PowerAtlas: Towards Electricity-Computing Co-Scheduling for Power Systems

Researchers introduced PowerAtlas, an LLM-agent framework for jointly scheduling electricity and computing for data centers.

It was evaluated with a Chinese provincial utility using real data-center data, and the system produced feasible schedules while respecting grid constraints and SLAs.

The team also released ECBench, a 2,000-instance benchmark for electricity-computing co-scheduling.

Results across multiple LLMs showed cost improvements and fewer line-flow violations, highlighting a practical path for AI-driven data-center growth.