Power-Flexible AI Data Centers: A New Paradigm for Grid-Responsive Compute

Overview

The work is a 2026 arXiv preprint (arXiv:2606.25098; DOI 10.48550/arXiv.2606.25098) submitted to IEEE. Its authors include teams linked to Emerald AI, EPRI, National Grid, NVIDIA and Oracle. The architecture links grid signals, workload scheduling and GPU power telemetry.

On a 130 kW NVIDIA Blackwell Ultra cluster in the UK, the system followed National Grid and EPRI dispatch events, including zero-notice emergency reductions, multi-hour curtailment and repeated dispatches, while preserving throughput for priority jobs. It achieved 100% compliance across 200+ distinct National Grid power events. In a surprise emergency dispatch the cluster reduced power by approximately 30% within 40 seconds; other tests achieved 40% reduction within approximately a minute.

Sustained curtailment tests requested 10%-40% reductions for two to ten hours; high-priority jobs kept nearly full throughput because flexible, lower-priority jobs were capped or delayed. In a second experiment, 10% of live inference traffic was migrated from an H100 cluster in Virginia to Illinois during a Dominion Energy winter peak event; a 375 W per-GPU limit with a 15-minute ramp-down and three-hour hold shifted about 3.1 kW to the Chicago cluster.

A prior field trial cited reduced AI cluster electricity consumption by 25% during peak demand while maintaining job service levels. The authors argue that such capabilities could let data centers connect as non-firm or flexible loads and so speed interconnection. The results rest on a small cluster and a limited set of experiments.

Key points

1. Achieved 100% compliance across 200+ distinct National Grid power events on a 130 kW cluster of NVIDIA Blackwell Ultra GPUs.
2. In a surprise emergency dispatch the cluster reduced power by about 30% within 40 seconds; other tests achieved 40% within about a minute.
3. Sustained curtailment tests requested 10%-40% reductions for two to ten hours; high-priority jobs kept nearly full throughput.
4. Migrated 10% of live inference traffic from Virginia to Illinois; a 375 W per-GPU limit with a 15-minute ramp-down and three-hour hold shifted about 3.1 kW to the Chicago cluster.
5. A prior field trial cited reduced AI cluster electricity consumption by 25% during peak demand while maintaining service levels.
6. The authors argue the capability could let data centers connect as flexible loads and speed interconnection.

Using it in design

• Add a 'flexible load' profile in the grid-connection module: committed reduction depth (%), response time (s), duration (h) and recovery, derived from the workload priority mix.
• Use it to evaluate non-firm or flexible interconnection options and to size how much firm capacity (plus UPS/BESS) is truly needed.
• Record requirements for GPU-level power-cap telemetry and a dispatch interface in the controls specification.

Who should read it

Electrical engineers designing grid interconnection and controls, AI infrastructure planners, and energy analysts.

Publication details

Authors

Chris Williams; Philip Colangelo; Ayse Coskun; Ethan Levine; Andy Neale; Ciaran Roberts; Shayan Sengupta; Nikhil Shirolkar; Varun Sivaram; Sarah Soares; Ethan Tiao; Scott Underwood; Daniel Wilson; Frank Sharp; Luke Wainwright; Harry Petty; Scott Wallace; Brandon Records

Publisher

arXiv preprint (Emerald AI / EPRI / National Grid / NVIDIA / Oracle; submitted to IEEE)

Year

2026

Document number

arXiv:2606.25098

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2606.25098

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation