National Grid has carried out the first trial of flexible electricity usage by a UK datacentre, in conjunction with operator Nebius. The trial used artificial intelligence (AI)-powered datacentre management software from Emerald AI’s software on a bank of 96 Nvidia Blackwell Ultra high-performance graphics processing units (GPUs) at a Nebius datacentre near London.

Over five days in December 2025, more than 200 real-time simulated “grid events” were sent to the site to test the Emerald software’s ability to dynamically adjust the datacentre’s power consumption.

Emerald AI’s platform was able to adjust power use to the requested level and cut demand by up to 40% while critical workloads ran as normal.

Key results included successfully reacting to spikes in demand during half time at football matches, followed by load-reduction requests for up to 10 hours that demonstrated an ability to help the grid navigate periods of low wind or extreme heat, and simulated a system stress event that saw it shed 30% of load in 30 seconds to help maintain grid resilience.

According to the partners involved in the trial, such capabilities could enable AI datacentres to add more than 2GW of capacity back to the grid when needed.

The aim is that AI datacentres can avoid being simply a source of electricity constraint to being more controllable in relation to the electricity grid, by managing peaks, making better use of existing infrastructure, and supporting the connection of different sources of energy to the grid.

“Most electric networks, most electric power systems, operate with probably 30% of capacity in place a year; there’s lots of capacity in the system, it’s a small number of hours a year when we’re at peak,” said Steve Smith, president of National Grid Partners, speaking at the Economist Impact Sustainability Week event in London.

“So, the trick is how you do it,” said Smith. “Because if you can throw more electrons at a fixed-cost system, you don’t need to put more infrastructure in, and the rates come down for everyone else.

“If you’re doing a small number of hours and you’re stretched, if we say, can you actually moderate your load when we need you to, then we don’t need to build lots more capacity.”

Also speaking at the Sustainability Week event, Varun Sivaram, chief executive of Emerald AI, said the trial showed that AI hardware at the Nebius datacentre could consume energy flexibly at a moment’s notice.

“When we got the signal in the middle of the night, we were able to reduce power within 30 seconds by over a third,” said Sivaram. “That’s also going to be the case with renewable energy, when there’s low wind, for eight hours, and the AI factory can reduce its consumption in such a way that we protect the critical workloads that run at 100% throughput.”

Sivaram explained that there are three ways to achieve flexibility of power consumption for AI workloads. The first is to slow some down or pause them. “Maybe a fine-tuning model run that doesn’t need to finish right this second, but it can be delayed by an hour,” he suggested.

The second way, he said, is by moving AI workloads. “You expect your answer from AI pretty soon, but we may be able to move it, as we did with a move between two different Oracle datacentres at the rate of 10 milliseconds of latency. There is a little bit of a latency penalty, but not relevant for that workload,” said Sivaram.

The third way, he said, is to monitor the datacentre to achieve flexibility. Here, Emerald operates as software intelligence to operate AI workloads – that can include by tagging them as different priorities – in an optimal way to give the grid what it needs while protecting the integrity of the workloads for the user.