TodayToday's edition
Emerald AI's Conductor software cut AI rack power about 25% for 30 minutes at Nscale's Iceland site, a vendor-reported grid flexibility pilot announced at NVIDIA GTC Berlin.
2 min readAI-assisted
The Nscale site in Iceland where Emerald AI ran its power-flexibility pilot with NVIDIA, as shown on Emerald AI's announcement page.
Image: Emerald AI
Source · Press image from the Emerald AI newsroom, used for news reporting.
Emerald AI said on Oct 21 that its Emerald Conductor software, working with NVIDIA's DSX power-management technology, cut the power use of AI server racks by about 25% at Nscale's AI factory in Iceland, held the cut for 30 minutes and then returned the racks to normal operation.
The company announced the pilot at NVIDIA GTC Berlin and published the results on its own site. The test responded to a signal like one a utility might send during peak demand. Emerald AI said the 30-minute event was consistent with EPRI Flex MOSAIC Class B parameters, a framework for how data centers can support the grid through flexible demand.
Emerald AI's argument is that grids are built for a small number of peak hours each year and have spare capacity the rest of the time. In its account, an AI factory that can ease its draw during those peaks can connect sooner on existing capacity, rather than waiting years for new power plants and lines. Emerald AI CEO Varun Sivaram said utilities need to see that AI computing can cut power on request before they will connect such sites faster. Voltgrid News reported earlier on Microsoft's dropped Nscale site as the UK grid slips.
For grid operators the number that matters is how much load can be shed, how quickly and for how long. A 25% cut for 30 minutes is a modest, short event. It shows the mechanism works on a real cluster, but it does not show how a site behaves through a multi-hour emergency.
It was the first time Emerald Conductor has fully integrated with NVIDIA DSX, and the company said it builds on six earlier Emerald AI demonstrations with NVIDIA. Emerald AI is a member of NVIDIA Inception.
The two figures come from separate tests, and both are vendor-reported.
Illustration (AI-generated)
NVIDIA's own GTC Berlin release described a separate early trial with Emerald AI, EPRI, Nebius and the UK's National Grid, in which Blackwell Ultra cluster power draw fell by up to 40% during simulated grid events while critical workloads kept running. The two figures come from different tests, hardware and baselines, so they should not be read as one result.
Nscale's earlier Iceland evaluation is also a different measurement. NVIDIA said DSX MaxLPS across four GB300 NVL72 racks gave 49% higher inference throughput and 37% more managed GPU capacity inside a 264-kilowatt provisioned budget, as Voltgrid reported. That is an efficiency result, not a grid-flexibility one.
All figures come from Emerald AI and NVIDIA, and no independent verification has been published. The source does not give the baseline load, the number of racks in the Iceland test or the size of the signal. It was a pilot responding to a utility-style signal, not a live grid event, and no utility or grid operator is named as a participant in the Iceland test. Whether utilities will accept demonstrations like this as a basis for faster connections is still open.
