The power infrastructure behind an AI factory can be thought of as a powertrain. This includes everything from the utility connection all the way from grid to chip and, depending on the architecture, high- and medium-voltage switchgear, transformers, low-voltage equipment, and resiliency schemes.

As workload densities and load profiles change, the pace of deployment is faster than ever. This makes the powertrain especially important for AI demands.

“Five to seven years ago, rack densities were around 7-15kW. We’re already deploying racks closer to 100-200kW, and we’re on a trajectory toward 400-600kW, eventually reaching megawatt-class racks” Schneider Electric’s Manav Patel explains to DCD’s Emma Strutton in a recent DCD>Talks episode. “At the same time, we’re dealing with highly dynamic computing loads and enormous pressure to bring capacity online.”

As VP of strategy and business transformation at Schneider Electric, Patel leads strategy for the low-voltage division at the company. When it comes to the powertrain, he’s looking at the situation from a strategy perspective and thinks the industry needs to focus on orchestrating entire systems so they are optimized at scale.

“As data centers have evolved, the importance of the overall system has become more apparent,” he explains. “This is no longer a component problem; it’s an architectural problem. The powertrain is how we turn grid capacity into reliable compute capacity.”

The operator’s juggling act

Operators have a lot to contend with, including issues of power availability, speed of deployment, resilience, and rising rack densities. Given that AI infrastructure generates plenty of revenue, time to power is a significant constraint.

“We need to design around how capacity will actually be energized,” Patel says. “Protection coordination, redundancy, transfer strategies, and system behavior under normal and abnormal conditions should be understood before energization – not discovered during commissioning.”

The objective isn’t simply to get power to the building. Rather, it’s to convert available megawatts into productive compute as quickly, efficiently, and reliably as possible.

“The long-term KPI is really time to compute,” Patel adds. “Power availability makes the project possible; execution determines when that power starts producing compute.”

Designing the powertrain as one system means bringing more people to the table. This is perhaps where the industry needs to change and adapt, rather than traditionally operating through a sequence of handoffs.

“The need for time to compute is becoming more urgent. The people who will build, inspect, operate, and service the powertrain need to be involved much earlier in the project lifecycle,” Patel says. “That means utilities and energy teams, consultants and engineering designers, equipment manufacturers, contractors, electricians, and authorities having jurisdiction all need to work together.”

He adds: “No single player in the value chain can solve this alone. Everyone needs to come together.”

Standardizing processes across the powertrain

The data center industry continues to move very quickly. As AI workloads and rack densities evolve, operators are looking at powertrain designs to prepare for the AI factories of tomorrow. For Patel, this starts with designing around the workload, rather than electrical equipment.

“Start with the compute roadmap,” he says. “The roadmap is evolving very quickly, so we need to understand it and anticipate how a data center’s compute might change in the future.”

Once that’s established, operators can then work backwards to the electrical architecture and future-proof it. Then, it’s about building infrastructure more like a product, prioritizing repeatable power blocks rather than reinventing electrical architecture at every site.

“The objective should be a system that’s globally repeatable but locally compliant,” Patel adds.

He says operators should then simulate before construction begins, creating a model early to gain a competitive advantage. This enables the next step, which is to operate based on data.

“Once the facility is running, the powertrain shouldn’t be static infrastructure; it should tell us what’s happening,” he adds. “Intelligent electrical equipment has provided a tremendous amount of information for decades. That creates an opportunity to understand asset health and move from time-based maintenance to condition-based approaches.”

As the industry cannot predict what architectures will look like in the future, Patel calls for designing a powertrain that can absorb change.

“In the future, architectures may move toward more centralized DC. Some of these approaches will coexist, depending on the site, workload, and phasing,” he explains. “We need to think about future-ready design for the AI factory and anticipate what it might look like five years from now.”

Watch the full DCD>Talks episode with Manav Patel here.