Particle.news
Download on the App Store

UK Trial Shows AI Data Centers Can Cut Power on Demand Without Hitting Critical Workloads

Participants frame the results as a blueprint for faster grid integration, including a planned 100MW Virginia site.

Overview

  • Over five days in December 2025 near London, more than 200 simulated grid events tested dynamic demand cuts at a Nebius facility.
  • The system reduced draw by up to 40%, including a 30% drop in 30 seconds and a sustained 10% reduction for up to 10 hours, while critical tasks continued uninterrupted.
  • The trial used Emerald AI’s orchestration software on 96 NVIDIA Blackwell Ultra GPUs, with partners including National Grid, NVIDIA, Nebius and EPRI.
  • NVIDIA says the findings will guide a 100MW “power‑flexible” facility in Virginia, positioning its infrastructure as a grid‑aware asset that can modulate demand in real time.
  • Organizers plan to share data with industry and regulators; National Grid Partners says demand‑flexing could speed connection timelines to about two years, though some hyperscalers may resist in favor of firm peak power.