Turning AI Data Centers into Grid-Interactive Assets: Results from a Field Demonstration in Phoenix, Arizona

Overview

This document is a joint arXiv preprint by Emerald AI, NVIDIA, Salt River Project and EPRI (arXiv:2507.00909, 2025), written by Philip Colangelo, Ayse K. Coskun and others. It reports the first field demonstration, under the EPRI DCFlex initiative, of software-only demand response for a 256-GPU NVIDIA A100 cluster in a commercial hyperscale cloud data center in Phoenix.

A controller (Emerald Conductor) and a simulator orchestrate training, fine-tuning and inference jobs using GPU power capping/DVFS and job pausing within flexibility tiers that bound the allowed performance loss. The tiers are Flex 0, no performance reduction; Flex 1, up to 10% average-throughput reduction over 3-6 h; Flex 2, up to 25%; and Flex 3, up to 50%. No hardware changes or energy storage are needed.

In the demonstration, during peak grid events (APS and SRP system peaks on 1 and 3 May 2025) the cluster cut power by 25% for three hours, with 15-minute ramp-down and ramp-up to avoid snap-back. In a re-enactment of CAISO's August 2020 emergency the system delivered an initial 15% curtailment followed by a further 10% step-down. 33 experiments (each 3-6 hours) with 212 jobs were run; the simulator predicted power with 4.52% RMSE relative to average experiment power.

The discussion notes that, according to cited studies, reducing AI data-center power by roughly 25% for up to 200 hours a year (far less than 1% of the time) could unlock up to 100 GW of new U.S. data-center capacity. The demonstration used a single cluster and a limited number of events; the results should not be generalised beyond that scope.

Key points

1. During peak grid events (APS and SRP system peaks on 1 and 3 May 2025) cluster power was cut 25% for three hours, with 15-minute ramps to avoid snap-back.
2. Flexibility tiers: Flex 0 no reduction; Flex 1 up to 10% over 3-6 h; Flex 2 up to 25%; Flex 3 up to 50%.
3. In a re-enactment of CAISO's August 2020 emergency the system delivered 15% curtailment then a further 10% step-down.
4. 33 experiments (each 3-6 hours) with 212 jobs; the simulator predicted power with 4.52% RMSE.
5. 256-GPU NVIDIA A100 cluster; no hardware changes or energy storage.
6. Reducing power ~25% for up to 200 hours a year could unlock up to 100 GW of new U.S. capacity (cited studies).

Using it in design

• Provide a demand-response calculator that maps a workload mix and flexibility tiers (allowed throughput loss) to achievable curtailment depth and duration, with ramp-time settings to avoid rebound.
• Use the result to evaluate flexible/non-firm interconnection capacity (e.g. curtailment of about 25% for up to 200 h/yr) in site studies.
• Note that software-only flexibility needs no extra storage but does need GPU power-cap telemetry and scheduler integration.

Who should read it

Electrical designers, energy planners and operators evaluating AI data center grid flexibility and non-firm interconnection capacity.

Publication details

Authors

Philip Colangelo; Ayse K. Coskun; Jack Megrue; Ciaran Roberts; Shayan Sengupta; Varun Sivaram; Ethan Tiao; Aroon Vijaykar; Chris Williams; Daniel C. Wilson; Zack MacFarland; Daniel Dreiling; Nathan Morey; Anuja Ratnayake; Baskar Vairamohan

Publisher

arXiv preprint (Emerald AI / NVIDIA / Salt River Project / EPRI)

Year

2025

Document number

arXiv:2507.00909

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2507.00909

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation