and bankrupt much of the merchant generation sector.

  While AI seems intuitively revolutionary, the path to sustainable profits commensurate with the trillions of planned upfront investment remains fuzzy.

  Goldman Sachs Group Inc global equity research head Jim Covello earlier this year said that, “this is the first technology transition in history where we are taking really expensive technology and asking it to replace really cheap solutions.” Even Bain & Co, generally bullish on AI, questions how it would generate the revenue necessary to justify its US$500 billion-a-year capital expenditure.

  An updated version of the Turing test might involve AI not convincing us that it is human, but rather convincing us that it is worth it. Energy is at the core of this. On that basis, Altman’s approach is exactly backwards.

  There is a case to be made that AI could help alleviate energy costs and emissions. Power grids are exceedingly complex exercises in optimization, becoming more complex as demand rises and we incorporate more intermittent wind and solar power. This is the sort of parameter-rich problem AI seems made to solve.

  Moreover, growth in electricity demand is not a bad thing in and of itself. Most of the increase in residential bills over the past decade relates to spending on distribution networks, a fixed cost that could be spread across more kilowatt-hours, as electricity consumption rises.

  For AI to play a useful role, the industry would need to embrace more flexibility in its demands. Agreeing to adjust datacenters’ power consumption for a handful of hours a year could effectively free up 100GW of existing capacity, a widely cited analysis from Duke University said. How is that for a productivity leap?

  However, rather than extolling the potential for efficiency, Altman’s giga-goal betrays an obsession with sheer size, and one that looks tailor made to stoke, not relieve, the economic, physical and political pressures building on the grid.

  Liam Denning is a Bloomberg Opinion columnist covering energy. A former banker, he edited the Wall Street Journal’s “Heard on the Street” column and wrote the Financial Times’ “Lex” column.
• Most Popular

  • 1

    Chinese technicians join airport MRT project

  • 2

    A train for whom? Is a new monorail helpful for Cairo?

  • 3

    Apple is giving Intel’s turnaround some momentum

  • 4

    Uncrewed US craft transits the Taiwan Strait

  • 5

    Starlux to expand long-haul flight offerings