The live project was carried out at a London data centre in December 2025. The results suggest data centres could become flexible grid assets rather than inflexible sources of demand as the digital economy expands, according to project partner National Grid.

The other project partners were Emerald AI (a portfolio company of National Grid Partners), the US Electric Power Research Institute (EPRI), Dutch technology company Nebius and US ‘Big Tech’ company Nvidia. Over a five-day period, Emerald AI’s software platform, Emerald Conductor, was used to manage a cluster of 95 Nvidia graphics processing units (GPUs) operating at Nebius’ new London facility.

Today, most large data centres operate with fixed ‘always on’ electricity demand. ‘As AI adoption grows and more facilities seek to connect to the grid, that fixed-demand model risks increasing network constraints and lengthening connection times,’ explained National Grid. ‘Demonstrating the ability of data centres to flex their power demand shows how they can ease constraints and unlock grid capacity, rather than be an extra source of inflexible demand.’

Using real-time signals from the electricity system, the platform adjusted power consumption without disrupting critical workloads.  

The London trial delivered significant demand reductions in testing scenarios. In several cases, electricity consumption was cut by more than one third in less than a minute, with reductions reaching up to 40%. Crucially, the computing workloads running on the GPUs continued uninterrupted during these adjustments.

The project also tested the system’s response to simulated grid emergencies. During scenarios designed to mimic lightning strikes or power plant failures, the AI platform was able to shed approximately 30% of load within 40 seconds, helping to stabilise the grid and reduce the risk of wider blackouts.

Another important capability demonstrated was peak smoothing. The software successfully responded to sudden spikes in electricity demand, such as those that occur during major televised events, by temporarily reducing data centre load to counterbalance the surge in consumer electricity use.

The system also maintained sustained load reductions for up to 10 hours, a feature that could support grid operation during prolonged periods of system stress such as extended low-wind conditions or extreme heat.

‘The success of the trial demonstrates that AI data centres can move from being a source of electricity constraint to a controllable grid asset,’ said National Grid. ‘By flexing demand in real time, they can help manage peaks, make better use of existing infrastructure, and support the connection of different sources of energy to the grid.’

With more than 6 GW of data centre capacity expected to be deployed in the UK by 2030, the project partners estimate that power-flexible facilities could potentially make more than 2 GW of capacity available back to the nation’s electricity system when needed.

The performance data from the London pilot will inform future rules for so-called ‘power-flexible’ data centre connections. Such frameworks could allow facilities that agree to adjust their demand on request to secure faster or larger grid connections, say the project partners.

The new power-flexible reference design, if adopted nationwide, ‘could unlock an estimated 100 GW of additional capacity on the existing US electricity system’, according to the project partners. This would be roughly equivalent to one fifth of the country’s annual electricity consumption.  

The trial also provides a potential operational blueprint for a much larger project already under development in the US. Project Aurora, a 96 MW ‘AI Factory’ operated by Nvidia in Virginia, aims to ‘establish a reference design and certification standard for power-flexible AI infrastructure’, according to company and its project partners Emerald AI, EPRI, Digital Realty and PJM.