Overview
- Over five days in December 2025, a Nebius facility near London ran more than 200 simulated grid events using Emerald AI software and successfully hit requested load reductions each time.
- Measured responses included shedding 30% of load in 30 seconds and sustaining reductions for up to 10 hours during low-wind periods, while critical workloads continued uninterrupted.
- Partners on the study included National Grid, Nebius, Emerald AI, NVIDIA and the Electric Power Research Institute.
- Trial partners say scaling this approach could return more than 2GW of capacity to the grid during peaks, easing constraints and helping integrate variable renewables.
- Emerald AI outlined three flexibility methods: pausing or slowing non-urgent runs, migrating tasks between data centers with modest latency trade-offs, and prioritizing workloads through continuous monitoring.