Follow the Curtailment: Data Center Location as a Grid Coordination Strategy

The U.S. data center industry applies a single siting framework to two fundamentally different workloads. Inference — the real-time serving of AI responses — must sit near users. Its geographic constraint is latency.

Training — the process of building foundation models — has no such constraint. A 500 MW training cluster in rural Nebraska produces the same model artifact as one in Ashburn, Virginia, but at roughly half the energy cost. This paper argues that current training infrastructure is systematically misallocated. Using a seven-factor composite scoring model across all 50 U.S. states, and a derived ratio of renewable curtailment to Fortune 500 density, we show that the states with the strongest case for training co-location — Nebraska, North Dakota, Iowa, Oklahoma, South Dakota — hold less than 1% of installed U.S. data center capacity.

The states that dominate the market score well for inference but poorly for training. The curtailment-to-F500 ratio maps both dimensions simultaneously. When composite model weights shift from customer-proximity to training-optimized, North Dakota rises 36 places in the national ranking, Nebraska rises 18, and Georgia falls 24. Texas converges near the top under both scenarios.

The analysis validates against Microsoft's announced investment pipeline and identifies Wyoming as a possible early signal of deliberate workload-based disaggregation. Two policy interventions would accelerate the shift: real-time wholesale market access for smaller operators in high-curtailment ISOs, and state incentives differentiated by workload type and curtailment-responsive operation.