Emerald AI, NVIDIA, National Grid, EPRI, and Nebius demonstrated that AI data centers can autonomously adjust power consumption during peak electricity demand while preserving high-priority workloads, NVIDIA announced.
The trial took place at Nebius’s AI factory in London using the Emerald AI Conductor Platform. Researchers ran production-grade workloads across a cluster of 96 NVIDIA Blackwell Ultra GPUs connected by the NVIDIA Quantum-X800 InfiniBand platform. GPU telemetry was gathered at seconds-level intervals through the NVIDIA System Management Interface.
EPRI and National Grid simulated grid stress scenarios, including lightning strikes, sustained periods of low wind power generation, and sudden surges such as the “TV pickup” phenomenon. In the simulated surge—modeled on the 1-gigawatt demand spike that occurred during the Euro 2020 England-Germany football match when viewers boiled kettles at half-time—the cluster ramped down power while keeping critical tasks operating at full throughput.
During emergency load-reduction trials, the cluster cut power by approximately 30 percent within 40 seconds. Emerald AI recorded 100 percent alignment across more than 200 power targets issued by EPRI and National Grid. Lower-priority computing jobs were temporarily slowed to accommodate the reductions.
“We did tests that go beyond the ones that have been done so far in the U.S. because we tested not just the GPUs, but also the CPUs and everything that sits around it — as well as the total power consumption of the IT equipment,” said Steve Smith, group chief strategy officer of National Grid.
Varun Sivaram, founder and CEO of Emerald AI, noted that load flexibility allows facilities to connect to grids faster by using existing capacity rather than waiting for infrastructure built around worst-case firm demand.
Grid signals are converted into workload adjustments across GPUs and racks using NVIDIA DSX Flex software alongside the Conductor Platform. A related tool, NVIDIA DSX MaxLPS, manages power dynamically within the facility to maximize token output within fixed energy limits.
The London demonstration followed earlier proof-of-concept tests in Arizona, Virginia, and Illinois. Emerald AI and NVIDIA plan to deploy the system in production at the Aurora AI Factory in Virginia, which is scheduled to open this year.
