AI is transforming the data center infrastructure equation. As advanced compute platforms push rack densities from tens to hundreds of kilowatts, and potentially 1MW per rack, operators must reconsider not only how power reaches the IT equipment, but also how the resulting heat is removed..
If you look at NVIDIA’s roadmap as an example, you’ll see the kilowatts go up radically. It’s all to do with processing tokens. The more efficiently you can process those tokens, the more money you can make Calvin Nicholson, Legrand.
Infrastructure built for previous generations of hardware is being pushed to its limits, while the requirements of future AI platforms remain difficult to predict. Decisions made today about power distribution, cooling capacity, and facility design will determine how readily data centers can accommodate tomorrow’s workloads..
These pressures are accelerating interest in higher-voltage DC architectures, power sidecars, liquid-cooled busbars, and increasingly application-specific cooling technologies. At AI densities, every power decision has thermal consequences, while every cooling choice affects wider facility design..
In a recent DCD>Broadcast episode, Legrand’s Calvin Nicholson and Nick Schweissguth explore how AI hardware roadmaps are reshaping facility infrastructure, and why preparing for them requires power and cooling to become part of the same conversation.. Turning power into tokens.
AI has disrupted data center operators’ traditional caution around unfamiliar infrastructure. As Nicholson jokes, data center managers “love new products as long as they’re ten years old.”. “If you look at NVIDIA’s roadmap as an example, you’ll see the kilowatts go up radically,” he says.
“It’s all to do with processing tokens. The more efficiently you can process those tokens, the more money you can make.”. Greater compute density can increase token-processing capacity, but only if the supporting electrical and thermal infrastructure can keep up..
Raising the voltage. Open Compute Project (OCP) rack architectures typically operate at approximately 48V to 50V DC, but delivering hundreds of kilowatts at those voltages requires extremely high currents.. Traditionally, it’s been about cooling the IT gear, but now the topic is how you cool the power and networking gear Nick Schweissguth, Legrand.
“The ugly truth of Open Compute and AI is that it just doesn’t scale to these densities,” says Nicholson. “Hence power sidecars and low-voltage DC, either plus or minus 400V or 800V DC applications. They need more power.”.
In transitional architectures, a sidecar can convert AC power to 800V DC before a second conversion brings it back down to approximately 50V inside the IT rack. Even then, current remains extremely high. Nicholson says Legrand is examining liquid-cooled busbars rated at up to 5,000A to support these environments..
Facilities may eventually distribute 800V DC more widely, reducing the number of conversions required, while Nicholson expects architectures to progress toward 1,500V DC as rack demand approaches 1MW.. Adoption will nevertheless depend on breakers, fuses, cables, and connectors keeping pace.
“The devil is always in the details,” Nicholson says. “There are parts that are either barely keeping up or hard to find.”. Cooling beyond the processors.
As electrical systems change, cooling can no longer focus exclusively on processors. Power-conversion equipment may be highly efficient, but even small losses generate significant heat when hundreds of kilowatts pass through it.. “You’re now looking at putting a rear-door heat exchanger on a power sidecar,” says Schweissguth.
“Traditionally, it’s been about cooling the IT gear, but now the topic is how you cool the power and networking gear.”. Liquid cooling is also extending to components such as busbars, while networking equipment may require different cooling strategies than the compute hardware alongside it because of differing thermal loads and design requirements..
Schweissguth expects the industry to move toward fully liquid-cooled IT stacks, but different platforms come with different specifications. Although OCP provides a degree of standardization, operators must still match their cooling strategy to the selected hardware and facility design..
“There’s disparity and diversity in the IT hardware stack and in its requirements,” he says. “A one-size-fits-all approach is not going to be sufficient.”. Designing for an uncertain tomorrow.
Operators may know which systems they intend to deploy today, but future compute platforms could have very different requirements.. There’s disparity and diversity in the IT hardware stack and in its requirements. A one-size-fits-all approach is not going to be sufficient Nick Schweissguth, Legrand.
“What do you do when that AI model is out of date, or you need a multi-trillion-parameter model?” Schweissguth asks. “How does that affect the existing facility? I don’t think we have answered that.”.
One option is to deliberately overprovision power and cooling infrastructure. A facility coolant network can be designed for greater throughput than the initial installation requires, creating capacity for future upgrades. However, the pace of AI development makes it difficult to determine how much will be enough..
Two-phase direct-to-chip cooling could offer another route to adaptability. By using a coolant’s phase change to remove heat, two-phase systems can operate at lower flow rates than comparable single-phase systems.. “You get about a tenfold benefit on your flow rate,” says Schweissguth.
“You can run two-phase liquid cooling off a two-inch pipe, whereas if you’re trying to run a higher-workload system with single-phase cooling, you may need a six-inch or eight-inch pipe.”. The technology may become increasingly important as chip power rises. “At 2,000W, you start to get into some hefty single-phase direct-to-chip cooling processes,” Schweissguth says.
“Once you get into 4,000W per chip, you’re talking about a different ball game.”. Adapting existing environments. Growing demand for compute means previous-generation GPUs and infrastructure can remain commercially useful even as more efficient platforms arrive..
“Nothing’s going to waste,” Nicholson says. “The demand for overall compute is still rapidly growing. If you implement one of the latest solutions and it works, it will continue to work for you.”.
The transition to higher-voltage DC is also unlikely to happen in a single step. Existing facilities may combine AC infrastructure in some areas with DC infrastructure in others. Early 800V implementations could use power shelves and sidecars before progressing toward on-site rectifiers, solid-state transformers, or more direct conversion of grid power..
“We know that we need higher voltage to the rack,” says Schweissguth. “It just depends on what those transitions look like and how that affects everything else downstream.”. It takes a village.
The number of interdependent systems makes it unlikely that any one supplier will provide every component required.. “It’s going to be impossible to try to get everything from one vendor,” says Nicholson. “There are too many different pieces, parts, systems, and specific things you may need.
It’s going to take a village.”. Can you produce fast enough? Can you adjust your production line quickly and efficiently to deliver gear?
You might have had a roadmap before. Now you’re developing in real time Nick Schweissguth, Legrand. Operators must work with design-build firms, integrators, IT manufacturers, and infrastructure specialists to align on project goals and infrastructure requirements.
Trusted relationships become particularly important while equipment and standards are still developing.. Vendors must also adapt quickly, responding to specifications and decisions made in real time rather than following a conventional product development cycle.. “Can you produce fast enough?
Can you adjust your production line quickly and efficiently to deliver gear?” Schweissguth asks. “You might have had a roadmap before. Now you’re developing in real time.”.
Not every data center will need these architectures. Many enterprise facilities will continue operating AC-powered racks at 20kW to 30kW for years. For the highest-density AI platforms, however, escalating power demand and heat flux are removing the option of treating electrical and thermal infrastructure independently..
Power is driving the architectural change, while cooling determines whether that architecture can operate. Planning the next generation of AI infrastructure therefore begins by bringing both into the same conversation.. To hear more about powering and cooling next-generation AI workloads, watch the full DCD>Broadcast episode with Legrand’s Calvin Nicholson and Nick Schweissguth, here..
More from Legrand. 08 Jul 2026. 25 Jun 2026
