AI Factories Meet the Power Grid
Imagine millions of people in the U K all getting up at half time during a big football match to switch on their kettles at the same moment. That is exactly what happened during the UEFA EURO 2020 match between England and Germany. The result was a sudden spike of about 1 gigawatt of power demand on the National Grid which is about the average output of a standard nuclear reactor.
Events like this are a nightmare for grid operators. They need to keep the power system stable even when demand jumps in seconds. Traditionally this means overbuilding power infrastructure so there is always enough capacity to handle the worst case spike. That approach is slow and extremely expensive.
Now a new kind of customer is appearing on the grid massive AI factories packed with GPUs and high performance servers. At first glance these data centers just look like another huge drain on the grid. But a project involving Emerald AI NVIDIA EPRI National Grid and Nebius is flipping that idea on its head. Instead of being a problem these AI factories can actually help stabilize the grid by intelligently flexing their power usage in real time.
Inside a Power Flexible AI Factory
Emerald AI has built a platform called the Emerald AI Conductor that turns large AI data centers into what they call power flexible AI factories. The idea is simple. When the grid is under stress the AI factory automatically dials down its power consumption in a controlled way. When the grid is stable it runs normally at full speed.
To prove this out the team deployed the Emerald AI Conductor Platform at Nebius new AI factory in London which is built on NVIDIA infrastructure. This site is one of the first of its kind in the U K and is a great showcase for how modern GPU heavy compute can interact intelligently with the energy system.
Under the hood the AI factory is running production grade AI workloads on a cluster of 96 NVIDIA Blackwell Ultra GPUs. These are connected through NVIDIA Quantum X800 InfiniBand networking and monitored using the NVIDIA System Management Interface which gives second by second telemetry on GPU power usage. That detailed data is crucial because it lets the Conductor Platform react very quickly and precisely to changing grid conditions.
EPRI and National Grid then simulated real world stress scenarios on the power grid. These ranged from lightning strikes to long periods of low wind generation. During each test the grid simulation sent signals to the AI factory telling it to cut or adjust its power draw. The Emerald AI Conductor responded automatically shifting how the workloads were scheduled and how much power the cluster was using.
One of the most interesting tests was a recreation of the U K TV pickup effect the type of surge that happens when millions of people simultaneously put the kettle on during a break in a major event. In this simulation as the fictional kettles began to switch on the AI cluster ramped down its power draw acting like a shock absorber for the grid.
Crucially this did not mean killing all compute. High priority AI jobs kept running at full performance while lower priority or more flexible tasks were temporarily slowed. That makes the setup practical for real AI operations where uptime and throughput for key services are critical.
Performance Results and Why It Matters
The London demonstration delivered some strong metrics. Emerald AI reported one hundred percent alignment with more than two hundred different power targets that EPRI and National Grid sent during the experiments. In plain language every time the grid asked the AI factory to hit a certain power level the system complied.
The tests included twenty two real time dispatch events where the grid effectively told the data center to adjust power usage on the fly. In some cases the AI factory was able to slash its power draw by about thirty percent in under forty seconds. For grid operators that sort of responsiveness is incredibly valuable. It means they can handle sudden swings in demand without necessarily firing up additional peaker plants or building out new long term infrastructure.
Steve Smith group chief strategy officer at National Grid highlighted another important point. These tests did not just watch GPU power in isolation. They included CPUs support systems and the total power usage of the I T equipment. That gives a realistic picture of what can be controlled in a real deployment and proves the broader value of the approach.
For cities like London where the grid is already under pressure to connect new large power users this is a big deal. Data centers and AI factories are hungry for electricity. Normally that would mean waiting years for infrastructure upgrades before they can run at scale. With a power flexible design these AI factories can plug into the existing grid more quickly because they offer something in return flexibility. Instead of always pulling maximum power they help smooth demand and free up headroom when the grid really needs it.
From a broader tech perspective this model changes the way we think about high performance compute. Traditionally power has been treated as a fixed cost and a fixed constraint. Here it becomes dynamic. GPUs and CPUs still deliver massive performance for AI workloads but they become part of a larger intelligent system that cooperates with the energy network.
Emerald AI and NVIDIA are already planning to move from demos to real deployments with the Aurora AI Factory in Virginia expected to open this year. If this approach scales it could shape how future GPU heavy data centers are built and powered worldwide. For anyone interested in high performance computing and modern infrastructure this is a sign of where the ecosystem is heading powerful AI hardware tightly integrated with equally smart energy systems.
Original article and image: https://blogs.nvidia.com/blog/power-flexible-ai-factories-energy-grid/
