Not so fast, says Emerald AI, which emerged from stealth last week with $24.5 million in seed funding led by Radical Ventures along with a slew of other big name backers, including Nvidia’s venture arm as well as former Secretary of State John Kerry, Google’s chief scientist Jeff Dean, and Kleiner Perkins chair John Doerr. The startup, founded and led by Orsted’s former chief strategy and innovation officer Varun Sivaram, was built to turn data centers from “grid liabilities into flexible assets” by slowing, pausing, or redirecting AI workloads during times of peak energy demand.
Research shows this type of data center load flexibility could unleash nearly 100 gigawatts of grid capacity — the equivalent of four or five Project Stargates and enough to power about 83 million U.S. homes for a year. Such adjustments, Sivaram told me, would be necessary for only about 0.5% of a data center’s total operating time, a fragment so tiny that it renders any resulting training or operating performance dips for AI models essentially negligible.
As impressive as that hypothetical potential is, whether a software product can actually reduce the pressures facing the grid is a high stakes question. The U.S. urgently needs enough energy to serve that data center growth, both to ensure its economic competitiveness and to keep electricity bills affordable for Americans. If an algorithm could help alleviate even some of the urgency of an unprecedented buildout of power plants and transmission infrastructure, well, that’d be a big deal.
While Emerald AI will by no means negate the need to expand and upgrade our energy system, Sivaram told me, the software alone “materially changes the build out needs to meet massive demand expansion,” he said. “It unleashes energy abundance using our existing system.”
Grand as that sounds, the fundamental idea is nothing new. It’s the same concept as a virtual power plant, which coordinates distributed energy resources such as rooftop solar panels, smart thermostats, and electric vehicles to ramp energy supply either up or down in accordance with the grid’s needs.
Adoption of VPPs has lagged far behind their technical potential, however. That’s due to a whole host of policy, regulatory, and market barriers such as a lack of state and utility-level rules around payment structures, insufficient participation incentives for customers and utilities, and limited access to wholesale electricity markets. These programs also depend on widespread customer opt-in to make a real impact on the grid.
“It’s really hard to aggregate enough Nest thermostats to make any kind of dent,”” Sivaram told me. Data centers are different, he said, simply because “they’re enormous, they’re a small city.” They’re also, by nature, virtually controllable and often already interconnected if they’re owned by the same company. Sivaram thinks the potential of flexible data center loads is so promising and the assets themselves so valuable that governments and utilities will opt to organize “bespoke arrangements for data centers to provide their services.”
Sivaram told me he’s also optimistic that utilities will offer data center operators with flexible loads the option to skip the ever-growing interconnection queue, helping hyperscalers get online and turn a profit more quickly.
The potential to jump the queue is not something that utilities have formally advertised as an option, however, although there appears to be growing interest in the idea. An incentive like this will be core to making Emerald AI’s business case work, transmission advocate and president of Grid Strategies Rob Gramlich told me.
Data center developers are spending billions every year on the semiconductor chips powering their AI models, so the typical demand response value proposition — earn a small sum by turning off appliances when the grid is strained — doesn’t apply here. “There’s just not anywhere near enough money in that for a hyperscaler to say, Oh yeah, I’m gonna not run my Nvidia chips for a while to make $200 a megawatt hour. That’s peanuts compared to the bazillions [they] just spent,” Gramlich explained.
For Emerald AI to make a real dent in energy supply and blunt the need for an immediate and enormous grid buildout, a significant number of data center operators will have to adopt the platform. That’s where the partnership with Nvidia comes in handy, Sivaram told me, as the startup is “working with them on the reference architecture” for future AI data centers. “The goal is for all [data centers] to be potentially flexible in the future because there will be a standard reference design,” Sivaram said.
Whether or not data centers will go all in on Nvidia’s design remains to be seen, of course. Hyperscalers have not typically thought of data centers as a flexible asset. Right now, Gramlich said, most are still in the mindset that they need to be operating all 8,760 hours of the year to reach their performance targets.
“Two or three years ago, when we first noticed the surge in AI-driven demand, I talked to every hyperscaler about how flexible they thought they could be, because it seemed intuitive that machine learning might be more flexible than search and streaming,” Gramlich told me. By and large, the response was that while these companies might be interested in exploring flexibility “potentially, maybe, someday,” they were mostly focused on their mandate to get huge amounts of gigawatts online, with little time to explore new data center models.
“Even the ones that are talking about flexibility now, in terms of what they’re actually doing in the market today, they all are demanding 8,760 [hours of operation per year],” Gramlich told me.
Emerald AI is well aware that its business depends on proving to hyperscalers that a degree of flexibility won’t materially impact their operations. Last week, the startup released the results of a pilot demonstration that it ran at an Oracle data center in Phoenix, which proved it was able to reduce power consumption by 25% for three hours during a period of grid stress while still “assuring acceptable customer performance for AI workloads.”
It achieved this by categorizing specific AI tasks — think everything from model training and fine tuning to conversations with chatbots — from high to low priority, indicating the degree to which operations could be slowed while still meeting Oracle’s performance targets. Now, Emerald AI is planning additional, larger-scale demonstrations to showcase its capacity to handle more complex scenarios, such as responding to unexpected grid emergencies.
As transmission planners and hyperscalers alike wait to see more proof validating Emerald AI’s vision of the future, Sivaram is careful to note that his company is not advocating for a halt to energy system expansion. In an increasingly electrified economy, expanding and upgrading the grid will be essential — even if every data center in the world has a flexible load profile.
’We should be building a nationwide transmission system. We should be building out generation. We should be doing grid modernization with grid enhancing technologies,” Sivaram told me. “We just don’t need to overdo it. We don’t need the particularly massive projections that you’re seeing that are going to cause your grandmother’s electricity rates to spike. We can avoid that.”
Kairos’ reactor technology uses a fluoride salt coolant.
Aryna Sabalenka attempts to cool down during her Ladies' Singles semi-final at Wimbledon on Thursday.Julian Finney/Getty Images
