Microsoft, Amazon, Google lock long-term power

Electricity is becoming the real scarce input in AI. Chips still matter, obviously. But the bigger shift now is that Microsoft, Amazon, and Google are starting to treat power the way they treated cloud servers a decade ago — as something you lock up years in advance if you want to keep growing. That is the news hiding underneath a bunch of separate deals. Microsoft signed a 20-year power purchase agreement in September 2024 to help restart Three Mile Island Unit 1 in Pennsylvania. Google signed a nuclear development deal with Kairos Power in October 2024 that aims for 500 MW by 2035. Amazon expanded its nuclear partnership with Talen in June 2025 to supply up to 1,920 MW for AWS data centers through 2042. ### Why are power deals suddenly the story? AI data centers are not normal server farms. Dense GPU clusters draw huge amounts of electricity and then need even more infrastructure to cool and stabilize that load. So the bottleneck has shifted. For a lot of new projects, the question is no longer “Can you get Nvidia boxes?” It is “Can you get firm power fast enough to turn the site on?” ### What did Microsoft actually lock up? Microsoft’s deal is the clearest signal. Constellation said the contract will support the restart of Three Mile Island Unit 1 — now renamed the Crane Clean Energy Center — and add about 835 MW of carbon-free power back to the grid. A 20-year term tells you this is not a hedge or a branding exercise. Microsoft is planning around long-lived AI demand. ### What about Google? Google’s move is earlier-stage but just as revealing. Its Kairos agreement is a master development deal for a fleet of advanced reactors, with the first commercial deployment targeted for 2030 and total capacity of 500 MW by 2035. Basically, Google is not just buying power from an existing plant. It is helping pull future generation into existence. ### And Amazon? Amazon went even bigger on headline megawatts. Talen said in June 2025 that it would supply up to 1,920 MW from the Susquehanna nuclear plant to Amazon Web Services data centers in Pennsylvania through 2042. That kind of duration matters because AWS is building campuses that need predictable power over decades, not just enough electricity to bridge the next demand spike. ### Why not just use the grid? Because the grid is slow. Interconnection queues, transmission upgrades, transformer shortages, and local permitting can drag out a data-center build even after the land is bought and the building is designed. The catch is that GPUs can be delivered faster than substations. So developers are prioritizing places with available capacity, on-site generation, or a utility willing to build around them. ### How big is the demand wave? Big enough that the numbers are getting silly. JPMorgan’s Jamie Dimon said the top hyperscalers could raise AI-related capital spending from about $450 billion in 2025 to $725 billion in 2026. Meanwhile, EPRI said data centers could consume up to 9% of U.S. electricity generation by 2030. Those two facts fit together semiconductors. ### So what changes now? Power-rich sites get more valuable. Utilities with spare capacity get more leverage. And smaller developers without long-term contracts get squeezed, because they are competing against companies willing to sign 20-year deals and even bankroll new generation. Inference here — but a pretty safe one — is that AI infrastructure growth will cluster where electricity can be guaranteed, not where land is cheapest. ### Bottom line? The AI buildout is starting to look like an energy story wearing a software costume. Microsoft, Amazon, and Google are telling the market the same thing with their checkbooks: if you want the next wave of compute, you need to own the power first.