Where the Electricity Comes From

The demand side of this problem has its own collection: how much electricity data centres use, how fast it is growing, and what regulators have done about it. This page is the other half — what is actually being built to serve that demand, and how long the physical equipment takes to arrive. Almost everything here reduces to lead time.

The clearest measure of what AI demand is worth to a generator is that it made a closed nuclear reactor worth reopening. Three Mile Island Unit 1 shut down in September 2019, prematurely, for poor economics — it had run at a 96.3% capacity factor in its final year, so the problem was price, not performance. Exactly five years later Constellation announced a restart as the Crane Clean Energy Center: 837 megawatts of capacity, approximately 835 returning to the grid, under a twenty-year power purchase agreement with Microsoft described as the largest in the company's history, expected in 2028 and subject to Nuclear Regulatory Commission approval. A plant uneconomic at merchant prices becomes economic when a single buyer will contract for two decades.

The stranger form of the same trade is buying reactors that do not exist. In October 2024 Google agreed to purchase electricity from a fleet of Kairos Power small modular reactors — up to 500 MW, first unit in 2030, deployments through 2035 — under a master plant development agreement with milestone-based accountability. Google described what it was buying as an orderbook of reactors: a commitment to a production line rather than to a plant, which is how a technology with high per-unit engineering cost is meant to become cheap. The design is fluoride salt-cooled, described as the first non-water-cooled reactor approved for construction in the United States in more than fifty years. For scale, Google has signed over 115 clean energy agreements totalling more than 14 GW since 2010, so 500 MW arriving in the 2030s is a strategically large bet and a numerically small one.

The fastest conventional route to firm capacity is a gas turbine, and those are now rationed by factory slots. Within the first quarter of 2026 GE Vernova's volume under contract rose from 83 to 100 gigawatts and its slot reservation agreements — reservations of future manufacturing capacity held ahead of firm orders — from 43 to 56, with 21 gigawatts of new agreements signed in three months. Roughly 10 gigawatts of 2029 capacity was available at the start of the quarter; by the end, roughly 10 remained across 2029 and 2030 together. Lead time on a heavy-duty machine is about three years. Data centres are 20% of that book and traditional utilities and industrials the other 80%, so this is not purely an AI phenomenon — but AI changed the rate, and new-order pricing is running 10 to 20 points higher per kilowatt than the previous quarter.

Generation is useless without the equipment that connects it, and that has the longest queue of all. Second-quarter 2025 survey averages: power transformers 128 weeks, generator step-up units 144 weeks, switchgear 44. Demand since 2019 rose 274% for generator step-up transformers and 119% for power transformers; prices rose 77% and 45%. Underneath it sits a replacement problem with nothing to do with AI — 40 million U.S. distribution transformers past their expected service life, 55% of in-service units over 33 years old. A transformer is not a sophisticated device. It is a large one, made of steel and copper by a small number of factories, and it sits on the critical path of everything else.

That last claim is disputed, and the dissent is carried here rather than dropped. A transformer broker quoted in the same reporting says flatly that there is no shortage, that standard substation units can be delivered in 12 to 14 months across all voltage classes, and that the scarcity is exaggerated to drive pricing — locating the bottleneck in utility qualification processes and approved-vendor lists rather than in factory capacity. Everyone arguing has an interest, including the analysts selling the research that reports a 30% deficit. Nothing read here settles it, but it changes what an average means: 128 weeks is consistent both with a real capacity shortfall and with a market where faster routes exist that most large buyers cannot use.

What is missing is named. The most legally consequential thing in this subject — the regulatory dispute over whether load co-located behind the meter at a nuclear plant should pay transmission charges — is absent entirely, because that source was not retrieved. So is on-site generation, the third strategy alongside contracts and grid connection, and the one with the sharpest local air-quality consequences. A larger small-modular-reactor commitment than the one described here was announced and not read, so this page understates the scale of what has been promised.