DCVC co-leads Emerald AI’s $150 million Series A round to transform data centers into intelligent grid-responsive assets

Last May in a hyperscale facility outside Phoenix, a GPU cluster dropped its power draw by 25 percent and held steady for three hours while the local grid rode out a demand spike. The jobs running on it kept their service level agreement (SLA) guarantees. No training runs were killed, no backup diesel generators turned on, the cluster drew less electricity, the surge passed, and the cluster ramped back up to full capacity.

That demon­stra­tion, which amounted to a showcase of Emerald AI’s capa­bil­i­ties, was reported in Nature Energy by a team led by Boston University professor Ayse Coskun — Emerald AI’s Chief Scientist. 

As AI computing continues to scale at breakneck speed, surging power demand has become the industry’s defining bottleneck, putting unprece­dented strain on grid infra­struc­ture, data center builders and operators, data center tenants, local communities, and the environment.

Emerald AI, our newest portfolio company, has the answer to the problem. DCVC Energy + Climate is proud to be co-leading a $150 million Series A financing round for the company. (New York Times coverage here.)

Over the past year, Emerald AI success­fully completed demon­stra­tions at five commercial data centers in Arizona, Illinois, Virginia, Oregon, and London, working alongside partners including NVIDIA, EPRI, Oracle, Nebius, and National Grid as well as regional utilities and grid operators to prove that AI data centers can flex power based on grid signals. ​“Our demon­stra­tions around the world proved that data centers can adjust their power use precisely when the grid needs relief, without compro­mising critical computing workloads,” says Dr. Varun Sivaram, Founder and CEO of Emerald AI. ​“Today that technology runs commer­cially at full data center scale, and this financing lets us take it everywhere AI is built.”

Emerald AI’s strategic position at the inter­sec­tion of energy and compute gives it visibility across the entire data center stack: from utilities to data center operators to compute customers. For utilities, Emerald AI’s software helps manage and dispatch signals to partic­i­pating flexible data centers during grid stress. For data center operators and compute customers, the Emerald Conductor platform can reduce power demand in response to grid signals by shifting compu­ta­tional workloads temporally — throttling or pausing workloads within service level agreement (SLA) guardrails — or spatially through geoshifting to networked facilities (amazing!). Conductor also enables resource flexibility by dispatching battery power or other onsite energy resources as needed, and creates an auditable data trail to establish how a given data center contributed to a demand response event. By unlocking flexibility at this scale, Emerald offers a lever for grid resilience.

Data center builders and operators face multi-year waits for grid inter­con­nec­tions, as utilities struggle to build enough generation and trans­mis­sion capacity to meet peak demand. Tenants and consumers face rising prices, as the resulting shortage of compute supply drives up the price of AI training and inference. For local communities and the environment, the cost looks different. Utilities look to spread the expense of new grid infra­struc­ture across ratepayers, raising electricity bills for nearby households. And to close the power gap, utilities and data center operators are turning to carbon-intensive solutions, such as diesel- or gas-powered generators or peaker plants.

Right now, one of the biggest roadblocks to expanding AI compute capacity is the traditional limitation that data centers require rigid, ²⁴⁄₇ power. Utilities treat data centers as ​“firm” customers — reserving capacity to guarantee 100 percent of contracted peak demand at all times, even during rare hours of maximum grid stress. If that weren’t so — if operators could dynamically throttle power demand down at a utility’s curtailment request during periods of peak load — then data centers would actually become flexible grid assets, available whenever a utility needs help shaving short-term demand. Operators could then move faster to build more data center capacity and get it connected to the existing grid.

Research at Duke University found that if utilities didn’t have to build enough capacity to cover all conceivable demand and could instead shed load on demand by just 0.25 percent to 1 percent, the U.S. grid could immediately absorb an additional 76 to 126 gigawatts of new large, flexible load. Emerald AI’s technology makes this level of throttling feasible inside data centers by intel­li­gently resched­uling compute with minimal impact on tenants. The result is AI infra­struc­ture that strengthens the grid rather than straining it. 

Emerald AI sprinted through its five demon­stra­tions in under two years, and is now pushing past the demo phase into commercial scaling with its first full-facility deployment. It now serves customers across the AI power value chain, from electric power utilities to leading AI companies to global data center operators, with additional large-scale deployments planned for later this year. Emerald AI partnered with Silicon Valley Power to launch the nation’s first Flexible Load Inter­con­nec­tion Program, which grants data centers expanded grid access in exchange for verified, dispatch­able flexibility. In Manassas, Virginia, Emerald AI is working with Digital Realty and NVIDIA to bring online the world’s first power-flexible AI factory, the nearly 100-megawatt Vera Rubin AI Research Factory, tested in collab­o­ra­tion with EPRI, Dominion, and the PJM Inter­con­nec­tion and slated to come online later this year. In addition, twelve Fortune Global 500 companies are investors in Emerald AI and have joined the company’s strategic advisory board, chaired by NVIDIA.

The big picture here is that the flexibility to temporally or spatially shift compu­ta­tional workloads will allow utilities and operators to solve the peak-sizing constraint, unlocking latent grid capacity and bypassing multi-year trans­mis­sion queues. By tapping into over 100 gigawatts of stranded capacity already on the grid, Emerald AI will not only allow data center construc­tion to proceed faster, but also spare utilities from having to build new emissions-heavy gas peaker plants, and spare data center operators from having to deploy fossil backup power sources behind the meter. This is a huge win for local communities and the environment as well.

Emerald AI accelerates the AI revolution while helping to limit the industry’s greenhouse gas emissions. It’s a combination that fits perfectly with DCVC Energy + Climate’s mission to decarbonize the economy while increasing energy abundance and industrial resilience.