A fraction of one percent. That is the share of a single year's data center construction spending that Amazon and Google's small modular reactor deals actually cover, according to The Bulletin of the Atomic Scientists' analysis of the SMR announcements piling up since 2024. Molly Langabeer and M.V. Ramana ran the math on every major deal. The press releases promise gigawatts. The contracts promise feasibility studies. That's it.
Google's October 2024 agreement with Kairos Power is a "Master Plant Development Agreement" targeting 500 megawatts by 2035. Amazon's deal with Energy Northwest pays for an "initial feasibility phase." No reactor gets built under it. Amazon holds an option on power from the project's first four modules, a combined 320 megawatts, with no obligation to buy if the economics sour. X-Energy pulled together a $500 million consortium, TerraPower raised $650 million against a Natrium unit now estimated at $9.4 billion for 345 megawatts, and NuScale's Idaho project died in 2023 when its price tag hit $9.3 billion and utilities walked away. A single 300-megawatt SMR produces about 2.6 terawatthours a year. A gigawatt AI campus needs more than three times that, running continuously, starting now.
Here is the number that matters to anyone speccing a CDU or a cooling tower this quarter: U.S. data center electricity demand hit 312.6 terawatthours in 2025, climbing 12.6 percent a year, while nuclear's share of global electricity generation slid to 8.8 percent, roughly half of what it was three decades ago. None of that changes because a hyperscaler signed a memorandum of understanding. The chillers going into service this year and next run on whatever the regional grid actually delivers, and increasingly that means off-grid gas turbines built specifically to skip the interconnection queue. Nuclear baseload is the slide-deck promise. Gas is the grid reality. Grid capacity is the bottleneck constraining where campuses get built and how their cooling plants get sized.
The one nuclear-and-data-center pairing actually operating today barely resembles the SMR pitch: existing plants like Talen's Susquehanna station, where 96 landowners split $586 million selling land to QTS next door. That is a reactor built decades ago, cooling a data center today. It has nothing to do with a 500-megawatt module that does not exist yet. Electricity from a new nuclear build runs about three times the cost of solar or wind. A review the Bulletin cites found 175 of 180 nuclear projects went over budget and behind schedule.
Nuclear may still power data centers in the 2030s. That is a decade away. Operators commissioning liquid cooling loops and evaporative towers this year cannot price their water and carbon math against a reactor fleet that does not exist yet. Cooling procurement teams are buying against the grid mix they will actually get: gas turbines, curtailed renewables, and whatever transmission capacity utilities manage to interconnect on schedule. Spec the CDU, the chiller plant, and the water budget for that grid.