Grid Integration of Gigawatt-Scale AI Data Centers under Connect-and-Manage

Abstract

The paper addresses how massive AI data centers (AIDCs) can connect to power grids without upfront transmission-infrastructure upgrades, via emerging "connect-and-manage" interconnection practices — the facility accepts real-time power-import reductions (curtailment) during grid-stress periods in exchange for skipping the years-long queue for a conventional firm-capacity interconnection.

Key Contributions

1. AIDC load decomposition into three operational classes with distinct flexibility profiles — frontier training, batch training, and inference serving — plus shared on-site battery storage that each class can draw on differently.
2. Hierarchical, three-layer control architecture: (i) learning-based power-request planning by the data center, (ii) robust acceptance evaluation by the transmission system operator (TSO), (iii) single-step real-time execution optimization.
3. Physical grid modeling: DC power-flow representation incorporating generator constraints and demand uncertainty, used to evaluate whether the TSO can safely accept a given AIDC power request.
4. Establishes explicit information boundaries between data center and TSO — the protocol is a sequential request/acceptance exchange, not a single centrally-optimized dispatch, which matters for real-world deployability where the two parties don't share full internal state.

Results

• Curtailment reduced from 9.1% to 2.8% under the proposed coordination framework versus a naive connect-and-manage baseline.
• 98.1% of frontier-training workload preserved despite real-time curtailment exposure — batch training absorbs most of the curtailment burden as the "grid-responsive" workload class, protecting latency/deadline-sensitive frontier training.
• On-site battery storage provides curtailment buffering via active discharge (to ride through a curtailment event) and charge deferral (to avoid drawing power when the grid signals stress).
• Tested on the IEEE 39-bus system using Australian electricity-market data.

Methodology

Sequential request-acceptance protocol between AIDC and TSO with explicit, limited information sharing; robust optimization used by the TSO for the acceptance decision; learning-based planning used by the AIDC for the request-generation stage.

Limitations

• The paper does not explicitly enumerate limitations, but the framework's reliance on learning-based prediction and robust-optimization guarantees implies sensitivity to historical-data quality and to how well the IEEE 39-bus/Australian test case generalizes to real transmission topologies (e.g., PJM, ERCOT) where connect-and-manage is actually being deployed.
• Validated only in simulation against a synthetic/test grid, not a live utility interconnection.

Full Content

This is a companion paper to Battery-Assisted Operation of Hyperscale AI Data Centers (arXiv:2605.14105, same lead author Xin Lu, submitted the same day) — together they cover the mechanism (this paper: how connect-and-manage interconnection itself works, and how curtailment risk is allocated between the data center and the grid operator) and the response (the companion paper: how on-site batteries let the data center absorb that curtailment risk operationally). Read together they are the most direct research-grade treatment yet of the specific mechanism now emerging in the wild at PJM, ERCOT, and other US grid regions to let gigawatt-scale AI load connect years faster than a firm-capacity interconnection would allow — directly relevant to the transmission/interconnect bottleneck this KB's LENS.md already flags as a required numeric cell ("US and EU interconnect queue size"). The 9.1%→2.8% curtailment result and the 98.1% frontier-training preservation figure are the two numbers worth carrying into any datacenter-power writeup: they quantify, for the first time in this KB, what "the data center eats some curtailment so it can skip the queue" actually costs in workload terms.

Source: Grid Integration of Gigawatt-Scale AI Data Centers under Connect-and-Manage (arXiv:2605.14109) by Xin Lu, Qianwen Xu. Submitted 2026-05-13.