As AI data center developers wait in long lines to hook up to the grid, a growing number are choosing to ditch that process altogether.
Microsoft last month cut a deal with Nscale, which plans to build an off-grid campus in West Virginia’s gas fields. The developers Fermi and Joule are also working on off-grid campuses in Texas and Utah, respectively, primarily powered by fossil gas. Those plans, especially from tech giants with ambitious climate goals, have grabbed headlines and analysts’ attention.
But industry experts — including Joule CEO Brock Andrus — agreed onstage at Latitude Media’s Transition-AI conference last week that these behind-the-meter, off-grid power projects will likely account for only a small fraction of the new generation of systems that ultimately come online to serve AI workloads.
Andrus said that Joule’s customers are still looking for a grid hookup as a “safety net,” even if that comes five years or more down the line. The company broke ground on its site in rural Utah in November, but has yet to publicly announce an offtaker. Joule lined up 455 megawatts of gas generators from Caterpillar and plans to pair them with battery storage to level out electricity; a nearby pipeline connection will supply the gas.
“With the exception of some really talented firms that are working on these off-grid solutions, it’s not going to be a huge part of the market,” predicted Tim Hade, co-founder of Brightfield AI and a former microgrid developer at Scale Microgrids. “It’s going to be sort of a specialty service.”
Gia Clark, director of power marketing at Engie North America, agreed. She said islanded solutions are a symptom of an outdated grid, but she expects most projects to eventually plug in to the larger system. Otherwise, they risk becoming stranded assets as technology advances, fuel prices rise, or climate policy shifts.
Hyperscalers are certainly building on-site, behind-the-meter generation. But it’s unclear how much of the total will be completely islanded from the grid. Last fall, McKinsey found that between a quarter and a third of incremental data center demand through 2030 will be met by BTM solutions. And a recent analysis by Cleanview found that there were 56 gigawatts of BTM power for data centers in the interconnection queue as of February, 90% of which was announced in the last year.
So far, few — if any — off-grid projects at GW scale have been successfully completed, while some developers have run into headwinds.
The Fermi shakeup
Fermi on Friday announced that CEO Toby Neugebauer had stepped down. The company didn’t provide an explanation for his exit, but Fermi has publicly struggled to find an anchor tenant for its Project Matador campus in Amarillo, Texas. Co-founded by former U.S. energy secretary Rick Perry, the company went public in October with the aim to build the world’s largest data center — but has made little progress. The company plans to power the 11-GW “Hypergrid” project with gas-fired power plants and, eventually, advanced nuclear power. Since then, the company’s shares have fallen more than 78%.
In a statement, Fermi said it has started searching for its next CEO and “is evolving from an entrepreneurial start-up culture into a public-company-caliber professional enterprise.”
Neugebauer, during CERAWeek in March, debated the pros and cons of off-grid data centers onstage. He argued that an off-grid site shields residential customers from the costs at a time when affordability is a major concern for Americans. Google’s global head of data center energy, Amanda Peterson Corio, warned that islanded data centers could actually lead to “unintended consequences.”
“When you’re building islands, you have to overbuild the system for the same amount of reliability,” Peterson Corio said. “That’s a lot of investment in gas that is used just a few hours a year and otherwise is sitting idle. That generation could be fed back into the grid at a cost completely paid for by data centers. That’s a missed opportunity.”
Hade, during the TAI conference last week, said the actual construction of a large, off-grid data center is also incredibly complex, particularly in electrical engineering. The first thing engineers want to know is what the load profile looks like, down to the millisecond, across an entire year.
“Then they go design the power electronics, and design the generation on the back of that,” Hade said. “Part of the problem is that no one knows what the electrical characteristics of a 1 GW data center for AI training is. Everyone’s trying to figure it out in real time.”
The electricity demands of AI models are unique in that a cluster of hundreds of thousands of GPUs can ramp from idle to peak power within milliseconds. This creates large load swings that traditional grid equipment and gas turbines aren’t designed to react to so quickly.
Neugebauer, before stepping down as CEO, also told Axios that he may have been naive about how complex the Fermi project is, including the cooling systems for AI chips and “where the supply chain is” for that equipment.
Hade said that a better strategy to help meet 100 GW of estimated load growth from data centers in the U.S. is to increase use of the grid — which operates at 50% of its capacity most of the year — as well as innovate with load flexibility and expand virtual power plants. That will require some energy market reforms at the federal and regional level, some of which are underway.
