Within the span of a single week, two of the world's largest cloud computing operators locked in two-decade contracts for nuclear power. Amazon secured 690 megawatts from Constellation's Calvert Cliffs facility in Maryland, which includes nearly 200 megawatts of additional generating capacity anticipated between 2030 and 2032. Meanwhile, Google finalized an agreement to inject 890 megawatts of new nuclear capacity into the United States grid. Constellation itself is pouring over $4.3bn into upgrading 11 different reactors spanning New Jersey, Pennsylvania, and Illinois. Notably, these arrangements do not involve constructing new reactors; instead, they purchase increased output from existing facilities, some of which predate the technology firms buying their power.

Such agreements would have been difficult to imagine just a few years ago. Nuclear power faced a severe political and public backlash following historic incidents at Three Mile Island, Chernobyl, and Fukushima, leading nations like Germany to phase out atomic energy entirely. Today, however, an undamaged reactor at Three Mile Island is being restored to active status under a 20-year commitment from Microsoft, while the US Department of Energy extends billions in loans to resurrect shuttered plants. The Palisades plant in Michigan, deactivated in 2022, is scheduled to resume generation by early 2027, marking an unprecedented commercial restart for a decommissioned nuclear facility.

For years, the technology sector believed that efficiency gains would solve the impending power crisis. Industry conferences routinely highlighted improvements in performance per watt, noting that newer artificial intelligence hardware performed calculations far more efficiently than older generations. Yet, these advances failed to curb overall electricity consumption. Every efficiency milestone was quickly absorbed by larger workloads, more intensive training runs, continuous background processing, and expanded video generation. This phenomenon mirrors historical observations like Jevons' paradox, where more efficient steam engines ultimately increased coal consumption by lowering the effective cost of power. Consequently, the International Energy Agency projects that global data center electricity consumption will more than double to approximately 945 terawatt-hours by the end of the decade.

Faced with unyielding demand, tech companies have pivoted from trying to shrink consumption to aggressively buying up available supply. However, financial resources cannot bypass physical timelines. Constructing a large traditional reactor typically takes a decade or longer, and while firms are investing in experimental small modular reactors, widespread deployment remains years away. Faster solutions, such as plant restarts and capacity upgrades, still require years to materialize. This creates a fundamental mismatch: the specialized chips powering artificial intelligence today may become obsolete long before the 20-year power purchase agreements backing them expire, leaving unanswered questions about how the industry will handle potential drops in demand.

This dynamic shifts the burden far beyond corporate balance sheets. A significant portion of these power deals reside within the PJM grid region, which serves millions of customers across multiple states. PJM utilizes capacity auctions to procure adequate power resources, with costs ultimately rolling down to residential consumers through monthly utility bills. Recent auctions have hit regulatory price caps, and market monitors indicate that data center expansion has driven tens of billions in capacity expenses. Some environmental organizations warn that typical households could face substantial monthly increases in utility bills if current trajectory lines persist. These escalating expenses are already provoking political pushback, with various states implementing specialized rate classes, minimum billing rules, or legislative debates to shield everyday ratepayers from the energy demands of the technology sector.