The Grid Got a New Option. Most of the Industry Missed What It Actually Proved.

On March 3, Emerald AI, EPRI, National Grid, Nebius, and NVIDIA published the results of a five-day live trial at a London data center. Most of the headlines focused on power reduction. The real story is simpler and more important: this trial eliminated one of the energy industry's favorite excuses.

A cluster of 96 NVIDIA Blackwell Ultra GPUs cut electricity demand by up to 40% in under a minute while running dozens of simultaneous AI workloads. Not in a simulation. Not in a lab. In a live Nebius facility, responding to more than 200 grid events across five days in December 2025.

That matters because the old argument was that AI data centers are fixed-load assets. Always-on consumers. A new burden on the grid that requires dedicated baseload, expensive upgrades, and long lead times. London made that position a lot harder to defend.

Emerald's software, called Emerald Conductor, accepted grid signals, prioritized workloads, and dynamically flexed compute demand through shifting, slowing, and clock-speed adjustment, all without disrupting the workloads themselves. The compute load behaved less like a passive consumer and more like a controllable grid asset. Peak smoothing during demand surges. Load shedding of 30% in roughly 30 seconds during simulated system stress. Sustained reduction held for up to 10 hours during low-wind and extreme-heat scenarios.

And this was not a one-off. In Phoenix, working with EPRI's DCFlex initiative and Oracle, the team modulated known workloads and hit a 25% power reduction over three hours. In Chicago, the system handled unknown, random workloads and adapted automatically. London raised the bar again: live facility, live grid signals, Blackwell hardware, and the most aggressive targets yet. That progression matters because it closes the gap between theoretical capability and operational credibility.

National Grid Partners President Steve Smith did not hedge: "High-performance data centres don't have to place additional strain on the grid. We've shown they can be connected and managed without major new network capacity, flexing their power up or down in real time to support the whole system."

The UK is preparing for over 6 GW of data center deployments by 2030. The trial suggests technology like Emerald Conductor could enable AI data centers to return more than 2 GW of capacity to the grid when needed. That is not a rounding error.

The U.S. implications go further. Arushi Sharma Frank at Luminary Strategies connected the London results directly to PJM and FERC. The trial provides physical evidence that AI loads can function as dispatchable resources, not just consumers. That reframes interconnection queues, co-location agreements, and grid planning across both deregulated capacity markets and vertically integrated utility territories. Instead of sizing grid capacity to worst-case peak, operators can use orchestration software to flex demand when the system is stressed. The concept is market-agnostic and location-agnostic. What matters is trustworthy, layered dynamic response.

The founder matters here because the company is moving at a pace that only makes sense when you understand who built it. Varun Sivaram is a Rhodes Scholar physicist out of Stanford and Oxford who served as Managing Director for Clean Energy at the State Department under John Kerry, where he created the First Movers Coalition, a public-private initiative that convened over 100 companies around $12 billion in commitments. He was Chief Strategy and Innovation Officer at Orsted, the $25 billion offshore wind company, leading 200 people across strategy, innovation, and M&A. Before that, CTO of ReNew Power, India's largest renewable energy company. TIME 100 Next. MIT Top 35 Innovators. World Economic Forum Young Global Leader. Three books, including one the Financial Times called "the best available overview of where the solar industry finds itself today."

That is not a startup founder learning the energy business as he goes. That is someone who built coalition frameworks, sat in the policy rooms, ran the technology organizations, and operated at the executive level inside companies that own grid assets. The distance between "this technology could work" and "here is National Grid, NVIDIA, and EPRI confirming it does" is the distance between research and deployment. Emerald crossed it in eleven months.

Total funding is $42.5 million after a seed extension led by Lowercarbon Capital, with NVentures, Radical Ventures, Salesforce Ventures, National Grid Partners, and John Doerr of Kleiner Perkins. The investor list is a deliberate assembly of the exact stakeholders who need this to work: the chip company, the grid operator, the climate-focused capital, and the allocator who has spent decades funding infrastructure that reshapes industries.

The AI energy conversation has been stuck in a loop. Demand surges. The grid falls behind. Build more. Wait longer. Pay more. That framing treats compute as a fixed, passive consumer. London proved compute can be an active participant in grid management. The question is no longer whether AI infrastructure can flex. The question is how quickly regulatory and commercial frameworks catch up to what has already been demonstrated.

The grid got a new option. The constraint now is not technical capability. It is institutional speed.

Sources: National Grid, Bloomberg, Data Center Dynamics, DCD Analysis, Luminary Strategies, Emerald AI, MCJ Collective, Council on Foreign Relations