UK-first trial shows an AI data centre acting as a flexible grid asset

What happened

National Grid announced on 2 March 2026 the results of a UK-first trial demonstrating that an AI data centre can act as a controllable, flexible asset on the electricity grid. Over five days in December 2025 at Nebius's London data centre, software from Emerald AI managed a cluster of 96 NVIDIA Blackwell Ultra GPUs through more than 200 simulated real-time grid events.

The results were specific. The cluster cut its power draw by up to 40 per cent at peak without disrupting critical workloads. In a simulated grid emergency it shed 30 per cent of its load in roughly 30 seconds.

It sustained reductions for up to 10 hours to mimic response to low wind or extreme heat, and smoothed demand spikes timed to football half-time surges. The trial was run under Ofgem's Network Innovation Allowance, with National Grid Partners, investor in Emerald AI alongside NVIDIA's NVentures, and EPRI's DCFlex initiative involved.

The technology

Emerald AI's Conductor software sits between the grid and the GPU cluster, orchestrating computing jobs against a power target. AI workloads are not uniform: training runs can be paused or slowed, some inference is latency-tolerant, and jobs can be re-sequenced — so the software throttles, defers and reprioritises work to shape the facility's electrical demand in real time, while ring-fencing workloads flagged as critical.

That converts a data centre from a flat, always-on load into something closer to a dispatchable demand-response asset: capable of fast emergency curtailment in seconds, sustained multi-hour reductions, and fine-grained peak shaving. Dr Varun Sivaram, Emerald AI's founder and chief executive, said the trial delivered "fast emergency curtailment and sustained, precise peak reduction". Josh Parker, NVIDIA's head of sustainability, said it proves the hardware "can act as a grid-aware asset, modulating demand in real time to support stability".

Why it matters for UK & Ireland

The UK projects more than 6GW of new data centre deployments by 2030, arriving on a grid where connection queues are measured in years and reinforcement in decades. The binding assumption behind those queues is that a data centre must be treated as inflexible peak load, with the network sized for its maximum draw. This trial attacks that assumption with evidence: National Grid suggests flexible operation could add more than 2GW of capacity back to the grid when needed.

The commercial logic follows. Steve Smith, president of National Grid Partners, said high-performance data centres "can be connected and managed without major new network capacity, flexing their power up or down in real time".

If a facility can guarantee a 30-second, 30 per cent load shed, the network can safely connect it sooner and at lower cost — the trial's backers explicitly link flexibility to accelerated connection timelines and reduced network charges. For Ireland, where Dublin-area connections are constrained enough that developers are building microgrids, demand flexibility offers a complementary route to the same end: more compute per megawatt of network.

There is also a system benefit. A fleet of grid-aware data centres that can absorb surplus wind and back off during scarcity behaves, in aggregate, like fast-responding storage — at high renewable penetration, that is valuable on both sides of the Irish Sea.

What to watch

Whether evidence becomes policy: the trial data is being shared with industry, regulators and policymakers to inform possible "power-flexible" connection rules, and any Ofgem or NESO move to formalise flexible connection terms for data centres would be the decisive signal. Watch also the scale-up path — Emerald AI's blueprint feeds a planned 100MW power-flexible AI facility in Virginia, and integration into NVIDIA's Omniverse DSX Blueprint for gigascale plants, and whether UK hyperscalers accept workload orchestration on commercial fleets, where contractual service guarantees, not physics, may prove the harder constraint. A five-day, 96-GPU trial is proof of concept; the test is whether flexibility holds at hundreds of megawatts under real market incentives.