arXiv study: Nuclear fusion could power AI data centers cost-effectively

As AI workloads explode, data centers are projected to consume a much larger share of U.S. and global electricity, straining grids and decarbonization goals. A new preprint from researchers Layla Araiinejad and Vineet Jagadeesan Nair, posted on arXiv (2608.10454), makes the case that nuclear fusion—long dismissed as decades away—could actually be the ideal power source for hyperscale AI. The paper compares fusion against intermittent renewables, fission, and firmed renewable systems across capacity factors, levelized cost of electricity (LCOE), grid interconnection constraints, and deployment pathways.

The authors argue that magnetic confinement fusion concepts, at Nth-of-a-kind commercial scale, may hit cost parity with advanced fission and firmed renewables for always-on AI training and inference workloads. Crucially, they propose co-locating fusion plants directly with data centers, eliminating transmission bottlenecks and improving grid resilience. Fusion's inherently safer profile and reduced radioactive waste also give it better long-term social and political viability than fission. The study concludes fusion should be prioritized in both policy and industrial deployment planning for next-generation computing infrastructure.