In the early days of enterprise computing, a typical data center rack drew roughly 1 kW, a load light enough to manage with minimal cooling infrastructure. Today, the average rack density has climbed to 27 kW, nearly four and a half times more than it was just five years ago (AFCOM 2026 State of the Data Center Report). That trajectory is already impacting data center development planning, as operators plan to deploy AI-capable infrastructure faster than their facilities can currently support it.

This kind of growth is exposing a bottleneck that has received far less attention than chips or new software: the physical infrastructure needed to deliver, convert and effectively manage that power. To close that gap, manufacturers must rethink the physical systems required to safely support ever-climbing rack densities.

A shift from compute to the physical layer

Much of the public conversation about scaling AI has centered on model performance, GPU availability and total energy demand. But manufacturers building the equipment inside AI infrastructure—including power delivery systems, conductors, connectors, transformers, cooling and control systems—face a more immediate constraint: fitting dramatically more power and heat management into a physical footprint that isn't growing at the same pace.

That pressure is amplified by what is happening on the grid side. Utilities providers already face long interconnection queues and permit delays, and much of the demand they’re planning for is still speculative. Some operators are turning to on-site generation to work around these constraints, but that only adds pressure inside facilities. Together, these dynamics mean manufacturers can't design toward a single, fixed density target. They need equipment that can flex as real demand becomes clearer.

Early industry responses show what this looks like in practice. Silicon photonics, which swaps electrical wiring for light-based connections to move data between chips, is helping some networking hardware manufacturers reduce power consumption as much as 3.5 times. Power architecture is evolving just as fast, with manufacturers adapting next-generation transformers and higher-capacity conductors built for greater rack density. These represent a new category of design challenges, and manufacturers engineering for them now won't have to retrofit equipment once standards catch up.

Treating power, cooling, and controls as one design problem

Traditionally, power delivery, cooling and controls have been engineered, sourced and optimized independently. That approach no longer works. At higher rack densities, all three systems are drawing on the same limited power and cooling capacity. A decision in one system quickly becomes a constraint on the others, turning what used to be manageable trade-offs into full deployment delays.

The manufacturers navigating this most effectively are building power delivery, cooling and controls into a single, systems-level architecture rather than treating them as separate components. A transformer engineered without the cooling system in mind, or a control system built without accounting for power delivery, only shifts the strain elsewhere in the stack. That interdependence is what sets this moment apart from past infrastructure upgrades, when individual components could be improved one at a time without much consequence for the rest of the system.

The competitive risks of standing still

AI-capable capacity increasingly determines what a manufacturer can promise its customers across turnaround times, production volumes and new product timelines. Those promises depend on infrastructure decisions made months or years earlier.

A transformer sized for yesterday's rack density or a cooling system with no headroom are limits that get set well before a retrofit is planned. Manufacturers who treat physical infrastructure as secondary are setting those limits without knowing it, and their customers are the ones who find out when a deployment slips.

As AI-capable infrastructure continues to scale at unprecedented levels, capacity planning must account for the physical layer with the same rigor applied to compute. The manufacturers positioning power, cooling and controls as part of that planning process now are the ones with room to grow into it.