Decommissioned nuclear plants could be converted into AI data centres, Nvidia says - Engineering and Technology Magazine

Nvidia's proposal to repurpose decommissioned nuclear sites for AI compute is not a real-estate play—it is a vertical-integration declaration. Once B200 racks exceed 40 kW, the build-first-connect-power-later model collapses. Grid interconnection points, cooling loops, and land area—these hidden assets come pre-installed at retired nuclear facilities. The architectural flip: compute embeds into power infrastructure, not the reverse. Upstream, liquid-cooling systems, specialized transformers, and nuclear-grade alloys will see demand shocks within two quarters. Compliance risks are real but mispriced: NRC decommissioning timelines stretch three to five years, and nuclear-liability premiums run two to three times conventional IDC rates. Nvidia's bet is that once energy becomes the binding constraint on AI training, friction costs get absorbed by compute premiums. Strategically, Microsoft and Meta will accelerate power-resource lockups, and Constellation Energy's valuation thesis gets rewritten. For TSMC in Taiwan, China and mainland fabs, the energy constraint plays out differently, but the compute-power coupling paradigm will leak into advanced-packaging thermal design. Within eighteen months, expect three to four US states to pass nuclear-datacenter incentive legislation, turning retired plants into compute real estate. The real long-tail effect will not sit in Nvidia's roadmap—it will sit on utility balance sheets.