A simple cup of tea once triggered a massive surge in electricity demand across the U.K. Now, that same phenomenon is helping reshape the future of AI and power grids.
During a UEFA EURO match between England and Germany, millions of people switched on their kettles at halftime. The result? A sudden spike of about 1 gigawatt — roughly the output of a nuclear reactor — recorded by National Grid.
These abrupt demand swings, known as “TV pickups,” are a long-standing challenge for grid operators. But as AI data centers grow rapidly, the pressure on power infrastructure is intensifying — raising a critical question: can these massive energy users actually help stabilize the grid instead of straining it?
Turning AI Into a Grid Asset
A new white paper from Emerald AI, developed with NVIDIA, Electric Power Research Institute, National Grid, and Nebius, suggests the answer could be yes.
The concept is simple but powerful: “power-flexible” AI factories that can automatically adjust their electricity usage during peak demand periods.
Instead of being rigid, always-on consumers, these AI facilities dynamically scale their power use — easing strain on the grid when it matters most.
For energy providers, that means fewer costly infrastructure upgrades. For consumers, it could help keep electricity prices stable by reducing peak load pressure.
Real-World Testing in London
The idea has already moved beyond theory.
In London, researchers deployed Emerald AI’s Conductor Platform at a new AI facility operated by Nebius and powered by NVIDIA systems. The setup included 96 advanced GPUs running real production-level AI workloads, connected through high-speed infrastructure.
Using detailed telemetry from NVIDIA systems, operators could monitor and control power usage in real time.
To test the system, EPRI and National Grid simulated extreme grid stress scenarios — from lightning strikes to low wind energy supply. One key test recreated the famous EURO “tea break” surge.
As millions of simulated kettles switched on, the AI cluster responded instantly — reducing its power consumption within seconds.
Crucially, the system maintained performance for high-priority AI tasks, while temporarily slowing less critical workloads. In effect, the AI facility acted like a shock absorber for the grid.
Impressive Performance Metrics
The results highlight how effective the approach could be at scale.
- 100% compliance with more than 200 power adjustment signals
- 22 real-time dispatch events handled successfully
- Up to 30% power reduction achieved in under 40 seconds
According to Varun Sivaram, CEO of Emerald AI, this shifts the role of AI data centers entirely.
“AI factories can become helpful grid assets,” he said, noting that flexibility also allows faster connection to existing power networks without waiting years for upgrades.
Why This Matters Now
The timing is critical. Cities like London are facing growing electricity demand driven by AI, cloud computing, and electrification.
At the same time, grid expansion projects often take years due to regulatory, financial, and physical constraints.
By making large energy users flexible, utilities can better manage demand using existing infrastructure — reducing the need to build systems for worst-case peak scenarios.
Steve Smith, strategy chief at National Grid, emphasized that the trials went further than previous tests by measuring total IT power usage, not just GPUs.
“We’ve proved the value that this technology brings,” he said.
A Faster Path to AI Expansion
Beyond grid stability, the approach could accelerate AI industry growth.
Flexible power usage allows new AI facilities to connect to the grid faster — removing one of the biggest bottlenecks for expansion.
In the U.K., where space and infrastructure are limited compared to the U.S., this could be a game-changer for competitiveness in the global AI race.
It also opens the door to more sustainable growth, aligning AI expansion with energy efficiency goals.
What Comes Next
After multiple successful demonstrations, Emerald AI and NVIDIA are preparing for real-world deployment at the Aurora AI Factory in Virginia, expected to open this year.
If scaled successfully, power-flexible AI factories could redefine how energy systems and digital infrastructure interact.
The bigger insight is clear: instead of being part of the problem, AI data centers could become part of the solution — helping balance the grid in real time while powering the next wave of innovation.