by Chris Dunlap
AI data centers are increasingly being treated as inflexible, firm loads in U.S. electricity grid resource adequacy planning, leading to capacity procurement requirements that surpass what the interconnection queue can actually provide. To address this, I present an eleven-parameter framework organized along three operational axes: spatial migration across grid regions, temporal deferral of non-interactive workloads, and dynamic voltage-frequency scaling (DVFS). This framework quantifies the flexibility of AI inference in terms relevant to capacity-accreditation for resource adequacy planning. Using data from four years of day-ahead pricing across five U.S. grid regions and 49.4 million production inference requests, I find that commitment depths reach 24.6% for mixed-use facilities and 40.0% for inference-dominant facilities at the single-facility level. These values decrease only slightly to 24.1% and 38.5% at the 10-GW fleet scale, since deferral and DVFS mechanisms contribute additively as they saturate. Finally, most remaining uncertainties such as destination utilization, contractual workload pinning, and post-stress recovery coordination are found to be institutional challenges, not physical limitations.