IN Brief:
- National Grid says power-flexible AI data centres could return more than 2GW of capacity to the GB system by 2030 as over 6GW of data-centre demand seeks connection.
- Emerald AI software controlled a 96-GPU NVIDIA cluster in London, cutting demand by up to 40% in under a minute while critical workloads continued to run.
- The trial points to a model in which large AI facilities are treated as dispatchable demand rather than fixed load.
A live London trial has shown that AI workloads in a grid-connected data centre can be curtailed at speed without interrupting critical operations, giving National Grid a new technical basis for treating hyperscale computing as flexible demand rather than a permanently fixed load.
The demonstration was carried out at Nebius’ new London data centre, where Emerald AI’s Emerald Conductor software managed a cluster of 96 NVIDIA Blackwell Ultra GPUs over five days in December 2025. More than 200 simulated grid events were sent to the site during the trial. In each case, the system adjusted power consumption to the requested level, with demand cut by up to 40% in under a minute while workloads continued to run as normal.
The test went beyond simple curtailment. National Grid said the software responded to short, sharp surges in electricity demand, including spikes associated with half-time breaks during major football matches, and also followed longer load-reduction requests lasting up to 10 hours. In a separate system stress scenario, the facility shed 30% of its load in roughly 30 seconds.
That combination is important. Grid operators are not dealing with a single pattern of demand stress, and data-centre flexibility only becomes valuable if it can respond to several operating conditions — fast disturbances, evening peaks, and longer periods of system tightness caused by weather or generation shortfalls. A site that can do all three starts to look less like a connection problem and more like a controllable system asset.
The timing is difficult to ignore. AI infrastructure is pushing power demand higher just as connection queues, reinforcement costs, and local capacity constraints are already slowing electrification projects across the UK. National Grid has said the country is preparing for more than 6GW of data-centre deployments on the grid by 2030. On its estimate, if similar flexibility were built into AI facilities at scale, more than 2GW of capacity could be returned to the system when required.
That does not remove the need for new infrastructure. It does, however, change the conversation around how new data-centre load is connected and managed. If high-performance compute can be scheduled and curtailed in a controlled way, operators may be able to use existing network capacity more efficiently and connect new sites sooner than a conventional fixed-demand model would allow.
The work also feeds into a wider DCFlex programme led by EPRI, which is testing how large data centres can support grid stability across different geographies and market structures. For the UK, the London trial offers something more concrete than theory: a field-tested demonstration that flexible AI load can be integrated into grid operations without shutting down the applications that created the demand in the first place.
