AI datacenters drive hidden 8% off-peak power cost spike

A research paper by Willa Gutowski, Charles Foltz, Nicholson Koukpaizan, Slaven Peles, and Eve Tsybina, published on arXiv (2607.28833, July 2026), quantifies a less visible consequence of AI-driven datacenter growth: surging costs during off-peak hours. While most attention focuses on peak electricity prices, the authors simulate a congestible 5000-bus system based on a modified IEEE 118-bus grid with fast-ramping datacenter loads and slow-ramping thermal generation. Their coupled simulations show that when datacenters rapidly change power demand, grid operators must preemptively ramp slow units, creating 'latent load pockets' that push marginal costs up by an average of 8% even outside datacenter peak times. In coincident peak scenarios, slow, expensive generators hit 100% loading—a severe inefficiency that standard peak-hour planning misses.

The paper tests two distinct load conditions and finds that the hidden costs stem from the operational decision to keep slow generators spinning in anticipation of datacenter load changes. The authors argue these findings are critical for grid reliability and market design. For utilities, the results underscore the need for faster-ramping resources, energy storage, or demand-response programs tailored to AI workloads. For datacenter operators, the 8% off-peak cost increase represents a real, avoidable expense that could be mitigated by shifting non-urgent compute jobs to times of lower ramping stress or by co-locating with flexible generation. This research adds a new dimension to the AI energy debate: it's not just about peak demand, but the systemic cost of grid flexibility itself. As AI infrastructure expands, coordinating datacenter load ramps with grid capabilities will become as important as total energy consumption.