Emerald AI raises $150mn at a $1.05bn valuation to flex data centre power

A two-year-old company is now worth $1.05bn. It writes software that makes data centres use less electricity on demand.

Emerald AI raised $150mn in an oversubscribed Series A, the company announced on Tuesday. Energize Capital and DCVC co-led it, and the round takes total funding past $220mn.

The investor list is the tell. Nvidia, Siemens, RWE and GE Vernova all took part. So did Aramco Ventures, Samsung Ventures, Salesforce Ventures, JERA Ventures and the CIA-backed In-Q-Tel, alongside John Doerr and Tom Steyer. Twelve Fortune Global 500 companies now hold stakes and sit on the company’s strategic advisory board.

What the software actually does

Emerald Conductor sorts computing jobs by how much delay each customer will tolerate. A utility signals that the grid is under strain. The platform then slows, pauses, caps or moves the jobs that can wait.

That is a different target from most efficiency software, which adjusts cooling equipment and leaves the computing alone. Emerald goes at the workload itself.

Founder and chief executive Varun Sivaram held climate policy roles in the Biden administration before starting the company in 2024. Its chief scientist, Ayse Coskun, is a Boston University computer science professor whose research opened the field.

The claim

Applied across the buildout, the approach could unlock more than 100 gigawatts of untapped capacity on the existing United States grid, the company says. That power would arrive years before anyone could build new generation and transmission.

Building grid infrastructure can take a decade. The release cites the International Energy Agency for its other number. Data centres will drive nearly half the growth in US electricity demand through 2030.

The intelligence driving the AI revolution could solve its own greatest bottleneck, Sivaram said, and that bottleneck is power.

Why a utility would want this

Silicon Valley Power serves roughly 55 data centres across 20 square miles in Santa Clara, including Nvidia and Intel. It is one of the most energy-intensive service areas in California.

All the spare capacity the utility once had is now spoken for, its director Nicolas Procos told Sri Muppidi at The New York Times. That leaves two options. Build out the system, or find creative solutions.

Silicon Valley Power picked the second. It has launched a Flexible Load Interconnection Program with Emerald, which grants data centres expanded grid access in exchange for verified, dispatchable flexibility.

What has actually been demonstrated

The strongest published result comes from a field test in Phoenix in May 2025. Emerald and its partners cut the power draw of a 256-GPU Nvidia cluster by 25% from its average base load. They held it there for three hours.

Across 33 experiments the system managed 212 jobs, and none broke its predefined service tier. Power prediction was off by 4.52% against average experiment power.

Marcus Schuler set out the limits for Implicator. Emerald personnel and partners wrote the study, and it covered one pre-profiled cluster. It slowed or paused batch training and fine-tuning, and left real-time inference, streaming and model-serving untouched.

Measuring anything beyond a single cluster will take larger deployments with full-site telemetry, the authors said.

The European test happened in London

Emerald has run five commercial demonstrations, in Arizona, Illinois, Virginia, Oregon and London. Partners included Nvidia, Oracle, Nebius, the research body EPRI and National Grid.

The London site puts the approach in front of a market where the constraint bites hardest. Britain has an Essex data centre that waited on a grid connection. And 63% of new European capacity is now going outside the big five markets, largely because of queues and land.

RWE and Siemens both invested, and RWE sits on the advisory board. European utilities are buying into an American company’s answer to a European problem.

Who controls the switch

A successful test does not create a commercial right to interrupt someone’s machines. Nobody has resolved that question, and it is the real obstacle.

Full utility control of the load-side breaker “is non-negotiable” in exchange for faster interconnection, Silicon Valley Power’s chief operating officer Chris Karwick has said. Operators resist handing that over, because an abrupt shutdown can damage expensive hardware.

No standardised binding agreement between a utility and a data centre existed as of late June, Schuler reported. Bill credits are often too small to justify delaying lucrative work. Faster grid access is the stronger inducement.

Steven Carlini, Schneider Electric’s chief advocate for AI and data centres, put the choice plainly. Data centres can slow, cap or shift workloads, he said, but whether they want to is questionable.

The emissions result is not automatic

Moving computing work into cheaper off-peak hours can raise emissions, where fossil generation supplies the marginal power. A Green Software Foundation policy review found as much in March.

It also found that most demonstrations in the field were under two years old and favoured workloads that were easy to defer.

What the investors are buying

The binding constraint on AI is no longer chips or capital, said John Tough, managing partner at Energize Capital. It is power, and software is the fastest way through it.

DCVC’s Zachary Bogue framed it as a change in category. The platform turns data centres into grid-responsive assets rather than energy-hogging liabilities, he said.

Interconnection rules written for last century’s factories keep excess power locked up, said Shawn Xu of Lowercarbon Capital. Rob Toews of Radical Ventures was blunter still: the bottleneck for AI is not compute, it is energy capacity.

The paperwork lagged the announcement

A Form D filed on 3 August recorded $90,229,639 sold toward a planned $150mn offering, Schuler noted. It named neither the valuation nor the lead investors.

Reuters put the round at $150mn and the valuation at $1.05bn, reporting on Tuesday. It cited a separate Morgan Stanley projection. Global electricity demand will rise by more than a trillion kilowatt hours a year through 2030, with data centres supplying nearly 20% of that growth.

The pitch behind the pitch

Emerald is selling a way out of a political problem as much as a technical one. The backlash against data centres has reached the point where Texas audits connections before granting them, and Greg Abbott now says the industry dug its own grave.

Some operators have concluded the grid is not worth the argument. Amazon is building a Texas facility that will not touch the grid at all, generating its own power on site.

Sivaram wants the opposite. His goal is to make data centres grid-flexible and, in doing so, transform them from villains into heroes, he told The New York Times.

Whether that works depends on something no funding round settles, which is whether operators will hand a utility the authority to turn their machines down.