A London-based datacenter has demonstrated its capability to significantly reduce power consumption of AI infrastructure in response to grid signals, while maintaining uninterrupted critical operations. Over a five-day trial conducted last December, a cluster of Nvidia Blackwell Ultra GPUs, part of a datacenter managed by GPU-as-a-service provider Nebius, underwent over 200 simulated grid event tests. These tests, orchestrated by energy company National Grid, aimed to verify the site’s dynamic power management response. Results indicated a power reduction of up to 40%, facilitating key workloads to proceed unaffected.
Emerald AI, supplying the necessary software, and the Electric Power Research Institute (EPRI), which oversees the Datacenter Flexible Load Initiative (DCFlex), collaborated on the project. Their approach predominantly involved pausing or rescheduling non-urgent tasks rather than shutting down infrastructure entirely. According to a whitepaper by National Grid, this strategic flexibility is akin to Google’s US operations, where non-essential AI processes are paused to aid grid stability.
To simulate realistic environments, a series of significant AI training workloads, such as gpt-oss and Llama models, were used. The cluster maintained full utilization throughout these tests. Grid signals detailing power reduction and response timeframes were provided by National Grid Electricity Transmission (NGET) and EPRI. Some scenarios tested immediate response capability without prior notice, simulating real-life demand spikes, such as those experienced during high-energy events like the halftime of major sports matches.
This trial serves as evidence that high-performance datacenters can be efficient partners in managing grid constraints, argues National Grid Partners’ president, Steve Smith. Critically, the potential of employing datacenters to ease grid demand is explored here, balanced against existing challenges like new power capacity lagging behind expanding datacenter infrastructure.