• Vertiv introducing a new medium-voltage BESS/UPS category that combines switchgear, battery storage and UPS intelligence in roughly 3 to 4 MW blocks
• Scott Armul said the move from “DC in the rack” to “DC in the pod” is being driven by density, not preference, as AI systems expand from 72 GPUs in a rack toward 576 and beyond

Vertiv is pushing a redesign of data center power architecture by moving the uninterruptible power supply upstream to medium voltage, arguing that the shift is needed to handle larger AI loads, improve grid interaction and support a new generation of higher-density compute racks.

At an investor event on May 20, Chief Product and Technology Officer Scott Armul said traditional low-voltage UPS systems were built for an earlier era of data centers and are now being strained by the size, volatility and synchrony of AI workloads.

And as utilities impose tighter interconnection requirements and grid capacity becomes scarcer, Armul said data centers increasingly need to behave like “grid connected and grid interactive” assets rather than simple power consumers.

Vertiv’s answer is a new medium-voltage BESS/UPS architecture that combines switchgear, battery storage and UPS intelligence in 3 to 4 MW blocks, allowing the system to buffer load swings, ride through voltage disturbances and integrate more effectively with on-site generation.

Meanwhile, Armul said rack-level power conversion is approaching physical limits at roughly 350 kW to 400 kW per rack because connectors, busbars and copper simply consume too much space as GPU density rises.

Vertiv’s solution is to move those power shelves out of the rack and into an external “sidecar” or power center that feeds the IT load at 800 volts DC. That frees as much as 8 to 16 units of rack space for GPUs, reduces conversion losses and supports densities from about 400 kW to as much as 900 kW, with a path toward megawatt-class racks. Armul said customer lab validation is already underway and that commercial launch is expected in early 2027.

“This is a story of DC power in the rack moving to DC power in the pod,” Armul said.

Cooling criticality

Armul also said the cooling side of AI infrastructure is becoming even more mission-critical than power in some respects.

“Liquid cooling has 1 to 2 seconds of thermal inertia to where if you are out of range or you’re out of tolerance, you are throttling a GPU or you are shutting down,” he said. “It is an incredibly higher degree of mission criticality on liquid cooling compared to air cooling.”

That urgency is helping drive a push toward warmer liquid loops, as noted recently by NVIDIA CEO Jensen Huang, who said NVIDIA’s latest chip designs would use water temperatures around 45 degrees Celsius.

Armul said every one-degree Celsius increase in a traditional chiller loop can deliver roughly a 3% efficiency gain by reducing compressor use and enabling more free cooling. “If I’m operating at 24, 27, 30 degrees Celsius on a primary chiller loop today and I can push that to 45, that is a significant energy savings and a significant peak power reallocation from a data center site to the IT,” he said.

Even so, air cooling is not disappearing. Armul said the latest high-performance AI environments may now be roughly 80% liquid and 20% air, moving over time toward perhaps 90%-10% or 95%-5%, but on a gigawatt campus that still leaves a large installed base of air systems that must be designed alongside liquid loops, adding further complexity to data center designs.