A live trial in London has shown a high-performance AI cluster can reduce electricity demand in real time in response to grid signals, with workloads continuing to run.
Emerald AI, EPRI, National Grid, Nebius and NVIDIA all collaborated to make the trial a reality, which uses AI technology to allow data centres to adjust their power usage in real time, without interrupting critical compute workloads.
The test was run at Nebius’s new data centre in London, using Emerald AI’s Emerald Conductor software to manage a bank of 96 NVIDIA Blackwell Ultra GPUs. Over five days in December 2025, the partners sent more than 200 real-time simulated ‘grid events’ to the site, requesting load reductions to specific levels.
According to the group, the platform responded each time, cutting demand by up to 40% while workloads continued to operate as normal. That could have a significant impact on future grid planning, and could ensure that data centres become part of the solution rather than being vilified for their intense energy usage.
A range of scenarios were tested, including reacting to spikes in demand during half-time of major football matches, following load-reduction requests for up to 10 hours, and simulating a system stress event by shedding 30% of load in around 30 seconds.
Steve Smith, President, National Grid Partners, noted, “As the UK’s digital economy accelerates, there’s concern that data centres could add pressure to an already constrained system. This trial proves the opposite can be true. High-performance data centres don’t have to place additional strain on the grid. With our partners, we’ve shown they can be connected and managed without major new network capacity, flexing their power up or down in real time to support the whole system. This approach will enable us to connect significant new demand more quickly and, help to lower network charges for customers over time.”
Dr Varun Sivaram, Founder and CEO of Emerald AI, added, “This trial demonstrates that AI infrastructure can be a dynamic force for the grid. With dozens of realistic AI workloads running simultaneously, we delivered fast emergency curtailment and sustained, precise peak reduction. The same approach we validated here can be applied to much larger AI factories, as the industry scales.”
The performance data will now be shared with industry, regulators and policymakers to inform future rules for ‘power flexible’ data centre connections, including whether sites could qualify for faster and larger connections where they agree to flex on request.
