Researchers at the University of Hong Kong have proposed a way for power grids and AI data centers to plan around sudden changes in electricity demand without exchanging their full operating data. In a simulated 14-bus grid, their checkpoint-aware plan avoided load shedding in all 100 tested scenarios; a plan that did not coordinate with the data centers shed load in every scenario, according to a new preprint.
Large AI training runs often save periodic checkpoints: copies of the model state that allow work to resume after a failure. When training computers pause together for a checkpoint, power use can fall sharply, then rebound when computing restarts. A grid operator needs to prepare for those swings, but it does not control the training jobs or normally see their private schedules.
In the proposed three-phase scheme, the grid operator first sends the data-center operator a set of power-use plans that satisfy modeled grid constraints. The data-center operator then adjusts training-computer power settings and routes flexible inference work while staying inside that set. It sends back a planned power trajectory and possible checkpoint windows. Finally, the grid operator schedules generators and reserves against uncertainty in renewable output and checkpoint timing. The exchange uses these limited summaries rather than the parties' full internal models; the paper evaluates coordination, not a deployed data-sharing system.
The main test used a modified 14-bus grid over a simulated day, with one 300-megawatt training center and two 50-megawatt inference centers. Researchers sampled 100 combinations of renewable output and checkpoint events. The coordinated robust plan shed no load in those scenarios, compared with load shedding in all 100 without coordination. A coordinated plan that assumed fixed checkpoint times avoided shedding in only 11. In this setup, dialing down training computers accounted for 84.1% of the data centers' modeled energy reduction, far more than inference adjustments.
In a second test on a modified New York grid, each strategy avoided load shedding because the modeled system had ample generation and transmission capacity. There, coordinating checkpoints chiefly reduced renewable curtailment. Compared with the plan assuming fixed checkpoint times, the robust plan had higher average curtailment but a lower worst-case value. These are scheduling simulations; the study did not measure whether data center operators could supply the planned power changes in live service.
Sources
- University of Hong Kong research preprint, 22 September 2026
- Full paper, including both simulated grid cases
- Illustrative transmission substation in Colorado, not a site in the study: Rsparks3, via Wikimedia Commons, CC0 1.0; unmodified.
