Artificial intelligence is transforming nearly every industry, but perhaps no sector is impacted as much as the data center market. 

Deloitte research estimates that by 2035, power demand from AI data centers in the U.S. could grow more than thirtyfold, reaching 123 gigawatts. For data center developers, this not only means they need to build more facilities faster but also to prepare for things to get hot. 

According to Stulz Director of Technology Dave Meadows, AI-centric data centers call for a new type of cooling technology that a significant number of current facilities may not be prepared for.

“In the U.S. we have thousands of data centers that were designed specifically as low-watt-density, air-cooled centers,” Meadows said. “Introducing high-watt-density, high-heat-flux chip equipment into those environments is very difficult because they were never built for it.”

Stulz provides cooling solutions and services for mission-critical facilities. Meadows walked Bisnow through some of the issues data centers may run into as they prepare for the AI-driven future and spoke on the solutions that will keep legacy centers running in this new era. 

Bisnow: What are some of the top challenges legacy data centers face as they work to meet today’s AI demands? 

Meadows: As I mentioned, retrofitting these facilities for modern requirements has been very challenging. Now, there are strategies in place. One of the most prevalent, and it's really just a stopgap measure to allow us to continue to use these facilities in the time of AI, is liquid-to-air coolant distribution units, or CDUs. With those, operators use liquid to actually cool the AI-specific servers, the Nvidia GB200s or 300s, for example, and they’re pulling that heat out and dissipating it into the traditional white space in a facility so the legacy cooling equipment can take that heat energy, remove it from the white space and get it outside.

It’s not the way we would design a brand-new data center, but it can get us to where we can utilize legacy facilities while meeting the challenge of the AI environment. 


Bisnow: What’s the best solution for legacy data centers that are hoping to move forward with modern cooling products but are facing these challenges? 

Meadows: I believe the best approach is a dual one, where you look at your infrastructure and see what can be taken offline and reworked as a more direct liquid-to-chip-cooled facility, as opposed to purely air-cooled. More than likely, you're going to have chillers in the facility that can be utilized. It’s just about determining how to utilize them under new conditions — the higher return water temperatures, maybe the greater piping runs that will have to be supported to get the chilled water to the CDUs. There are strategies that can be used. It’s got to be well thought out, though.

One of the real challenges today is that we have to build a cooling structure very specific to the information technology equipment we're trying to cool, because it has very special requirements for the flow rate, the supply temperature, coolant type, and that all has to be worked out in the design phase. It’s going to require more engineering involvement than back in the day when everything was air-cooled and we really didn't care what kind of server was there. It needs to be addressed through thoughtful engineering. 


Bisnow: How is Stulz working to help data centers make this complex transition?

Meadows: We're building up our engineering teams, both our application engineering team and our design engineering team, to not only develop new products that work with AI data center cooling infrastructure needs but also to help operators transform their traditionally air-cooled data center into an AI-specific one. This requires customization on the part of the cooling equipment, because you're trying to put a square peg in a round hole to a certain extent. We have to make custom alterations within the equipment so it achieves the goal that we're trying to reach, which is to move the very high heat energy and high heat flux from that white space. 

We’re also assisting the engineering community, specifically the consulting engineering community, by sharing our experiences and what we're seeing in this space. A lot of design engineers are seeing liquid-cooled AI data center requirements for the first time. While we’ve been working on them for several years now, we actually built our first CDU in the U.S. over 20 years ago to cool hybrid liquid-air-cooled racks. We make it a point through our applications engineering team to share our knowledge and help facilities get to the next level of heat removal. 


Bisnow: How should operators prepare for the future of data center cooling? 

Meadows: When it comes to data centers, futureproofing is very difficult — our crystal ball just doesn’t seem to be working very well in this ever-changing field. One of the big challenges we will be facing soon is that Nvidia has made it very clear that we're going to need to go to an 800-volt distribution architecture to power these very high-watt-density racks. We are talking 600 kilowatts and beyond, and to do that, we need special cooling systems.

Water-based systems will not be able to handle this without a serious risk of fire and explosion due to electrolysis of the coolant in the case of a leak in the proximity of 800-volt DC power. This is why there is a push to develop products that can remove high heat flux, high rack density heat, without having the danger of having water in a high-voltage environment. As an industry, we are looking at what we call direct refrigerant-to-chip, where we're using a safe, low GWP refrigerant instead of a water-based solution to remove heat. It’s very early days for this technology, but it’s very promising. 


Bisnow: What’s one thing you wish more people understood about the data center industry? 

Meadows: I would have to say the water issue. We fixate way too much on the water that's used on-site at data centers, and we don't really look at the water that's used in the production of the power that is being consumed in these data centers. We need a more nuanced approach outside of “water is always bad.” Sometimes, we can use water for very efficient cooling methodologies, including direct evaporative cooling or indirect evaporative cooling, but people seem to be under the impression that any cooling method involving water is harmful.

In the meantime, however, large amounts of water are used at the coal-burning power plant, the natural gas power plant, or even the nuclear power plant that's providing that power to that data center. The water used in power generation is often called “source water,” which is different from “site water,” which is typically water consumed at the data center through evaporative cooling. If we can take some of the power requirement off the grid through judicious use of on-site water, we have the potential to use less total water for the data center. There are many variables to look at in the source water-versus-site-water argument. The Green Grid recently came out with the water usage impact metric, which measures a data center's consumed water relative to the local water stress level, and I think that would be a good place to start this important conversation surrounding how much water these facilities really use and where that water comes from. 

This article was produced in collaboration between Stulz USA and Studio B. Bisnow news staff was not involved in the production of this content.

Studio B is Bisnow’s in-house content and design studio. To learn more about how Studio B can help your team, reach out to studio@bisnow.com.