AI Factories Can Stabilize Power Grids

When millions of British viewers turned on their kettles during halftime of the England-Germany Euro 2020 match, the power grid experienced a sudden surge equivalent to adding an entire nuclear reactor's output in minutes. This 'TV pickup' phenomenon illustrates the delicate balancing act grid operators face daily, a that grows as energy-hungry AI factories come online. Now, a new approach turns these massive power consumers into flexible assets that can actually help stabilize the grid during peak demand periods.

Emerald AI, in collaboration with NVIDIA, EPRI, National Grid, and Nebius, has demonstrated how AI factories can autonomously adjust their power consumption to relieve grid strain. Their solution, the Emerald AI Conductor Platform, allows these facilities to temporarily reduce energy use when the grid is under stress, acting as a 'shock absorber' for sudden demand spikes. This capability means AI factories can connect to existing power infrastructure faster without waiting for years-long grid upgrades, while also helping limit the need for expensive new power plants that drive up electricity rates.

In a December demonstration at Nebius's new AI factory in London, researchers tested this technology under realistic conditions. They ran production-grade AI workloads on a cluster of 96 NVIDIA Blackwell Ultra GPUs connected through the NVIDIA Quantum-X800 InfiniBand platform. The NVIDIA System Management Interface provided seconds-level power telemetry, while EPRI and National Grid simulated various grid stress scenarios, from lightning strikes to prolonged periods of low wind power generation.

One key test recreated the exact energy surge from that Euro 2020 football match. As millions of simulated tea kettles were about to be turned on, the AI cluster successfully ramped down its power consumption without disrupting high-priority AI workloads. The system achieved 100% alignment with over 200 power targets that grid operators instructed the cluster to follow. 'We did tests that go beyond the ones that have been done so far in the U.S. because we tested not just the GPUs, but also the CPUs and everything that sits around it as well as the total power consumption of the IT equipment,' said Steve Smith, group chief strategy officer of National Grid.

The demonstration showed that flexible AI factories can help grid operators manage sudden demand swings more efficiently using existing capacity. This reduces the need to overbuild permanent infrastructure to meet worst-case peaks, which ultimately helps keep electricity rates more affordable for consumers. 'With this technology, AI factories become friendly and helpful grid assets,' said Varun Sivaram, founder and CEO of Emerald AI. 'Simultaneously, the AI factories get connected much faster to the grid because they can tap into existing power grids.'

For the UK specifically, this technology addresses a critical bottleneck. London's power grid faces constraints in infrastructure upgrades needed to connect large customers like AI factories. By making these facilities flexible, the UK can support AI industry growth without requiring massive grid expansions. 'We have enormous skills and potential in AI,' said Smith. 'We're never going to be on the scale of the U.S. in terms of data centers, but relative to the size of the country, we could be and we're certainly seeing that interest from many of the hyperscalers.'

The London demonstration followed successful proof-of-concept trials at AI factories in Arizona, Virginia, and Illinois. With four demonstrations completed, Emerald AI and NVIDIA are preparing for real-world deployment at the Aurora AI Factory in Virginia, scheduled to open this year. This progression from testing to implementation suggests the technology is moving toward practical application rather than remaining purely experimental.

The approach does have limitations. The system prioritizes certain AI workloads over others during power reductions, meaning some flexible jobs experience temporary slowdowns. Additionally, the technology requires specific hardware and software integration, including NVIDIA's power monitoring systems and Emerald AI's Conductor Platform. While the demonstrations have shown promising across multiple locations and scenarios, widespread adoption will depend on industry acceptance and regulatory frameworks that incentivize flexible power consumption.