Flex appeal: UK datacenter cuts AI power draw 40% on command
A UK datacenter has successfully demonstrated it can reduce the amount of power drawn by AI infrastructure in response to grid events, without disrupting critical workloads.
The trial, conducted over five days last December, involved a cluster of Nvidia Blackwell Ultra GPUs installed in a datacenter near London operated by GPU-as-a-service biz Nebius.
This involved more than 200 simulated grid event notifications, sent to the site to test its ability to dynamically adjust the cluster's power consumption. This was achieved successfully, cutting power demand by up to 40 percent while key tasks continued to run as normal, according to energy provider National Grid.
As well as National Grid and Nebius, the project involves Emerald AI, which supplies the software, and Electric Power Research Institute (EPRI) with its Datacenter Flexible Load Initiative (DCFlex).
A whitepaper provided by National Grid reveals that power control is largely achieved by pausing or deprioritizing jobs running on the GPUs, or shifting workloads to a later time, rather than the blunt instrument of powering down parts of the infrastructure.
Google is already doing this in the US, and said last year it will pause non-essential AI workloads to protect power grids.
While some AI workloads - such as inference - are latency sensitive, others including training and fine-tuning are more throughput-intensive. These latter tasks also typically include natural "flex points" like checkpoint intervals, where processing can be paused, the whitepaper explains.
To approximate production-grade conditions, Emerald AI and Nebius chose a set of commercially representative AI training workloads, including the gpt-oss, Llama, and Qwen models. The cluster was kept continuously utilized running these.
For the power experiments, National Grid Electricity Transmission (NGET) and EPRI submitted grid signals through an event submission portal, specifying the notice period, power reduction percentage, ramp-down duration, ramp-up duration, and overall event duration.
Some of the tests involved "surprise" signals, where no advance notice was given, or specifying no ramp time, requiring immediate response. Some also emulated real-world spikes in demand seen when thirsty Brits put the kettle on for a brew during half time of major football matches.
The tests were carried out as the Nebius "AI Factory" was being brought online, and involved a 130 kW compute cluster, roughly equivalent to the power consumption of 400 UK households.
According to the whitepaper, the cluster achieved 100 percent compliance with all requested power targets and ramp rates, suggesting these systems could help alleviate the problems caused by AI's huge power consumption.
By replacing the rigid "firm load" models of the past with measurement-based flexibility, grid operators and policymakers can create new options for delivering capacity efficiently, the report concludes.
"As the UK's digital economy accelerates, there's concern that datacenters could add pressure to an already constrained system. This trial proves the opposite can be true. High‑performance datacenters don't have to place additional strain on the grid," claimed National Grid Partners president Steve Smith.
That assumes you can get the power in the first place. As The Register has reported previously, new generating capacity isn't being added at the same rate datacenters are being built, and some developers complain they face a wait of years to get a grid connection and for local substations to be upgraded. ®
