TL;DR
A GPU cluster near London cut its power consumption by up to 40% on command during a five-day trial with National Grid, achieving 100% compliance across more than 200 simulated grid events. The approach pauses or deprioritises AI jobs rather than shutting down infrastructure.
A UK data centre has demonstrated it can slash the power drawn by AI infrastructure by 40% in response to grid signals, without disrupting time-sensitive workloads.
The trial ran over five days last December at a facility near London operated by GPU-as-a-service provider Nebius. A cluster of Nvidia Blackwell Ultra GPUs responded to more than 200 simulated grid event notifications, achieving 100% compliance with all requested power targets and ramp rates.
How It Works
Rather than simply powering down hardware, the system — built with software from Emerald AI — pauses or deprioritises lower-priority jobs and shifts workloads to later time slots. Latency-sensitive tasks like inference continue running normally, while throughput-intensive work such as training and fine-tuning gets rescheduled around natural “flex points” like checkpoint intervals.
The project also involved the Electric Power Research Institute (EPRI) through its Datacenter Flexible Load Initiative (DCFlex).
To approximate real production conditions, Emerald AI and Nebius ran commercially representative training workloads including gpt-oss, Llama, and Qwen models. National Grid submitted signals specifying notice periods, power reduction percentages, and ramp durations — some with zero advance notice, requiring immediate response.
The tests even simulated real-world demand spikes, such as the surge when millions of viewers put the kettle on during half-time of major football matches.
Scale and Implications
The 130 kW compute cluster involved is roughly equivalent to the power consumption of 400 UK households. While modest in scale, the trial’s 100% compliance rate suggests the approach could work for much larger deployments.
“High-performance data centres don’t have to place additional strain on the grid,” said National Grid Partners president Steve Smith.
A whitepaper from National Grid concludes that replacing rigid “firm load” models with measurement-based flexibility gives grid operators new options for delivering capacity more efficiently.
Looking Forward
The trial offers a practical response to concerns about AI’s growing energy footprint in the UK. However, as The Register notes, the approach assumes data centres can get connected to the grid in the first place — new generating capacity is not keeping pace with construction, and some developers face years-long waits for grid connections and substation upgrades.