UK transmission owner National Grid has a uniquely British phenomenon to manage: the TV pick-up effect, which sees sports fans rush from their televisions to their kettles in nationwide synchronisation.
After the final whistle of the Euro 2020 match between England and Germany this tea-based surge totalled 1.6GW, more than the output of the Sizewell B nuclear power plant.
While these rapid increases have traditionally been met with additional generation in the form of stored hydroelectric power, in the future these spikes could be accommodated by data centres acting as virtual power plants.
As part of a recent flexibility trial involving National Grid, a new AI factory in London successfully adjusted its energy use to match the inverted demand profile of the Euro 2020 match without impacting its critical workload.
The TV pick up was just one of 22 tests set by National Grid. The data centre was able to successfully meet 200 separate power targets and ramp rates by temporarily pausing its non-critical workflows.
The data centre industry hopes that the success of the trial will help open the door to the next generation of data centres being treated as power-flexible, grid-responsive assets around the globe.
This would mean that data centres could be connected to the grid within existing capacity headroom, resulting in faster connections for this critical industry.
However, despite the data centre industry appearing to embrace the premise, some experts believe data centre flexibility should be mandated for the full benefits to be realised – something the UK government has just announced it is considering.
Flexible factories
Most data centres connected to the grid today deliver a time-critical workflow, such as cloud-based services for a call centre, and therefore have a constant demand profile, leaving little opportunity for providing flexibility.
So-called ‘AI factories’ are different. They have a mixed workflow, with a proportion focused on producing large-scale intelligence through AI training and batch inference. These workflows are less time-critical and can be momentarily paused, reducing the demand load of the data centre.
The DC Flex project, a global initiative run by US organisation the Electric Power Research Institute (EPRI), is currently building a picture of the potential for AI factories to act as grid-responsive assets through a series of trials.
One of these was held at Nebius’ new AI factory in London. Over a five-day period in December 2025, a cluster of 96 Nvidia Blackwell Ultra GPUs (graphics processing units) was put through its paces by National Grid.
A set of commercially representative AI training workloads were selected to run on the cluster for the duration of the trial to replicate production-grade conditions.
This workload, and therefore power demand, was then managed by software developer Emerald AI’s Conductor platform, a ‘smart mediator’ software layer that sits between the grid operator and the IT infrastructure.
In response to dispatch signals submitted through a portal designed by Emerald AI, the software autonomously selects and pauses lower-priority work, thereby providing flexibility while ensuring service-level agreements are still met.
National Grid was able to specify the notice period, power reduction percentage, ramp-down, ramp-up and overall event duration through the portal. These signals were submitted remotely without any prior knowledge from Emerald AI.
The Emerald Conductor platform was able to achieve 100% compliance with 22 distinct real-time dispatch events, which included more than 200 power targets and ramp rates. The signals were designed to be rigorous and representative of realistic grid operating conditions under grid stress events.
As well as the ‘TV pickup effect’, otherwise known as peak load relief, which saw the cluster drop its demand from 120kW to 80kW, the data centre was also asked to respond to an emergency load reduction event based on a situation that developed on the 9th August 2019.
On that day there was a sudden unexpected reduction in generation from both Hornsea Offshore Wind Farm and Little Barford gas station. Together these units shed 1.38GW of generation in seconds and triggered the grid’s last resort protection mechanism, which saw 1GW of demand automatically disconnected to balance the system.
In the test the cluster was subjected to a surprise emergency curtailment signal which required immediate material reductions in load. The smart mediator platform was able to reduce the power consumption by 30% within 40 seconds. In another surprise test it managed to increase this to 40% within roughly one minute.
The tests also demonstrated the data centre’s ability to reduce its consumption over sustained periods of time response to a new signal every five minutes.
National Grid requested a 10-40% reduction to be sustained for between two and 10 hours. Emerald Conductor achieved this while maintaining 98.8% performance on its highest priority jobs.
This means data centres are potentially better suited to providing sustained load reduction than conventional flexibility sources such as battery storage, which are more effective at short-duration balancing.
The trial also demonstrated the data centre’s ability to follow frequent fine-grained signals over time by following a carbon intensity signal at five-minute granularity. Derived from Great Britain’s active generation mix over time, the signals required power to be reduced during high intensity periods and increased when system emissions were lower.
For more data centre news, sign up to our newsletter.
