Across the data center industry, the race to support AI is increasingly being measured in megawatts. As workloads become more demanding and compute densities continue to rise, operators are having to think carefully about how power is delivered and, crucially, how effectively it can be used.

While the conversation often starts at the facility level, the real challenge begins much closer to the rack.

For nVent, the rack itself is becoming a critical point of control. As AI workloads introduce rapid fluctuations in power demand, the ability to monitor and respond in real time will increasingly determine how effectively secured capacity can be utilized.

“The ceiling isn’t really about the silicon,” says David Wood, senior product manager of smart power at nVent. “It’s about power, and specifically, how intelligently you can deliver it.”

With more than 20 years of experience in the data center industry, Wood has witnessed the evolution of power infrastructure first-hand. Now, in a recent DCD>Broadcast episode, he shares why the next step will be about giving the rack the intelligence to respond to increasingly dynamic workloads.

The rack-level power problem

Data center capacity is often discussed in terms of the overall power available to a facility. Operators may have 20, 50 or even more than 100MW at their disposal, but this doesn’t necessarily mean all this capacity can be put to work. As Wood explains:

“The constraint almost never shows up at this scale. It’s evident inside a single rack.”

At the rack level, provisioned power is either converted into useful compute or becomes stranded because something further downstream can’t respond quickly enough to the demands placed upon it.

Traditional enterprise workloads were relatively predictable, allowing infrastructure to be designed around stable patterns of consumption. AI workloads are an entirely different beast.

GPU clusters can rapidly move between idle and full power draw as workloads shift between training, inference, and communication cycles. Crucially, these changes can happen within milliseconds, creating a level of volatility that conventional infrastructure wasn’t designed to manage.

“This volatility is where the real damage happens,” says Wood. “Whether it’s thermal spikes, phase imbalances, or circuit breaker trips – when the power delivery can’t keep up, your own infrastructure becomes the bottleneck.”

This reality is driving a fundamental shift in power infrastructure. Rather than sizing systems to meet peak demand, operators increasingly need to manage power dynamically as workloads change.

From monitoring to real-time intelligence

The problem with responding to these events effectively is that they can happen faster than traditional monitoring systems can detect them.

Conventional power distribution units (PDUs) typically rely on polling, where readings are requested at intervals of a second or more. This can provide a useful overview of consumption, but it risks missing the more transient events that matter most.

“If you’re polling every second or every few seconds, and you have a millisecond-scale swing, that reading is a snapshot of something that’s already over,” says Wood.

nVent’s approach is based on subsecond sampling and live streaming of power data. Beyond providing a record of what’s happened, the system is designed to provide visibility into how the workload is behaving as it happens.

“You can’t automate what you can’t see in real time,” adds Wood.

The engine turning data into action

This is where nVent’s Intelligent PDU Platform with built-in event engine secures its place. Traditional metering can deliver accurate information about power consumption and quality, including billing-grade measurements, but this insight is still retrospective. The event engine is designed to take this information and put it into action to trigger an immediate response.

“The event engine can act on what’s happening right now because it runs on the PDU itself,” explains Wood.

The principle is relatively straightforward. An operator establishes a trigger, defines the condition under which it should be sustained, and then specifies the action that follows.

For example, a phase imbalance could be allowed to persist for a defined period before it exceeds a specified threshold and triggers an automated response. This helps prevent the system from reacting to momentary noise while ensuring it responds when it really matters.

Beyond the PDU

For nVent, the longer-term opportunity lies in connecting power intelligence with the wider data center environment. Once the rack can understand its own power behavior and respond accordingly, this intelligence can be linked with other infrastructure systems, including critical cooling and workload management. This creates the possibility of coordinating systems around what the workload actually needs.

“We’re looking at it from the ground up at the rack level, to help secure better coordination across the entire facility,” says Wood.

nVent’s position is strengthened by its presence across both power and cooling. Rather than treating the two as entirely separate systems, the company is exploring how they can work together to anticipate workload requirements and respond accordingly.

On a broader level, rack-level intelligence can also feed into building management systems and data center infrastructure management software. What starts out as a local response can become part of a wider optimization strategy across the data hall.

This can help deliver more usable capacity from the same footprint, fewer nuisance trips, and closer coordination between power, cooling, and compute.

The shift towards self-healing infrastructure

The ability to respond automatically is central to another concept Wood highlights: self-healing infrastructure. The term describes a move away from systems that simply identify a problem and notify an operator, towards more proactive infrastructure.

“Most of the infrastructure components and solutions that we use today, especially in power, tell you something went wrong,” says Wood. “What self-healing means is that the power distribution itself can do something about it before you even see the alert.”

Environmental monitoring provides another example. If sensors detect a temperature increase within a cabinet, predefined thresholds could trigger a response such as shutting down non-critical loads, reducing the heat being generated, and giving cooling systems time to react.

Inside high-density environments containing huge volumes of high-value equipment, the stakes are elevated. Consequently, for operators, the appeal of self-healing infrastructure is intrinsically tied to protecting valuable physical and computational investments.

The bottom line

As rack densities continue to rise, however, intelligent power management will also need to evolve alongside the underlying architecture.

Wood points towards a future in which higher-voltage DC power is delivered directly at the rack, with architectures moving towards 400V and potentially 800V.

By reducing the number of power conversion stages between the source and the compute hardware, operators can potentially reduce losses while freeing up physical space to dedicate to compute.

“It’s a real architectural shift, and it doesn’t happen overnight,” says Wood. “But it’s coming, and rack intelligence has to move with it.”

Ultimately, the argument for intelligent power comes down to the fundamental reality that secured capacity is only valuable if it can actually be deployed effectively.

“The risk has never really been about running out of power. It’s about stranding the power you already have and not making use of it because the rack can’t act fast enough,” he continues.

This makes the rack an increasingly important part of data center power strategies. With real-time visibility, automated responses, and close integration between cooling and workload management, power distribution can become an active participant in infrastructure optimization.

For Wood, the ultimate goal is to ensure that the enormous investments being made in AI infrastructure deliver as much useful compute as possible.

“We invest a lot of money in our data centers today, and it’s really important to be able to make the most of it,” he concludes. “nVent is here to help extract the greatest value out of that investment and to be responsible in our environmental approaches by utilizing our resources most efficiently.”

To hear more from David Wood on intelligent power for AI operations watch the full DCD>Broadcast episode, here.