Artificial intelligence (AI) is changing businesses’ perception of data centers. Traditionally, these facilities were often seen as a cost center, and CEOs didn’t want to build them. Now, there is growing enthusiasm for their development. They have become a key investment, impacting individuals, industry, and wider society.

During DCD’s AI Week, Greg Stover, global director of hi-tech development at Vertiv, commented: “Historically, the focus was on lowering costs only. Nowadays, data centers are about raising ROI; they are an investment with a return, a business line function, a driving differentiator for companies.”

With growing demand and the increasing power of AI, CEOs now want to and can build AI factories – creating knowledge, information, and intelligence. Together, these centers can generate revenue and, according to Vertiv’s partner Nvidia, “change the world.”

AI changes everything

Jim McGregor, founder and principal analyst at Tirias Research, commented: “AI is changing everything: how we learn, how we work, and how we play. It is impacting our lives. This is a societal change, and technology and innovation are enabling it more than anything else.”

This places the data center industry at a pivotal moment, not just in terms of the scale of data centers themselves or even the industry. It’s the accelerating speed of technological development and the introduction of new technologies.

For a long time, says Martin Olsen, senior vice president of products and solutions at Vertiv, the approach was far more incremental than it is today. The focus was on efficiency from a cost perspective. Today, efficiencies remain high on the agenda – but more from a return on investment (ROI) standpoint, given the immense power involved in compute and infrastructure.

Not everyone wants to operate data centers at a gigawatt scale. However, the demands of AI are prompting questions amongst data center investors and enterprises that may not have considered data centers for a decade – at least not on-premise – about how to incorporate what AI is really becoming: a utility.

Olsen explains: “It's like electricity, like water, and it's infrastructure, right? You have got to think of it that way, because if you don't have it, well, the lights won't be on, and you can't compete anymore.”

Has AI reached a plateau of demand? McGregor finds that it has not because of the widespread impact that AI has on virtually everything we do. AI was once about completing simple tasks, such as managing mobile device batteries. It has now expanded to include AI agents and the use of large language models (LLMs), becoming a routine part of daily life and business functions: there is a need to plan for future demand, new products, and services.

AI: Demand beyond expectations

Wall Street's questions about AI reaching a demand plateau are nowhere near reality. McGregor notes that forecasts predict a 19x increase in AI demand between 2023 and 2024. In reality, however, the increase was closer to 500 percent or more.

“It was monumental,” he remarks, adding that “the number of tokens is growing from 660 trillion tokens generated in 2024 to over 80 quadrillion tokens by 2030, and that's not even including the growth that we're going to see in images, video, and gaming.”

A year or two ago, meeting the demand for graphics processing units (GPUs) was a challenge, with delivery times of 12-18 months. Nowadays, McGregor says a full rack of GPUs can be installed within six to nine months. The trouble is that building data centers has become the real elephant in the room, which can take up to 36 months. There is a wide array of factors to consider, including infrastructure, power, cooling, and networking.

Challenging to keep up

Why does this make it more challenging to keep up with AI demand? Well, the moment anything changes – or when a hyperscaler or a colocation data center requires a data center to operate in a certain way or with specific technologies – the design, construction, and development of the data center become customized.

“We've driven that model for so long that data centers are all customized,” McGregor says, and subsequently, “it takes several years to build a data center.” He continues: “We can't afford that. We need to bring that down to 12 months or less.”

“There can be some pre-planning in terms of land and whatnot, but it's still a lengthy process, and it's usually customized. We've got to get away from that,” concurs Olsen.

Faster deployment required

The fact that AI demand is surging requires the industry to adopt a faster deployment approach that is modular in design, upgradable, and prefabricated to meet data center needs. Data centers must also shift toward being a ‘unit of compute’ by integrating power, cooling, network, and compute into standardized blocks.

Deployment can be accelerated using digital twins, which are critical for real-time simulation and collaboration when pre-verified designs are employed. Olsen and McGregor both agree that industry-wide collaboration and standardization are essential to enable agile, scalable, and environmentally responsible AI-ready critical digital infrastructure.

They emphasize that the future of AI infrastructure depends on data centers achieving energy sovereignty while continuing to work closely with utility companies to maintain power supply as needed. This includes operating microgrids and higher-voltage direct current (DC) distribution to overcome power limitations as AI drives increasing demand.

Future-proofing data centers

A common theme throughout Vertiv’s sessions during AI Week is the need for data centers to make future-ready designs, management practices, and operations. This is crucial as they are gradually transforming into megawatt- and gigawatt-scale AI factories, which require improvements in cooling and power efficiency.

Nvidia, a close partner of Vertiv, is working to address the current data center challenges with an eye toward future needs. “It's basically not a data center, it's an AI factory that we're building, focused on a product; just like any factory, it's producing tokens,” says Dr. Ali Heydari, director of data center cooling and infrastructure at Nvidia.

Cooling is a central part of planning and implementing future-ready data centers and AI factories. “We're seeing racks that are probably good for 50-60 years. They have gone from four kilowatts to maybe 10, 15, and 20kW at most before suddenly jumping to 120, then reaching 600kW and beyond a megawatt,” Heydari adds.

Stover commends Nvidia for understanding the present and future challenges that partners like Vertiv face. “Our job is to consider how we optimize those building blocks, so that as we look at new opportunities – whether five megawatts, 50MW, 500MW, a gigawatt, or seven gigawatts – we can match power and cooling solutions to deploy them rapidly.”

In their discussion, they highlight that liquid cooling should include two-phase systems to support high-density racks. Nvidia is leading efforts to achieve this by working on collaborative projects. To this end, Stover and Heydari also emphasize the role of the Nvidia Partner Network (NPN) in developing scalable reference architectures for consistency and alignment across key stakeholders.

Digital twins and simulation models are seen as essential for optimizing thermal and power designs, cutting time and cost in AI factory deployment. Moving beyond the notion that data centers are primarily cost centers is the realization that mechanical, electrical, and plumbing (MEP) systems are now recognized as value creators. They are essential for enabling AI infrastructure to function efficiently and intelligently.

Engineering thermal resilience for next-gen workloads

The increasing demand for AI and for AI applications is creating a lot of heat. Since they came onto the market, GPUs have taken some of the heat load, pushing heat density to what Steve Madara, vice president of thermal and data centers at Vertiv, calls the “next level.”

“GPUs really forced us to look at using liquid cooling as the means to remove heat from the chip and the server, to get the heat out of the data center,” he says. In fact, he notes that liquid cooling is rapidly replacing air cooling due to soaring GPU power densities.

To address this challenge, Vertiv has collaborated with Nvidia to create the Vertiv-Nvidia GB200 NVL72 platform. According to a joint press release, it provides ‘end-to-end critical power and cooling reference designs for Nvidia Blackwell architectures up to seven megawatts, with OCP infrastructure option’ focused on the need for highly efficient, scalable thermal systems.

Madara suggests future cooling trends will include hybrid models before transitioning to fully liquid-cooled racks, using two-phase fluids to manage higher thermal loads. Meanwhile, there are very few large-scale facilities that are liquid-cooled – “you can count them on your hands right now,” he says.

Nevertheless, there have been valuable learning experiences around piping, flushing, and filling systems, commissioning, and understanding what services are needed to maximize the benefits of liquid cooling. The benefits include raising temperatures on the secondary loop to improve peak PUE and annualized efficiency, as well as recovering some of the stranded capacity. AI pods, he notes, will be roughly 95 percent liquid-cooled and five percent air-cooled for heat rejection.

Best practices in commissioning, fluid cleanliness, and collaboration with partners are essential to support next-gen AI deployments. Thermal fluctuations also need careful management. Madara emphasizes exploring heat reuse opportunities and notes critical considerations:

“There are critical considerations such as the cleanliness of the system, the secondary fluid network (SFN), and all equipment, as well as questions about how we manage load. On the power side, a pod that's one to two megawatts could fluctuate in power requirements.”

Gio Albertazzi, CEO of Vertiv, advises: “Start at the GPU and decide how many generations of GPUs you think your business needs to support to be future-ready, and consider normalization of each GPU and pod configuration because it lends to speed.” These decisions will influence all design factors, including power and cooling.

Energy independence strategies for AI beyond the grid

Without power, no data center can operate, and AI – along with the many applications and services that enterprises, organizations, and individuals rely on – cannot function. It is therefore only logical for AI data centers or AI factories to invest in their own power sources, power generation facilities, and grids to maintain a high level of uptime while reducing the impact of power outages.

Utilities face challenges, such as unpredictable GPU load demands. This drives the need for behavioral models and close collaboration across the supply chain to support a modular, repeatable, and scalable deployment approach, capable of accommodating shifting power demands from CPUs to GPUs, for example, from 50MW to one gigawatt.

Peter Panfil, vice president of global power at Vertiv, explained that traditional power grids struggle to keep up. This has led companies to adopt a ‘bring your own power’ approach, and a move toward onsite generation in various forms. Utilities also face challenges with unpredictable GPU loads, further highlighting the importance of behavioral modeling and close collaboration across the supply chain.

More recently, therefore, much of the emphasis is on grid integration, decarbonization, and power reliability to support the high performance of AI. This is especially critical because AI factories require rapid deployment strategies with infrastructure designed around GPU pod configurations and liquid cooling.

Driving innovation

Finally, Albertazzi emphasizes that Vertiv’s strong partnerships with Nvidia and other collaborators are key to driving investments in innovation, as well as for pushing the boundaries of innovation itself. The creation of AI factories, rather than purely traditional data centers, is a vital part of this process, making them invaluable investments rather than mere cost centers.

With the rapid growth of AI infrastructure demands and rising rack densities, this approach is expected to deliver ROI. Speed of deployment can be supported by Vertiv’s focus on modular, prefabricated solutions and liquid cooling to meet high-density, high-efficiency requirements, which translates to higher, more immediate revenue and ROI.

However, Albertazzi notes that the transformation of data centers is far from complete: the industry requires stronger supply chain resiliency and leadership in design, while commissioning and manufacturing are critical for scaling infrastructure quickly and reliably. To further support this evolution, Vertiv believes power and cooling systems must advance with hybrid models, real-time orchestration, and energy storage to sustain growing AI workloads well into the future.

Explore all content from AI Week, including Vertiv’s sessions, here.