AI demonstrates data centre demand management

The trial, part of EPRI’s DC Flex initiative, answers a key question – can AI help to moderate the surging energy demand of the data centres that are generating it?

The answer is yes, partially at least, and paving the way for the technology to support the grid connection of these hyperscaler data centres by playing a significant role in delivering flexibility to the grid.

The trial led by GB’s National Grid utilised Emerald AI’s software, Emerald Conductor, to manage a cluster of 96 NVIDIA GPUs at cloud provider Nebius’s new data centre in London.

Also of interest
How data centre-driven power demand is shaping big tech power plays

Over five days in December 2025, more than 200 real-time simulated ‘grid events’ were sent to the site to test Emerald Conductor’s ability to dynamically adjust the data centre’s power consumption.

Each time, the power use was successfully adjusted to the requested level, cutting demand by up to 40% in under a minute while critical workloads continued to run as normal, National Grid has reported.

Steve Smith, President of National Grid Partners, comments that as the country’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,” he said.

“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."

Drilling into the results more deeply, several different demand adjustments were explored.

Peak smoothing was demonstrated with the successful reaction to sudden spikes in demand during half time of major football matches.

Long duration shifting was demonstrated following load reduction requests for up to 10 hours, which would help the grid navigate periods of low wind or extreme heat.

System management also was demonstrated with simulation of a stress event, shedding 30% of its load in roughly 30 seconds to help maintain grid resilience.

Based on these results, as the UK prepares for over 6GW of data centre deployments on the grid by 2030, National Grid estimates that the technology could enable AI data centres across the UK to add more than 2GW of capacity back to the grid when needed.

The findings also are intended to serve as the operational blueprint for the Nvidia operated almost 100MW power flexible ‘Aurora’ AI factory being developed in Virginia.

More broadly they are also being integrated into the Nvidia Omniverse DSX blueprint for gigascale AI factories – dubbed a ‘grid aware by design’ approach to advancing these assets with the codesign of the building and its power and cooling requirements with Nvidia’s AI infrastructure stack.

David Porter, EPRI Vice President of Electrification and Sustainable Energy Strategy said that field demonstrations are a unique element of DCFlex, designed to accelerate the adoption of solutions through real world validation.

"With each DCFlex demonstration, we learn more about operationalising data centre flexibility to support AI-driven load growth while ensuring the reliability of the electric system.”

DCFlex advances

DCFLex is a major international initiative to demonstrate how data centres can support and stabilise the grid and develop a common understanding of flexibility between power providers, grid operators and data centre stakeholders.

The workstreams focus on grid-informed flexible data centre designs, utility market and programme structures that advance data centre flexibility, grid planning for large flexible loads, data centre energy supply and distribution connected data centres.

Earlier demonstrations have been launched in Lenoir, North Carolina with Duke Energy and Google, in Phoenix, Arizona with Salt River Project, Nvidia, Emerald AI and Oracle – these two addressing how computational flexibility can address grid needs – and in Paris, France with RTE, Schneider Electric and Data4 – this exploring data centre response to power quality events.

The intention with these is to cover diverse geographies and market structures, data centre size, ownership models and on-site generation technologies, among others.