The AI Data Center Boom: Technology, Capacity, and Society

Estimated reading time: 8 minutes

AI may begin with a chip, but its impact ripples across the entire electronics supply chain. From advanced substrates and copper foil to power distribution, thermal management, and data center construction, the race to build AI infrastructure is reshaping manufacturing priorities worldwide, including for PCBs. That is where Schneider Electric becomes important.

Schneider Electric’s Role in AI Infrastructure

As a global energy technology leader, Schneider Electric exists to “electrify, automate, and digitalize every industry, business, and home, driving efficiency and sustainability for all.” One of its most important business areas is supporting data center (DC) facilities and cooling the IT inside them—the infrastructure that enables AI to run.

A modern AI data center is a different animal from data centers built before cloud computing and AI. They must be massive in size to meet the demand that AI and high compute present. They consume a tremendous amount of power, prompting their comparison to large cities and small countries. Importantly, this results in a tremendous amount of heat that must be dissipated at the server rack level as opposed to the facility building level, making traditional air cooling inadequate for the task. By 2030, AI is expected to drive approximately 60% of new data center deployments, necessitating significantly more data center capacity than exists today.

At this time, most large AI data center activity remains in the United States, driven by hyperscalers such as AWS, Meta, Google, and Microsoft. Schneider is involved in most new projects under development worldwide through its partnerships with leading technology providers, hyperscalers, and co-location operators. The company collaborates directly with NVIDIA and others to stay ahead of chip and silicon evolution.

A Closer Look at DCs, Cooling, and Supply Chain

During a recent visit to Schneider Electric's liquid cooling operations in Buffalo, New York, I visited with Andrew Bradner, senior vice president of the cooling business, to discuss how AI has transformed data center design, why liquid cooling has suddenly become a necessity rather than an option, and what these changes mean for the broader electronics ecosystem that supports the AI revolution.

Marcy LaRont: Andrew, it is nice to meet you. Tell us about yourself and your career. How long have you been with Schneider?

Andrew Bradner: I have been with Schneider for nearly 24 years, joining American Power Conversion (APC) in 2002. APC, based out of West Kingston, Rhode Island, was acquired by Schneider Electric in 2007. Prior to that, I owned a human resource management startup software company in California, a precursor to today's software-as-a-service platforms.

I am originally from Rhode Island and moved back several years later to be closer to family. I wanted to continue working in technology, and the closest thing I could find was APC, where technology was part of operations. APC and Schneider have taken me on a journey all over the world, to Shanghai, Beijing, and Barcelona. I am now based in Italy, moving through different parts of the business.

I've been in the data center business for over 20 years, five of which were leading our prefabricated data center business. When you talk about supply chain and the evolution of how data centers are built, I've seen that from the modular side. I have been leading the cooling business for more than seven years.

LaRont: Italy is Schneider’s engineering and industrial hub, but given the big builds in the U.S., you’ve been spending time here. What does that mean in relation to the Motivair facility we toured here in Buffalo?

Bradner: Buffalo is our North American hub for liquid cooling, while Conselve, Italy, remains the location of our largest global cooling facility in terms of employees, manufacturing capacity, and engineering resources. We also maintain engineering and manufacturing centers in Bangalore, India and Zhuhai, China.

The Evolution of Data Center Cooling

LaRont: Now that you’ve been in the server business for more than 20 years, what is the most significant change you have witnessed?

Bradner: The most important change in the server journey is this pivot we began to see about four years ago. Historically, servers operated at power levels that could be effectively managed with fans and heat sinks. For years, CPUs and GPUs remained in the 200–300 watt range, making air cooling a practical and economical solution.

LaRont: We learned in some of the technical sessions here that modern liquid-cooling solutions use less water and power than air cooling. That being the case, why didn't the industry just go to liquid cooling first?

Bradner: Remember, these data centers at the time weren’t built for the optimum data center design. They were built around the chip technology they were supporting. So, the data center design was very agnostic to the IT itself. At that time, air cooling was a logical solution. But now, with the incredible amount of heat that modern AI data servers generate, we must take a different approach to both rack and data center design.

LaRont: Is this a trend that continues to ramp with each new chip design?

Bradner: Yes, the generational shift between chipsets is immense. If you are deploying NVIDIA and deciding whether to deploy a GB200 NVL72 or a GB300 NVL72 rack system, each choice shifts how you need to distribute power and cooling.

Today, liquid cooling has become a necessity. Yet just five years ago, we were still looking at liquid cooling from a sustainability angle. We knew it would happen at some point, but weren’t sure when.

Liquid cooling existed before AI, particularly in parts of Asia where climate and efficiency concerns made it attractive. However, the public launch of ChatGPT in late 2022 was an inflection point, accelerating AI adoption dramatically and turning liquid cooling from an efficiency strategy into an operational necessity.

LaRont: Is there more upfront OEM involvement in data center and infrastructure design than there had been previously?

Bradner: Yes, these changes are being driven by the OEMs, which have significantly affected the white-space design of servers. These server racks are being built plug-and-play ready for liquid cooling. (“White space” refers to the area inside a data center that is available for IT equipment: servers, storage, networking gear, and racks.)

We used to have thousands of server vendors. At one point, the guiding intelligence was that only 25% of servers would ever be liquid-cooled. Now, new server designs are released only with liquid cooling built in. That's what has driven the industry’s shift toward liquid cooling.

This also changes the supply chain side. Today, only a few qualified organizations can build the right kind of server rack and data center facilities. OEMs like NVIDIA are working with that market, helping to bring up more qualified vendors.

Retrofitting Existing Data Centers

LaRont: As these new data centers emerge, what will happen to all the old server farms? Will they always have utility as air-cooling facilities, be retrofitted for liquid cooling, or go defunct and become idle?

Bradner: The older server facilities were generally built for enterprise computing, not the newer cloud-based applications that require more power and generate more heat. It's difficult to retrofit such facilities, but not impossible.

On the Motivair tour, you saw some heat-dissipation units at the end of the production line that were being developed into liquid-to-air coolant distribution units (CDUs). Those are often seen in brownfield or existing air-cooled data centers that don't have facility water going into their white space.

LaRont: Theoretically, then, at least some of those older data centers could still be useful.

Bradner: Yes, where you could deploy liquid cooling into those environments, modernization will be a big part of what happens. The access to power that those sites already have is primary for any data center location. You can leverage some of the electrical infrastructure even if you must rebuild into a more modern facility.

You may need to modernize some of the cooling infrastructure and look more at the physical build. What's the slab density? If it's multi-story, what weight can it manage? The weight that must be supported and the amount of steel and concrete infrastructure involved are different and much more substantial than what you would have for an existing enterprise compute data center.

We are seeing the larger companies do this at some of their older cloud computing sites.

LaRont: Will circularity become a bigger part of the conversation again soon?

Bradner: Yes, circularity is important, and part of Schneider's mission. When it comes to facility modernization, we try to minimize the work required and use what is already in place as much as possible. Even the retrofits are focused on the smaller parts that can be changed out rather than the housing and metalwork.

We are also paying attention to the expected lifecycle of the data center as chip technology continues to advance. Extending a center’s lifecycle, especially as AI is applied to more and different use cases, is an important question.

Where AI Data Centers Are Being Built

LaRont: Where do you see the greatest geographic push for new data center builds?

Bradner: North America currently leads AI deployments, driven by access to advanced GPUs and significantly higher rack densities than many other regions. It is the difference between 132–600 kilowatts historically, to racks now pushing around 1 megawatt, then to skidded solutions enabled by modular CDUs ranging from 105 kW to 2.5 MW, with Motivair by Schneider Electric’s liquid cooling portfolio scalable to 10 MW and beyond to meet the demands of next-generation AI data centers. So, the scale of deployments is very different. We're starting to see that change in Southeast Asia and Europe as chip availability shifts.

LaRont: What is an optimum power arrangement for a modern AI data center?

Bradner: The answer is not simple, and differs geographically, by climate, and by state. AI data center development is underway in Texas, Virginia, Nevada, and Phoenix. Each state has different rules and regulations as to how it can build and where it can access power. Some can build their own substations; others cannot.

Outside the U.S., Saudi Arabia is a good example of a country funding access to power, much of it through gas turbines that drive desalination, providing low-cost power for that region. The Nordics have strong access to renewable energy, especially hydro. So, each place has different dynamics that drive the “optimum” data center design for power.

LaRont: What do you think about Phoenix specifically? We have a lot of semiconductor development there, but we have issues with water and power. Our reality is that we are 115°F for several weeks out of the year, so, to the average person, development here seems counterintuitive.

Bradner: Phoenix has good access to power and land. The trade-off is that you need more mechanical cooling year-round. But the desert is dry, which helps cooling, and in the evenings, temperatures can drop 20 or more degrees. Those big swings are advantageous for natural cooling.

AI data centers are working with hot water temperatures of 40°C (104°F). A location that lets you take advantage of free cooling hours and days throughout the year is very important.

A regular person is considering the worst-case temperatures in the middle of the day. They want to keep their house around 70°F. That isn’t needed for a data center. If it drops to 90°F at night, a data center can take advantage of that delta, providing cooling without using compressors or water to cool the chip. That wouldn’t be comfortable for you and me, but it’s fine for the IT equipment.

Building a Resilient Supply Chain

LaRont: With manufacturing sites in many countries, what does Schneider’s primary supply chain look like? Do you have local supply chain partners for your different manufacturing locations?

Bradner: We rely on a combination of global and local suppliers, emphasizing redundancy through what Schneider calls the "Power of Two": multiple manufacturing locations and multi-sourced supply chains. We try to have some portion of the build that's locally sourced, but that isn’t always possible. If we face an issue such as a natural disaster, our approach mitigates risk. Motivair experienced that last year in the U.S., but their affected local supplier also had a factory in Germany that could supply back to the U.S. It didn't disrupt Schneider’s ability to build and ship product.

LaRont: Andrew, it has been a pleasure meeting and talking with you. This has been a real learning experience for me. Thank you.