The data centers of today are a very different place from the ones operators were running just a few years ago.

As AI workloads arrive, they fundamentally change the demands being placed on data center infrastructure. Rack densities that once sat in the 15-30kW range are now reaching around 200kW, bringing liquid cooling into the mix alongside traditional air cooling and adding new layers of complexity to power and cooling systems.

And beyond individual systems, as operators look across power infrastructure, cooling infrastructure, and the connections between them, managing the facility as a whole becomes incredibly complicated. At the same time, campuses are scaling rapidly – some reaching 100MW and others moving toward gigawatt scale.

According to Gwenael Guillemoto, global vice president, secure power & cooling and data center services at Schneider Electric, amid this whirlwind of events, high-performing data centers are differentiated by how effectively they are operated, not by technology alone.

Power and cooling translated into infrastructure demands

AI workloads are placing new demands on power and cooling systems. Those much higher rack densities and the increasing cooling complexity are compounded by energy constraints, operational risk, cybersecurity concerns, and pressure to minimize downtime.

For the AI operator looking to make big returns on their facility, power and cooling infrastructure is a strategic business asset. But high-quality equipment and abundant capacity do not, by themselves, translate into high-performing data centers.

There are simply more assets to manage, more data being generated, and fewer skilled people available to deal with it all.

“If you look at any GPU deployment, when you start to run the LLM model, you have a spike of electricity demand. The cooling needs to follow, and you cannot monitor this at the asset level. It doesn't work. You need to have the system view.”

That system view becomes particularly important as operators face growing challenges in finding the people needed to run increasingly complex facilities. The Uptime Institute’s 2025 global data center survey found that 46 percent of operators reported difficulties finding qualified candidates for vacant roles. In its 2026 survey, that figure increased to 53 percent.

A large amount of available capacity has little value if operators cannot understand how their infrastructure is performing, identify emerging issues, and act on them quickly. Infrastructure and operations are therefore becoming fundamentally interconnected, and the key aspects of monitoring and maintenance need to be considered as part of the infrastructure itself.

Build operational resilience into the facility from day one

From the moment operators start to design the physical infrastructure, they ought to think about how that infrastructure will be managed once it is in operation. When designing the digital architecture, the software platform – including all the sensing capabilities where you connect your assets – needs to be integrated so you can monitor and manage efficiency across the asset lifecycle.

That means pushing capex and opex closer together. Guillemoto expects this to become even more important as AI infrastructure scales and speed becomes paramount, making traditional approaches to operations too difficult to sustain. Within five years, he predicts, the industry will be thinking less about simply building data centers (capex) and more about how to operate them efficiently (opex):

“You are ultimately going to spend more on the operational lifecycle of the asset, so that’s why you need to design from day one how you are going to address data center operations from a commissioning and operational standpoint.”

Connectivity is one of the first considerations. Operators need to connect their assets and capture the data required to understand how those assets are performing, as Guillemoto explains:

“The more you are able to standardize the architecture, the more you can standardize the connectivity, and the better you are going to operate your data center.”

Once connected, the assets can generate a much clearer picture of what is happening across the facility. Operators can track everything across maintenance activity, equipment stress, energy use, and other data points over time, helping them understand the health of individual assets and decide when intervention is actually needed.

That is where condition-based maintenance comes in. Instead of sending technicians to equipment just because a scheduled maintenance date has arrived, operators can use real-time information to decide what needs attention, when it needs attention, and what expertise is required.

“When we start the deployment, we start to connect those assets to Schneider EcoStruxure’s digital platform, and on top of that, we deploy all of the EcoCare advisors which drive the end-to-end efficiency.”

AI-powered performance over a lifetime

While connectivity and visibility are important, Guillemoto says it is not enough to monitor individual pieces of equipment. AI workloads can create changes across a facility in a matter of milliseconds.

That is particularly significant because most data centers contain assets from multiple manufacturers, so the operational approach needs to work across the facility rather than being tied to a single brand or system.

As AI data centers scale, the volume of data involved is an increasingly significant consideration. According to Guillemoto, a single phase of a typical multi-phase data center project can encompass around 4,000 assets generating significant volumes of operational data. As further phases are deployed, this creates an increasingly rich data environment, opening significant opportunities for analytics, AI-enabled services, and lifecycle optimization.

Over time, that data creates a historical marker of how the infrastructure has operated, giving teams a reference point when conditions change. Operators can use that history to understand efficiency, identify patterns, and anticipate what might happen next.

Adding predictive analytics and AI takes that a step further, helping identify potential issues before they become failures.

“Connected assets, consolidated data, 24/7 monitoring and expert predictive insights give operators the visibility to manage infrastructure complexity and scale amid talent shortages,” says Guillemoto.

The principle is to connect the assets from the moment the data center begins operating, bringing the resulting information together in a shared digital environment. From there, operators can start to build a picture of how the facility is performing as a whole.

That continuous monitoring also changes how operators think about the health and lifespan of their equipment. Running infrastructure below its potential performance or maintaining it on a fixed schedule can lead to unnecessary interventions and additional wear.

According to Guillemoto, the condition-based approach combined with predictive analytics can also have a measurable impact on operating costs.

“Analysis conducted jointly with customers using operational data collected over several years showed that condition-based maintenance, powered by predictive analytics and AI, can reduce maintenance opex by 20 percent compared with a traditional calendar-based approach,” Guillemoto says, adding:

“If you can extend the life of any asset by 10 percent or one year, this is a massive gain in the financial equation of the customer. You are also able to decrease the intrusive, on-site maintenance by 40 percent.”

The same principle applies at the scale of the industry. Guillemoto points to the hundreds of billions of dollars expected to be invested in AI data center infrastructure and argues that even a small improvement in how those assets are operated could translate into substantial savings.

Those gains also have a sustainability dimension. More efficient operation means better use of the infrastructure already built, reducing unnecessary maintenance and helping operators get more value from the energy and resources invested in the facility.

Extending the useful life of an asset, even by a relatively small amount, can change the economics of the original investment. For operators managing increasingly large, more complex AI facilities, that matters. Turning data into a better understanding of the infrastructure ultimately leads to better decisions about how to operate it.

Can AI infrastructure really run itself?

If today’s AI infrastructure is more intelligent than ever, it might seem reasonable to assume that, once deployed, it should largely be able to manage itself.

Guillemoto doesn’t see it that way. AI infrastructure cannot simply be left to run unattended when the demands placed on it can change so quickly. Digital intelligence can help operators understand what is happening, but it still needs to be connected to the physical systems responding to those changes.

Consider an LLM workload that may sit relatively quietly before a sudden increase in compute demand accelerates electricity consumption, requiring the cooling system to respond just as quickly.

“If you have an LLM model, at some point you work on the model, but when you press the button, the condition moves from zero to 100 percent in a millisecond.”

AI agents can monitor what is happening across the electrical infrastructure, cooling loops, and other systems, but Guillemoto argues that the real value comes from closing the loop between that digital intelligence and the physical infrastructure.

“You cannot disconnect the two,” says Guillemoto.

People remain an important part of that loop. Remote experts can monitor asset conditions across a facility and guide on-site teams, while technicians carry out the physical work. The aim is to make sure the right person is sent to the right asset, with the right information and expertise, when they are actually needed.

This is also where predictive operations are heading. Not every asset has a predictive model in place today, but Guillemoto expects that to change, with predictive capabilities becoming available across most assets over the coming six to 12 months.

The direction of travel is therefore toward more connected assets, more intelligence and more automation, but not an entirely autonomous data center. People still play an indispensable role in turning that intelligence into action.

Who will come out on top?

As AI infrastructure scales, success depends on a whole host of factors – but crucially, it is more than just capacity and deployment speed. For Guillemoto, three measures matter: time to first token, tokens per watt, and tokens per dollar or euro.

The first measures how quickly infrastructure can be deployed and compute brought online, the second looks at energy efficiency, while the third captures the true operational cost of running that compute over the facility’s lifetime.

The focus, therefore, shifts from what gets installed to how infrastructure is operated.

That starts with standardization and designing for operations from day one. Connectivity, sensing, and digital monitoring need to be built into the physical infrastructure, giving operators the visibility and data to understand performance from the moment a facility goes live.

“There is an upfront investment, but the real question is the lifetime economics of the infrastructure. As billions in capex are deployed, operators increasingly need to understand and predict the cost of running those assets over their entire lifecycle. Condition-based maintenance can help create a more predictable operational cost model, while connecting efficiency gains with the broader business and sustainability case.”

Ultimately, the competitive advantage may go to those who can operate, measure, and continuously improve AI infrastructure, not simply build it.

For more information, please visit: Condition‑based maintenance services for data centers | Schneider Electric.