AI data centres have a power problem that is becoming a grid problem. Emerald AI thinks it has a solution, and it has spent the past year proving it works.
The company's white paper sets out a case for power-flexible AI factories: facilities that can autonomously reduce their electricity draw within seconds when the grid comes under strain, without interrupting production workloads. The approach is designed to let large AI customers connect to the grid faster, by offering flexibility in exchange for capacity, rather than waiting years for permanent infrastructure upgrades.
The London test
The most demanding demonstration took place at Nebius' new London AI factory, built on Nvidia infrastructure. Emerald AI ran production workloads on a cluster of 96 Nvidia Blackwell Ultra GPUs connected over the Nvidia Quantum-X800 InfiniBand platform, with seconds-level telemetry drawn from the Nvidia System Management Interface.
EPRI and National Grid simulated grid stress events and sent signals prompting temporary power ramp-downs. Emerald AI recorded 100% alignment with more than 200 consecutive power targets. National Grid's group chief strategy officer Steve Smith noted that the tests went further than previous US trials, covering not just GPUs but CPUs and all surrounding IT equipment, as well as total power consumption across the facility.
Additional trials were conducted in Arizona, Virginia and Illinois.
Why this matters beyond the data centre
The white paper frames the problem around abrupt demand spikes, using the example of the TV pickup kettle surge, the sudden jump in electricity demand when large numbers of people simultaneously boil kettles during a television break. AI factories present a version of the same problem at industrial scale, with the added complication that their power draw can shift dramatically depending on workload.
Power-flexible facilities that can respond to grid signals turn that volatility into a resource. Instead of requiring utilities to build permanent capacity to handle peak AI demand, grid operators can call on AI factories to reduce load during stress events, easing pressure on existing infrastructure.
"With this technology, AI factories become friendly and helpful grid assets," said Varun Sivaram, founder and CEO of Emerald AI.
What comes next
Related reading
- Nvidia and energy partners unveil flexible AI data centres to ease grid pressure
- Microsoft and Nvidia team up to bring AI to nuclear power permitting and design
- Nordic Semiconductor expands 'ultra low power' AI chip lineup
Emerald AI and Nvidia say they are moving beyond demonstrations toward commercial deployments. The Aurora AI Factory in Virginia is planned for deployment this year.
The London results give that timeline credibility. Hitting every power target across 200 simulated stress events is not a proof of concept. It is an operational track record.
The recap
- Emerald AI white paper demonstrates power‑flexible AI factory capability.
- 96 NVIDIA Blackwell Ultra GPUs used in Nebius London demonstration.
- Aurora AI Factory in Virginia planned for deployment this year.
