An explainability-aware adaptive energy planning framework for AI data center expansion

Artificial intelligence (AI) is turning data centers into a strategic electricity and governance problem, with demand uncertain, geographically concentrated, and split between schedulable training and latency-sensitive inference. We develop an adaptive energy planning framework that states supply for AI data centers as a multi-criteria portfolio-allocation problem over an explicit feasible set (firm-capacity coverage through capacity credits, deployability limits set by lead times and technology readiness, emissions limits) and reports Pareto-efficient portfolios rather than a single optimum. The demand model adds a closed-form treatment of the energy cost of explainability, in which mandatory explanation of regulated decisions creates a schedulable auditable-inference sub-load sized by the governance regime. We implement the framework for the United States and China with a transparent scoring rubric, a full portfolio enumeration, and sensitivity analyses over demand growth, technology costs, lead times, preference weights, and the explainability parameters. In the 2030 base case, every efficient U.S. portfolio pairs natural gas with firmed renewables and none built from resources deployable by 2030 reaches 0.10 tCO 2 /MWh, whereas China’s efficient set spans coal-backed and nuclear-plus-renewable portfolios depending on the weight given to environmental performance. Counting physical rather than contractual delivery drops a certificate-backed grid purchase from first to below the top 1,800 portfolios in China. The explainability overhead is method dependent, from negligible for exact or amortized explainers to tens to hundreds of gigawatt-hours per year for perturbation-based methods at one million audited decisions per day. Governance choices thus become quantifiable planning variables.