Solving for AI’s power problem

The AI race has long been defined by how quickly firms can stand up GPUs. Now, Elon Musk and other leaders say they expect physical bottlenecks, including a shortage of electricity needed to power data centers, to slow down competitors. As utilities warn of multiyear waits for new connections, “speed to power,” or how quickly a data center can get electricity flowing, has become an industry mantra.

The Information editor and writer of The Electric newsletter Steve LeVine moderated a discussion with three leaders pursuing different solutions for smarter architecture, from the grid to the server rack.

• William Mao, Vice President of Energy Infrastructure, Delta Electronics
• Allan Schurr, Chief Commercial Officer, ERock
• Brian Janous, Co-founder, Cloverleaf Infrastructure

The capacity myth

According to Janous, the problem is less about finding gigawatts and more about optimizing them. The grid is underused by about 45% on average, Janous said, because it has to accommodate peaks in usage.

“When a utility says ‘there’s no power for five years,’ what they’re saying is: we don’t have the capacity to meet the coldest winter nights and hottest summer days,” he said, adding that for most hours of the day, there’s enough energy to meet data center needs. (Some notable startups including GridCARE, Emerald AI, Cloverleaf itself as well as Nvidia are trying to figure out how to use such spare capacity.)

Off-grid versus microgrid

Some developers are building their own power plants rather than wait on public utilities’ timelines. But Schurr said going off-grid isn’t necessarily a good long-term solution for data centers.

“It’s a very high-carbon, high-cost commitment,” he said. Instead, he promoted the idea of building a microgrid—a localized power source powered by any number of sources—as a bridge until grid connections arrive. After the data center joins the public grid, the microgrid can still cover peak demand and keep the lights on if the grid goes down.

“It takes three to four years to build a transmission substation,” Schurr said. “We’re building 350 MW microgrid in 16 months.”

Betting on the grid

Cloverleaf looks for sites with spare grid capacity and develops “powered land” around them. Janous said the company has delivered about 7 GW of sites in 18 months, with another 10–12 GW in the pipeline.

Janous agreed with Schurr that going it alone has pitfalls, noting the grid provides services that are hard to replicate, like balancing supply and demand to keep power stable, and restarting the system after a blackout.

“If you have to recreate that behind the meter [in other words, off the public grid,] and think, ‘all I’m doing is building a single power plant,’ you’re in for a rude awakening,” he said.

Simplifying the path to the rack

Mao has watched power infrastructure go from an afterthought to a hot topic.

“For a long time, the electron was not so interesting. Now the space is actually becoming pretty sexy because there’s a lot of demand for the sexy electrons,” he said. That demand puts a spotlight on what happens after power reaches the data center’s door. Delta is tackling that challenge between the grid to the rack.

Mao explained that power typically passes through multiple conversion stages before reaching a server rack. Delta’s solid-state transformers convert medium-voltage AC to the 800-volt DC used by racks in fewer stages, potentially reducing energy losses, as well as reducing the number points of failure.

“Not only do we reduce the number of steps in the distribution chain, but also increase the reliability by removing some of the intermediate power layers,” Mao said.

Solving for AI surges

In the past, data centers traditionally drew steady, predictable power. AI workloads, noted Mao, have changed the game.

“Today with AI, loads no longer measure in kilowatts but in megawatts, and they fluctuate in milliseconds,” he said. Those swings strain the grid.

Managing them takes more sophisticated controls and buffering, he said, adding that Delta’s control platform is built to coordinate different power sources, from the utility grid power, natural gas generation, and to solar and fuel cells. Mao argued that the cleanup should fall to the data center itself, not the utility or its other customers.

Looking ahead, Mao expects the most common setup for large sites to pair natural gas generators with energy storage. Nuclear will play a role too, he said, pointing to a 20-year deal Microsoft signed with Constellation Energy to buy power from a restarted Three Mile Island reactor in Pennsylvania for its data centers.

“You have to have these large blocks of power to move the needle substantially,” he said.