AI Heavyweights Team Up to Promote Demand Flexibility

Nvidia, Google, and data center software startup Emerald AI are teaming up to lead the AI Energy Management Alliance, a trade group dedicated to promoting flexible load for AI data centers, the organizations announced on Wednesday.

“There are a lot of AI trade associations, data center trade associations, energy trade associations. This is the only one that is laser focused on flexible AI data centers,” Varun Sivaram, founder and chief executive of Emerald AI, told reporters in a briefing earlier in the week.

The group represents the evolution of an older trade association, the Advanced Energy Management Alliance, which was founded in 2014 and advocated for demand response for large electricity customers. Energy policy veteran Frank Lacey will lead the reconstituted group, which will also include other energy and AI heavyweights among its members, such as Anthropic, NRG, and Constellation Energy.

The pursuit of policies and technologies that can enable data centers to reduce their draw on the grid during moments of peak demand has been something of a holy grail for energy policy practitioners and hyperscalers like Google. That’s because much of the cost of building out and maintaining the grid — including greenhouse gas emitting gas-fired powered plants — is for meeting those peak hours.

“The savings to consumers if we had effective flexibility is enormous,” Abraham Silverman, former general counsel at the New Jersey Board of Public Utilities and assistant research scholar with the Ralph O’Connor Sustainable Energy Institute at Johns Hopkins University, told me. (He is not involved with the alliance.) “It’s when you get up to the hottest or coldest day of the year that you need that extra transmission line or need to build a new one,” which then drives up costs for everyone, Silverman said.

A recent Johns Hopkins analysis of the grid operator PJM Interconnection, which covers large portions of the Mid-Atlantic and Midwest, found that “requiring data centers to accept occasional power interruptions saves over $15 billion per year.”

State and local regulators have shown openness to a variety of approaches that could get data centers on the grid faster while minimizing impact on the grid. “Just about every state has some either legislative or regulatory process for looking at this,” Silverman said.

The case for flexibility picked up steam last year thanks to an academic paper co-authored by energy systems expert Tyler Norris, who at the time was a researcher at Duke University’s Nicholas School of the Environment and is now Google’s head of energy market innovation. Norris argued that much of AI data center electricity demand could be served by the existing grid with modest flexibility.

“The limiting factor for new digital infrastructure isn't capital or silicon; it's power,” Norris told the reporters during the briefing. “But the biggest near-term barrier isn't a lack of electricity. Multiple studies have found that if new loads are able to reduce their draw from the grid for a small fraction of the year — less than 100 hours during peak periods — we can add dozens of gigawatts of new load to the existing U.S. power system.”

Google says it has 1 gigawatt of demand flexibility integrated into its existing utility contracts, while Emerald, which recently fetched a valuation of just over $1 billion, is working on a 100-megawatt data center with Digital Realty and Nvidia in Virginia. That facility “is intended to demonstrate a model that future AI factories around the world can adopt,” Josh Parker, the head of sustainability at Nvidia, told reporters on the call.

“What we want to do is to better utilize that infrastructure,” Parker added. “Every watt wasted is a watt that could have been used to generate tokens, which could lead to life-saving treatments, or economic productivity, or even energy efficiency in other sectors that dramatically improve our sustainability outcomes.”

The effort is especially noteworthy because it explicitly seeks to make building data centers easier amidst mounting and diffuse skepticism from the communities that may host them and the public as a whole. Part of the case for flexible load is to solve for the mounting utility bills widely predicted to accompany AI’s expansion.

“The real goal,” Sivaram said, is “more community-friendly and grid-friendly data centers — data centers that are good grid citizens all across the country.”

Concern over the energy system’s ability to meet the demand from AI has reached the federal level. In June, the Federal Energy Regulatory Commission asked the six large independent power markets to come up with reforms to help protect the grid and consumers from the huge predicted rise in demand from data centers. Those include coming up with “new transmission services to reflect large load flexibility,” as FERC Chair Laura Swett put it. A trade group focused on flexibility could push forward these conversations in a coordinated way, ideally bringing together state and federal regulators, Silverman told me.

Right now, any effort to reform data center interconnection tends to ping pong back and forth between states, the federal government, and the regional transmission organizations, with major players trying to have their case heard at whatever level they think will be most favorable to their interests. For example, Microsoft is contesting a Virginia rule over data center cost allocation, claiming it stands to get in the way of federal rules.

“This new trade association could be very helpful in bringing together the companies for whom flexibility is a competitive strength and give them a voice,” Silverman told me.