Nuclear computing contracts ignite the electricity battle and put Gulf plans before a pricing test
The biggest cloud-computing firms raced to legacy power grids to secure electricity flows for their data centers after recognizing that the real bottleneck for AI expansion lies in generation capacity. Within a week, two of the three largest cloud-computing giants in the world signed twenty-year contracts to purchase nuclear electricity. Amazon acquired 690 MW from the Calvert Cliffs plant owned by Constellation in Maryland, including 190 MW of additional generation capacity expected between 2030 and 2032. At the same time, Google struck a deal adding 890 MW of nuclear capacity in the United States, while Constellation is injecting more than $4.3 billion to upgrade 11 reactors in Illinois, Pennsylvania and New Jersey.
A major paradox emerges in the details of these deals: they do not involve building a new nuclear reactor, but rather purchasing output from existing reactors, some of which predate the technology companies themselves. The move reshapes the energy landscape after decades of waning confidence in nuclear power following historic incidents at Three Mile Island, Chernobyl and Fukushima. Microsoft contracted to buy output from a functioning reactor at Three Mile Island for twenty years, and the U.S. Department of Energy provided loans up to $4.4 billion to support the restart of three closed plants to recover about 2.25 GW, including the Perry Mansfield plant in Michigan slated to return to service by March 2027, which would become the first U.S. commercial nuclear plant to resume operation after entering decommissioning.
During the past decade, the technology sector assumed that improving chip efficiency, shrinking models and advancing cooling would contain electricity consumption, but reality proved the opposite. The surge revived Jevons’ paradox, noted by economist William Jevons in 1865, when higher steam-engine efficiency led to greater coal use because lower cost created new applications. Today, even when the DeepSpeed model demonstrated the ability to train a competing model with a modest reduction in expected computing cost, industry spending did not fall; laboratories instead accelerated plans for larger training runs, longer-thinking models, continuously operating agents and video-generation systems. The International Energy Agency estimates that global data-center electricity demand will double to exceed 945 TWh per year by 2030.
Financial calculations revealed a critical timing gap between hardware lifecycles and energy-investment cycles.Large reactors require one or more contracts to move from decision to actual production, and the small modular reactors that tech firms are investing in will not begin delivering electricity before the early 2030s, with very few having been built worldwide. Even the fast-track upgrade and restart option means waiting until 2027, 2028 or 2029, while AI chips being installed this year may be outpaced by much newer technologies decades before the twenty-year electricity contracts expire, and the industry does not clarify what will happen to those contractual obligations if demand later declines.
The crisis has moved beyond data-center walls to hit ordinary consumers directly, particularly within the PJM network that serves about 67 million people across 13 U.S. states and the District of Columbia. Electricity-capacity auctions have hit the regulatory price ceiling in three consecutive rounds, with data-center demand contributing roughly $29.4 billion in capacity costs across the last four auctions, including $6.3 billion in the most recent auction alone for facilities that are largely not yet built. Estimates from the Natural Resources Defense Council suggest that a typical household in those areas could see a monthly electricity bill increase of up to $70 by 2028, sparking political opposition that led Virginia to create a separate pricing tier for data centers and Ohio to impose a minimum billing floor, while federal regulators in October 2026 declined to require centers to generate their own power and left Google’s initiatives in this area as a fragile voluntary commitment.
The American lesson carries a double strategic weight for anyone planning computing infrastructure in the region.Regional capitals are leading unprecedented investments to attract supercomputing hubs, with Qatar Investment Authority committing to a $20 billion joint venture with Brookfield that starts with an initial phase of about 300 MW and can expand to roughly 1.5 GW, equivalent when completed to about 12 % of Qatar’s total installed electricity capacity. At the same time, Abu Dhabi is targeting plans to build up to five GW of data-center capacity, and Saudi Arabia is planning around two GW through the Humain initiative.
The practical shift imposed by these facts on decision-makers in the Gulf and its surroundings is to settle the alternative-cost question before locking in large commitments, because power earmarked for computing cannot be used simultaneously in other industrial or civil sectors. The region now has an opportunity to resolve issues that have stalled Western markets: set industrial electricity tariffs in advance, require developers to fund additional generation capacity, and craft contract terms that reconcile the rapid obsolescence of technical equipment with long-term agreements, thereby ensuring economic returns that exceed mere energy sales and a limited number of operational jobs. The real challenge is not supplying power to AI centers but protecting public grids and defining operating conditions before commitments become hard-to-undo constraints.