Analysts warn hyperscale AI compute could nearly double global data‑center electricity use

Data centers are turning into power stories. Not side stories — core stories. The new thing here is not just that AI uses a lot of electricity. It’s that the scale is now big enough to reshape how hyperscalers buy power, where they build, and which energy projects get financed first. The clearest marker came from the IEA in April 2025: global data-center electricity demand is on track to more than double from 415 TWh in 2024 to about 945 TWh by 2030, with AI doing most of the pushing. (iea.org) ### Why are data centers suddenly an energy problem? Classic cloud data centers were already large loads, but AI clusters are a different animal. Training and inference stacks pack far more chips into each hall, which raises both server power draw and cooling demand. That is why the IEA expects AI-optimized data centers to more than quadrup(iea.org). (iea.org) ### What does 945 TWh actually mean? Basically, it means data centers stop looking like a niche utility customer and start looking like a country-scale load. The IEA frames 945 TWh as slightly more than Japan’s current annual electricity consumpti(iea.org)nt. (iea.org) ### Why does “hyperscale” matter so much? Because a handful of companies can move entire power markets. When Microsoft, Google, Amazon, or Meta decide to expand a region, they are not buying a little green branding. They are trying to secure multi(iea.org)rom around 30 GW in 2025 to 90 GW or more by 2030. (mckinsey.com) ### So what are they buying now? More direct, bigger, longer contracts. In February 2026, TotalEnergies signed two long-term PPAs to deliver 1 GW of solar capacity to Google’s Texas data centers over 15 years — about 28 TWh in total. That is not a symbolic renewable credit deal. It is infrastructure-sized procurement tied to real load growth. (totalenergies.com) ### Why isn’t solar alone enough? Because AI loads do not care whether the sun is out. Data centers need reliability every hour, and the catch is that cheap megawatt-hours are not the same thing as firm capacity. That is why the ma(totalenergies.com)scribing a move away from plain fixed-price PPAs toward structures built around delivery certainty. (pv-magazine-usa.com) ### Is that why nuclear is back in the conversation? Yes — especially existing nuclear. Microsoft’s 20-year deal with Constellation helped justify restarting Three Mile Island Unit 1, adding about 835 MW of carbon-free capacity if the project returns as planned in 2028. Amazon went even bigger in June 2025, (pv-magazine-usa.com) (constellationenergy.com) ### What about SMRs? They are still more promise than volume, but investor attention makes sense. Amazon signed SMR-related agreements in October 2024, and the logic is obvious — hyperscalers want clean power that is both local and always o(constellationenergy.com) (aboutamazon.com) ### Bottom line The real warning is not “AI uses a lot of power.” Everyone knew that. The warning is that electricity is becoming a gating input for AI growth itself. If compute demand keeps climbing, the winners may be the companies that secure power first — not just the ones that secure chips.