BIANCA GIACOBONE | The middle layer of the AI infrastructure boom — the unglamorous one, made of power electronics and their supply chain — has always been critical. But as companies race to secure power, they are also increasingly racing to secure the pieces of equipment that keep that power reliable.
And there’s especially demand for the most advanced versions of electronics like transformers and uninterruptible power systems, or UPSs. Deals for this hardware, especially the versions designed specifically for the large loads waiting to come online, are now framed as strategic.
Yesterday, for example, Crusoe announced that it’s partnering with ON.energy, which builds power systems technology for large loads. Starting this year, the AI infrastructure company and lead developer of the Stargate campus in Abilene, Texas, will deploy five gigawatts of ON.energy’s UPSs at its campuses across the country.
The systems — medium-voltage equipment comprising a backup battery, two power conversion systems, and a transformer — are designed specifically for AI data centers. They sit between the data center and its power source and decouple the two, acting as a shock absorber against both internal GPU load swings that would otherwise destabilize the grid, and external voltage faults that risk damaging sensitive equipment.
As I wrote for the site, this buffering matters because grid-connected AI data centers are emerging as a threat to grid reliability not just because of how much power they draw, but also because of how they behave once connected. AI training requires data centers to ramp their power up and down in waves, which creates dramatic fluctuations in power demand. If unmitigated by systems like the UPS, the fluctuations can propagate back into the transmission system, adding stress to the grid.
Plus, when there are voltage faults or other problems that could end up damaging delicate and expensive hardware, data centers can suddenly drop off the grid entirely. This risks creating large load-loss events, during which a utility suddenly loses GWs of demand all at once, which creates a demand-supply imbalance that can damage grid and generation assets.
(This is precisely what the North American Electric Reliability Corp. warned of when it issued a rare Level 3 alert in May.)
Preventing these issues, as the AI UPS systems do, is starting to become a prerequisite for grid interconnection. ON.energy’s technology, for instance, meets ERCOT’s latest large load interconnection requirements, which mandate that loads of 75 megawatts or more must “ride through” during disturbances, remaining connected or restoring at least 90% of power draw within seconds.
The same pressure to secure innovative hardware is showing up for other types of large loads as well. Latitude Media has learned that IONATE will announce a collaboration this week with General Motors to install its Hybrid Intelligent Transformer technology at one of the automaker’s manufacturing plants in Michigan.
IONATE says the smart transformers can give GM “real-time visibility into power quality and consumption,” helping to prevent power surges and outages. They will be integrated into existing substations and facilities, essentially helping the automaker squeeze more power out of its existing grid connection.
These deals do two things at once: First, they show how companies are refining their power connections, securing them, and making the most of them. But they're also a signal to the rest of the industry that these players have their hands on the hardware, right at a moment when securing everything from electricity to land to the technology to get a facility powered up can be a competition. Voltus senior VP Tim Hade said on last week's episode of Open Circuit that shortages of grid equipment especially is rampant, with some items facing 200-week lead times.
That's going to be a problem for everyone. But as Hade points out, it will be felt most acutely by the many off-grid data center developers promising to build on what he considers unrealistic timelines. "People think about this as like you build a power plant here and then you build a data center here and like voila, you're off-grid," he said. "And that's the furthest thing from the truth. In fact, I would argue that 80% of the difficulty of building one of those projects is the power electronics that sit in the middle."
As Hade himself concedes, many of these developers are keenly aware of the problem. Joulent, an energy developer building power plants for AI data centers, which I wrote about earlier this month, took a different route and partnered with National Grid. This way, as Brian Boland, Joulent's CFO and head of strategy, told me, it hopes to access the utility's electrical expertise and vendor relationships.
"The electrical side of these solutions is very complex," he said. "All the interfaces between the power plant and the data center, and the opportunity that exists in that space, are something that National Grid is extremely experienced in." Joulent also plans to leverage National Grid's vendor relationships for high- and medium-voltage grid equipment, which Boland says has become "the longest lead time item” in the space.
But not everyone will be able to secure a multi-gigawatt deal or a partnership with a major utility. And that casts doubt both on whether we'll actually see the 40 GW of behind-the-meter capacity by 2028 — and more broadly, on whether the AI infrastructure buildout can scale as fast as the industry wants.
