Microsoft, Amazon face power queues

Microsoft and Amazon are running into a very old-economy problem. They can afford the chips, the land, and the buildings — but getting enough electricity to a new AI data center can take years. That is the real story here. The bottleneck is no longer just Nvidia supply. It is grid capacity, transmission upgrades, transformers, and utility approvals. Microsoft’s AI buildout is already running into a broader capacity gap, with demand outpacing available power, cooling, and data-center construction capacity. Why does power matter so much now? Because AI racks are much denser than older cloud gear. A normal enterprise data center used to be mostly a real-estate problem. An AI cluster is an industrial-load problem. Goldman Sachs has been blunt about the shift — new AI chips may consume more power than expected, and the limiting factor is increasingly whether operators can get reliable electricity and cooling on time. The same note points out that solar and batteries can come online relatively quickly, while new gas generation can take around five years. So what are Microsoft and Amazon actually doing? They are moving upstream into power procurement. Microsoft signed a 20-year agreement with Constellation in September 2024 that supports restarting Three Mile Island Unit 1 — now the Crane Clean Energy Center — and would add about 835 MW of carbon-free power to the grid. Constellation said the project is tied directly to Microsoft’s data-center power needs. Amazon has gone even bigger in Pennsylvania. Talen and AWS expanded their arrangement into a 17-year power purchase agreement announced in June 2025 for up to 1,920 MW from the Susquehanna nuclear plant, with deliveries ramping over several years and full volume expected by 2032 at the latest. The deal also includes work around broader Pennsylvania data-center expansion. Why not just connect those campuses to the grid like everyone else? Because the queue is the problem. In markets like PJM, which covers a huge chunk of the Mid-Atlantic and Midwest data-center corridor, interconnection backlogs have become a serious choke point. PJM only just reopened its standard interconnection process after a long pause to clear backlog, and its first reformed cycle drew 811 generation projects totaling 220 GW. At the same time, FERC this week ordered PJM to revise rules for large data centers connecting directly to power plants, which shows how fast AI load is colliding with old tariff structures. That is why you keep hearing about “behind-the-meter” power, co-located generation, and even small modular reactors. The idea is basically to avoid waiting for the full grid to catch up. Some operators are also leaning harder on cooling changes — especially closed-loop systems that use less water, and more aggressive thermal designs — because power and water constraints are now linked. Investors have started pressing Amazon, Microsoft, and Google for more disclosure on both. The catch is that none of these fixes are instant. Restarting a nuclear plant takes years. Building gas turbines takes years. Transmission upgrades take years. Even if a company can finance the whole stack tomorrow, the physical grid still moves at utility speed. That is why this matters beyond two companies. AI is turning cloud builders into de facto power planners, and the winners may be the firms that secure megawatts first — not the ones that order the most GPUs.