Nuclear plant costs in the U.S. have repeatedly exceeded projections. Here we use data covering five decades and bottom-up cost modeling to identify the mechanisms behind this divergence. We observe that nth-of-a-kind plants have been more, not less expensive than first-of-a-kind plants. Soft factors external to standardized reactor hardware, such as on-site labor supervision, contributed over half of the rapid cost rise from 1976-1987. Relatedly, reactor containment building costs more than doubled from 1976-2017, due only in part to safety regulations. Labor productivity in recent plants is up to thirteen times lower than industry expectations. Our results point to a gap between expected and realized costs stemming from low resilience to time-and site-dependent construction conditions. Prospective models suggest reducing commodity usage and automating construction to increase resilience. More generally, rethinking engineering design to relate design variables to cost change mechanisms could help deliver real-world cost reductions for technologies with demanding on-site construction requirements.