Original title: Spatial LLM Workload Shifting Needs Foresight: Model Commitment for AI Data Center Operation under Power Grid Constraints
Bojun Du, Hongyang Jia, Tonghui Li, Qingchun Hou +3 · September 2026
Data centers running large language models sometimes run low on electricity. When power runs short, operators want to send computing tasks to other sites. However, current systems wrongly assume that any center with enough computer power can instantly run those moved tasks. In reality, models take time to set up, causing failed transfers and dropped requests. To fix this, researchers created a planning method called model commitment. It plans ahead by deciding when to load copies of models and where to route requests across sites, tracking power limits, prices, and processing delays. In tests using real-world data, this method helped operators serve 100% of incoming requests during changing grid conditions while cutting total operating costs by 29.0%.
Imagine you run a food truck that prepares gourmet sandwiches at multiple locations. Each truck has limited ingredients and can only serve customers in its own neighborhood. When one truck runs out of a key ingredient, you want to reroute customers to another truck. But that other truck might not have the sandwich already prepared—it takes time to set up and start making the specific sandwich. Here, the food trucks are data centers, the sandwiches are AI models, and the ingredients are electricity. The paper plans exactly which sandwich each truck prepares and which customers go where, so no one is left hungry even when ingredients run short.
AI data centers are essential for running large language models, but they consume enormous amounts of electricity. When the power grid is strained, operators currently have no reliable way to move workloads to other data centers without risking service disruptions. This new planning method ensures every user request is served even during shortages, while also cutting operating costs by 29%—saving money and keeping services dependable.
The abstract does not explicitly state limitations. Results are based on case studies using real-world data, not live deployment. The computational complexity of solving the mixed-integer linear program for large-scale systems is not discussed.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Spatial LLM Workload Shifting Needs Foresight: Model Commitment for AI Data Center Operation under Power Grid Constraints — Bojun Du, Hongyang Jia, Tonghui Li et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.