The cost of a delayed data center is enormous: roughly $3 million per megawatt in revenue loss per month according to some industry estimates. In popular markets, the queue to get connected to the grid can stretch for many years, which means those opportunity costs pile up.
So as much as they might want to avoid complicating contracts and systems that have provided reliable compute for years, data centers are being forced to embrace flexibility.
The open question, however, is how much and what kind. According to experts operating at the AI-energy nexus, who spoke onstage at Latitude Media’s inaugural Flex Summit last week, answering it is complicated by the fact that people who work primarily in the energy world underestimate the complexity of data centers, and vice versa.
“I think a lot of conversations between those two counterparties miss the mark because folks don’t fully recognize that they’re talking to somebody different than they might think they are,” Jake Elder, SVP of research and innovation at Energy Impact Partners, told the roomful of people who mostly worked in distributed energy.
Flexibility can exist either inside or outside the walls of a particular data center. A data center could, at least in theory, flex everything from cooling or power distribution to compute itself. They can also tack on on-site energy systems, anything from batteries to generators, in order to draw less from the grid in times of high demand.
But regardless of how it’s achieved, the idea is that flexibility — or being “a good grid citizen,” as compute flexibility startup Emerald AI calls it — comes with benefits: quicker interconnection, most immediately, as well as keeping the grid stable even as more loads get hooked up.
As data centers grapple with how to trade flexibility for speed, practically speaking, the industry is finding that not all flexibility approaches are created equal. As Peter Hans Hirschboeck, founder and managing director at impactECI, put it, “of the types of flexibility, the energy industry doesn’t understand that compute is probably the hardest.” (Incidentally, Hirschboeck is also an advisor to Emerald AI.)
Elder concurred, adding that the majority of projects that he’s seen make the speed-for-flexibility trade have relied on on-site power, either front-of-meter or behind-the meter. That’s because “it doesn’t require any real conversation with the end customer.”
“Your engineering teams don’t have to get comfortable with messing with your cooling processes. You don’t have to deal with any of the SLA conversations around workloads,” Elder said. “You can just kind of turnkey transact.”
One prominent example is Aligned Data Centers, which signed a deal in 2025 with Portland General Electric to bring a 31-megawatt, 62-megawatt-hour battery alongside its facility, in exchange for a faster interconnection. While this is one of just a few public deals, Elder said he’s seen a number of private bilateral agreements to do the same thing.
(There’s also a growing push to use virtual power plants as a part of a data center’s capacity stack, though that’s more about demonstrating that there’s enough room on the grid for the data center to get connected, rather than flexing the facility’s power draw on the day-to-day.)
That said, the industry is hampered by the lack of a shared definition for flexibility. What duration of curtailment should be required? Should there be response time or location standards? When it comes to formalizing flexibility as a part of the grid planning process or even interconnection studies for a single facility, the uncertainty is a problem. That’s according to Seyed Madaeni, co-founder and CEO of Verse.
“Do I cut the load for two hours, four hours, six hours, eight hours? Because that impacts the amount of capex that we need to bring in,” Madaeni said. “There’s not a one-size-fits-all [because] every location is different. There might be a capacity problem. There might be a transmission problem.”
Depending on how much flexibility is needed — would the facility need to decrease power draw for a few minutes, or a few hours? — the solutions might look different. According to one Energy Impact Partners analysis, if a service territory has 200 hours of disruption per year, with a mix of lengths and depths, about half could be met with workload flexibility, while batteries could make up most of the rest, Elder said, with “maybe one outlier event per year” that requires something more dispatchable like fossil capacity.
“But that’s just one indicative study,” he added. “It’s very much going to depend on the site and the market. And over time, as we clip the far left side of the load duration curve, it’s just going to get harder.”
In practically evaluating flexibility solutions for a data center, Hirschboeck considers their cost (including revenue loss), capability (short- versus long-duration), benefit, and second-order effects like the pollution that comes from a gas generator.
Ultimately, the panelists largely agreed that the majority of data centers coming online by 2030 will incorporate at least one if not several versions of flexibility. According to Nima Amirzadeh, chief operating officer of LōD Technologies, “Flexibility is not a one or other solution. It’s a combination of a software layer, a battery layer, and a generation behind the meter, or some sort of backup generation.” No one can afford to be picky about their solution of choice, because getting connected to the grid is simply much harder than it ever has been before. If data centers don’t embrace flexibility, as Madaeni said, they’ll be forced to go off-grid instead.
But in the meantime, there are kinks to be worked out, both technically and definitionally.
“There is absolutely a path for operations here,” Madaeni added. “I think what’s not clear, which creates uncertainty, is what are the rules around this.”
