Electricity, not chips, is now the thing standing between a hyperscaler and its next model. The GPUs can be bought. The transformers, the substations, the gigawatts of baseload — those take years, and years are the one thing nobody building AI infrastructure in 2026 has to spare.
The numbers behind that statement are unusually well documented for once.
176 → 325-580 TWh — US data center electricity use, 2023 actual versus the US Department of Energy’s projected range for 2028.
4.4% → up to 12% — Data centers’ share of total US electricity consumption over the same window.
9.3+ GW — Combined nuclear capacity committed across just three hyperscaler deals disclosed since mid-2025: Microsoft, Meta, and Amazon.
+12% — CBRE’s projected 2026 rise in the cost of securing data center capacity across Europe’s five largest hubs.
The grid can’t move fast enough
Harvard Kennedy School’s Belfer Center frames this as a genuine mismatch in build timelines: data center construction can be completed in a few years, while grid interconnection and transmission upgrades routinely take a decade or more once permitting is factored in. That gap is the entire story of AI infrastructure right now — every other constraint (capital, chips, even talent) has effectively been solved by throwing money at it. Power hasn’t, because you can’t out-spend physics and permitting timelines.
The result is a market splitting in two: projects that secured power and grid access early are proceeding on schedule, and everyone else is stuck in an interconnection queue watching their hardware orders age.
Nobody has said this more plainly than the people running the budgets. In a joint, on-the-record interview, Microsoft’s Satya Nadella put it directly: the constraint isn’t a compute glut, it’s power — going as far as to say Microsoft has GPUs sitting in inventory it can’t install because it doesn’t have “warm shells to plug into.” OpenAI’s Sam Altman, in the same conversation, framed the other side of it: without a real breakthrough in energy technology, AI does not get to its next stage — and anyone locked into expensive long-term power contracts today could get burned if a genuinely cheap energy source shows up at scale.
Why nuclear is showing up in the same press releases as GPUs
Three deals disclosed since mid-2025, each announced directly by the companies involved, add up to a real number:
- Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1 (now the Crane Clean Energy Center) — 835 MW, targeted for 2027.
- Meta announced agreements with Vistra, TerraPower, and Oklo in January 2026 unlocking up to 6.6 GW: more than 2.1 GW from Vistra’s existing Beaver Valley, Perry, and Davis-Besse plants, up to 2.8 GW plus 1.2 GW of storage from new TerraPower Natrium units, and up to 1.2 GW from an Oklo microreactor campus in Ohio.
- Amazon expanded its supply agreement with Talen Energy to up to 1.9 GW from the Susquehanna nuclear plant by 2032, under a 17-year, $18B power purchase agreement — disclosed in Talen’s own SEC filing.
That’s over 9.3 GW from three companies alone, before counting anyone else’s deals. Separately, the US Nuclear Regulatory Commission issued TerraPower’s Natrium reactor in Kemmerer, Wyoming its construction permit in March 2026 — the first approval for a non-light-water commercial reactor design in more than 40 years — and Canada’s nuclear regulator authorized construction of the first BWRX-300 small modular reactor at Ontario Power Generation’s Darlington site.
None of this replaces natural gas as the near-term bridge — turbines can be sited and turned on in a fraction of the time a nuclear restart or a new SMR design takes — but it’s a real signal that hyperscalers are now underwriting new nuclear supply directly rather than just buying existing clean power off the grid.
The bottleneck has a geography, and it’s not evenly priced
Power isn’t a single global commodity, and that’s reshaping where AI infrastructure actually gets built. The International Energy Agency’s Electricity 2026 report puts electricity prices for energy-intensive industries in Europe at roughly double US levels and about 50% above China and India — a gap that has held for two years running. The same IEA report projects global data center electricity consumption will keep climbing sharply through 2026. On top of that baseline cost gap, CBRE’s own research projects the cost of securing data center capacity in Europe’s five biggest hubs (Frankfurt, London, Amsterdam, Paris, Dublin) will rise another 12% in 2026 alone, driven by supply constraints against a wall of new AI-related demand.
Capacity is gravitating toward wherever power is cheapest and fastest to contract, and Europe’s AI ambitions are running into its energy prices before they run into any model or chip constraint.
The industry is now regulating its own externality
The most telling move of the year wasn’t a chip announcement. In March 2026, the White House announced that Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI had signed the Ratepayer Protection Pledge — a commitment to build, bring, or buy the generation they need and cover the cost of the power-delivery upgrades their data centers require, rather than let those costs land on household electricity bills. Reading between the lines, that’s seven companies conceding that their power draw has become a public-utility-scale problem rather than a private procurement one.
What this means if you’re not a hyperscaler
For everyone operating a level below the Amazons and Microsofts of the world — mid-size AI companies, infrastructure investors, enterprise buyers signing multi-year compute contracts — the practical takeaway is simple: compute availability is no longer just a pricing question, it’s a siting and power-contract question. Where your provider’s power comes from, and how firm that supply is, now matters as much as which chip is in the rack.
The next twelve months will separate the infrastructure bets that priced in the power constraint from the ones that didn’t.
Sources: US Department of Energy — DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers, Lawrence Berkeley National Laboratory — 2024 United States Data Center Energy Usage Report, Belfer Center, Harvard Kennedy School — AI, Data Centers, and the U.S. Electric Grid, BG2 Pod — Satya Nadella interview, Constellation Energy — Crane Clean Energy Center announcement, Meta — Meta Announces Nuclear Energy Projects, Talen Energy — SEC filing on Amazon supply agreement, US Department of Energy — NRC Issues Construction Permit for TerraPower’s Natrium Advanced Reactor, Canadian Nuclear Safety Commission — Darlington New Nuclear Project, IEA — Electricity 2026: Prices, IEA — Electricity 2026: Demand, CBRE — Data Centre Capacity Pricing in Europe Set for Double-Digit Growth, The White House — Fact Sheet: The Ratepayer Protection Pledge
Opinions are my own.