Emerald AI: Making Data Centers Responsive to the Grid

Emerald AI: Making Data Centers Responsive to the Grid

The AI build-out keeps asking the grid for one thing: more. More power. More access. More substations. More transmission. More patience from communities already wondering why their electricity bill is rising while a new data center goes up down the road.

Emerald AI was built around a different question. What if the people building AI infrastructure did not have to wait for the grid to become larger before they became better partners to it?

Dr. Varun Sivaram founded Emerald AI in 2024. Professor Ayse Coskun serves as chief scientist. Together, they are building software for the moment AI infrastructure meets physical reality: grid constraints, peak demand, utilities, operators, customer commitments, and an AI workload that cannot simply disappear because the system is having a rough afternoon.

Emerald Conductor helps data centers identify eligible compute workloads and onsite energy resources that can respond when the grid is under pressure, while protecting agreed performance requirements. The premise is not that every workload should pause. It is that the right workload, at the right moment, can flex without breaking the operating promise made to the customer.

That distinction is the company.

Emerald has chosen to prove the idea in the real world. The company documents demonstrations in Arizona, Illinois, Virginia, Oregon, and London. In Phoenix, it says a 256-GPU cluster reduced power use by 25% for three hours during peak demand while operating within defined quality-of-service requirements. In California, it describes a commercial, multi-megawatt, full-data-center deployment. In Virginia, it is working with NVIDIA and Digital Realty on a planned 96MW Vera Rubin AI Factory in Manassas.

Those projects explain why Emerald matters. A software company can call itself infrastructure all day long. Eventually it has to perform inside infrastructure someone else is responsible for keeping alive. Utilities have to trust the response. Operators have to trust the computing performance. Customers have to trust that their work will still be done.

Emerald cites work alongside NVIDIA, EPRI, Oracle, Nebius, National Grid, utilities, and grid operators. It is not attempting to replace the people who run power systems or data centers. It is attempting to give them a common operating language before the tension between AI demand and grid capacity becomes impossible to manage politely.

The company’s $150 million Series A, co-led by Energize Capital and DCVC, is a chapter in that story, not the story’s premise. Readers looking for the transaction record can visit DevCuration’s Funding Announcement. Emerald says flexible computing could unlock more than 100 gigawatts of capacity on the existing U.S. grid. That is modeled opportunity, not capacity already under contract.

The more interesting truth is already on the ground: Emerald has shown that AI data-center demand can become responsive under real conditions. Now it has to make that response dependable enough to become part of how the next generation of AI infrastructure gets built.

What to watch next

Emerald’s next proof points are contracted deployments, utility-program expansion, repeatable performance across markets and workloads, and public evidence of which partners become customers.