#Summary: **Trial shows AI data centres can participate in demand-response and stabilize the grid, reducing the requirement for grid upgrades**
A collaboration between Emerald AI, NVIDIA, Nebius, National Grid and EPRI has demonstrated that AI data centres can act as flexible grid assets, automatically curtailing power consumption during peak demand periods. In a proof-of-concept trial at Nebius’s London AI factory — running 96 NVIDIA Blackwell Ultra GPUs — the team simulated real grid stress scenarios, including the famous “TV pickup” surge from the Euro 2020 England-Germany match, when millions of kettles switching on simultaneously created a roughly 1 GW demand spike.
The Emerald AI Conductor platform responded to grid operator signals by rapidly ramping down the cluster’s power draw — achieving a 30% reduction in under 40 seconds — while preserving throughput on high-priority AI workloads. Lower-priority jobs were temporarily slowed instead. Across more than 200 power targets and 22 real-time dispatch events, the system achieved 100% compliance.
Critically, the London trial went further than earlier US demonstrations in Arizona, Virginia and Illinois by managing not just GPU power but also CPUs and all surrounding IT equipment. National Grid’s Steve Smith said the results proved the technology’s value for a country eager to attract hyperscaler investment without waiting years for grid infrastructure upgrades.
The implication is significant: if AI data centres can function as demand-response participants — absorbing grid shocks rather than amplifying them — operators can connect large new loads to existing infrastructure more quickly, reduce the need to overbuild for worst-case peaks, and help keep electricity rates stable for ordinary consumers. The next step is real-world deployment at the Aurora AI Factory in Virginia later this year.
1 Comment
#Summary: **Trial shows AI data centres can participate in demand-response and stabilize the grid, reducing the requirement for grid upgrades**
A collaboration between Emerald AI, NVIDIA, Nebius, National Grid and EPRI has demonstrated that AI data centres can act as flexible grid assets, automatically curtailing power consumption during peak demand periods. In a proof-of-concept trial at Nebius’s London AI factory — running 96 NVIDIA Blackwell Ultra GPUs — the team simulated real grid stress scenarios, including the famous “TV pickup” surge from the Euro 2020 England-Germany match, when millions of kettles switching on simultaneously created a roughly 1 GW demand spike.
The Emerald AI Conductor platform responded to grid operator signals by rapidly ramping down the cluster’s power draw — achieving a 30% reduction in under 40 seconds — while preserving throughput on high-priority AI workloads. Lower-priority jobs were temporarily slowed instead. Across more than 200 power targets and 22 real-time dispatch events, the system achieved 100% compliance.
Critically, the London trial went further than earlier US demonstrations in Arizona, Virginia and Illinois by managing not just GPU power but also CPUs and all surrounding IT equipment. National Grid’s Steve Smith said the results proved the technology’s value for a country eager to attract hyperscaler investment without waiting years for grid infrastructure upgrades.
The implication is significant: if AI data centres can function as demand-response participants — absorbing grid shocks rather than amplifying them — operators can connect large new loads to existing infrastructure more quickly, reduce the need to overbuild for worst-case peaks, and help keep electricity rates stable for ordinary consumers. The next step is real-world deployment at the Aurora AI Factory in Virginia later this year.