The Energy Wall Facing Frontier Models: Why Big Tech is Buying Nuclear Power Plants for AI Data Centers
The race to build frontier AI models is colliding with a far older constraint than chips or talent: electricity. Microsoft, Amazon and Google are now locking in 20-year nuclear power deals because the next generation of training clusters is pushing data-center demand into the gigawatt class, where grid interconnection queues, transmission bottlenecks and carbon constraints become binding. The result is a new industrial logic: compute is no longer just a software problem, but a power-asset acquisition strategy. That shift is visible in deals tied to Three Mile Island and Talen Energy, where Big Tech is seeking firm, round-the-clock baseload supply to support always-on AI workloads. The thermodynamic reality is unforgiving: as model size, parameter count and training runs scale, so do cooling loads, power density and the cost of downtime. Nuclear is not a perfect answer, but for hyperscalers facing constrained grids and volatile renewables, it is increasingly the only asset class that can match frontier AI’s appetite for reliable megawatts.