New paper reduces AI data center curtailment by 69% with smart grid coordination

A new paper from researchers Xin Lu and Qianwen Xu addresses a critical challenge in the energy-AI intersection: how to integrate gigawatt-scale AI data centers (AIDCs) into the power grid without expensive infrastructure upgrades. The work, published on arXiv (2605.14109), formalizes the coordination between AIDCs and transmission system operators (TSOs) under emerging 'connect-and-manage' practices that allow immediate grid connection at the cost of real-time curtailment during grid stress.

The authors develop physical models for both sides of the point of common coupling. On the AIDC side, they decompose workloads into frontier training, batch training, and inference serving subclasses sharing on-site battery storage, capturing different temporal flexibilities. The transmission network is modeled with DC power flow and budget-constrained demand uncertainty. Because the TSO's acceptance mechanism is opaque to the AIDC, they propose a three-layer hierarchical architecture: a learning-based planning layer generates power requests, the TSO evaluates via robust acceptance, and a single-step execution optimizer enforces feasibility. Case studies on the IEEE 39-bus system with Australian market data show curtailment drops from 9.1% to 2.8% while preserving 98.1% frontier training workload, batch training serving as the primary elastic resource, and on-site batteries proving vital for curtailment buffering.