As the data center industry rushes toward AI-driven computing, its foundations in compute, storage, and networking are being redefined. DCD’s recent 2026 Trends and Outlooks Broadcast saw industry leaders evaluating the physical and economic realities of modern data centers as AI shapes every layer of digital infrastructure.

Martin Olsen, leader of Vertiv’s data center segment strategy and deployment, operates at the critical intersection between silicon partners, such as Nvidia, and Vertiv’s own research and development teams. As he puts it:

“I make sure that we stop some of the clever pilots from running in perpetuity and turn them into scalable, productive solutions or products that our customers can actually use.”

He is joined by Jim McGregor, founder and principal analyst at North American custom market research firm Tirias Research, serving customers across the technology stack.

Together, McGregor and Olsen examined the rise of AI not merely in terms of models, tokens, or outputs, but through the lens of physical infrastructure in 2026 and beyond.

AI-accelerated infrastructure trends

AI demand is as insatiable as ever, and the industry is still only in the early stages of development and maturity. While the number of AI experts, use cases, and models continues to grow exponentially, the infrastructure required to support them is increasingly constrained and in deficit. Power density is one of the most pressing challenges. As Olsen notes:

“We’re now operating at 140kW, and with Nvidia just announcing their Vera Rubin generation that will take us straight to 250kW, the graphic processing units (GPUs) must be clustered so closely together because of the connectivity between them, which really puts a strain on the data center.”

These tightly packed configurations create significant challenges: an enormous amount of power must be delivered within an exceptionally small footprint. Feeding such concentrated loads reshapes data center design, forcing chillers and large-scale electrical equipment to sprawl outward, and introducing challenges in layout, construction, operations, and maintenance.

At these densities, traditional air cooling alone is insufficient, making hybrid solutions with liquid cooling a necessity – adding another layer of complexity as operators begin to work the two into the same environment.

As cooling technology is forced to evolve in parallel with increasing power demand, Olsen highlights Nvidia CEO Jensen Huang’s CES 2026 announcement outlining a move toward warmer liquid-cooling technologies. While this will not eliminate the need for chillers entirely, it opens the door to broader adoption of alternative heat-rejection technologies, such as dry coolers and trim coolers, improving overall efficiency and returning valuable capacity to the data center.

Together, these pressures underscore how AI is accelerating infrastructure trends and pushing data center design into entirely new territory.

The next steps in density management

Densification is coming from every aspect of the industry at once – the memory, networking, processing, and AI acceleration – each advancing on its own trajectory but requiring greater alignment. As McGregor explains:

“We have to bring these elements closer together because we’re running into limitations in semiconductor processes, power distribution, thermal limits, and so many more areas within the same rack architecture. For the first time, our industry is finally looking toward the future and trying not to be reactionary,” says McGregor.

This shift marks a move away from processor-centric thinking toward a full-stack approach. Olsen likens it to purchasing a car or a smartphone, where instead of having customers assemble individual components themselves, consumers buy an integrated product designed to work as a whole.

Applying the same logic to data centers, a GPU alone is only one part of the equation. Without tightly coupled memory, networking, software, and applications, it’s “just a piece of electronics,” as Olsen puts it.

Vendors like Nvidia are addressing this by designing and developing tightly integrated hardware and software platforms deployed as complete systems. Of note is the company’s NVLink, which intelligently enables up to 72 GPUs to operate as a single processor.

This level of interconnection turns networking into a primary constraint, with AI factories depending on high-bandwidth, ultra-low-latency connections within racks and across entire sites. Olsen notes that this requirement places strict physical limits on scale:

“In a gigawatt-scale AI factory, you’re effectively limited to a network radius of around 100 meters – that’s about as far as the system can run optimally. That constraint is what’s driving the tight clustering of compute, power, and cooling infrastructure, including the large chiller systems we now call ‘sprawl’.”

Networking is forcing everything closer together and driving rack densities well beyond 240kW, toward 600kW and even a megawatt per rack. We’re currently designing for a customer at just over a megawatt per rack, which is very significant.”

The crux of Olsen’s point is that you can’t compartmentalize the power, cooling, white space, and software management pieces. They must be engineered as a single, end-to-end system, spanning everything from grid connection to silicon to heat rejection and reuse.

This is crucial because while hyperscalers like Google and AWS may have the resources to integrate these layers internally, most enterprises do not. But smart solutions such as pre-integrated architecture can help mitigate this constraint.

Vertiv OneCore, for example, is a fully prefabricated AI factory architecture scalable up to a gigawatt, aligned with Nvidia’s DGX and DSX AI factory designs. According to Olsen, the goal is to deliver a cohesive, full-stack physical infrastructure that minimizes risk, avoids costly surprises, and keeps the data center operating as a unified system.

Beyond single facility infrastructure

Power

As AI systems grow in scale and complexity, the full-stack conversation extends well beyond facility infrastructure, edging into orchestration, software, and operational intelligence. As McGregor observes, the industry has reached a point where AI will increasingly be required to manage AI itself.

When it comes to scaling these systems, however, the most immediate constraint is the ability to transmit and distribute power to the right place at the right time. And addressing this challenge is prompting renewed interest in alternative energy sources, including small modular reactors (SMRs). As Olsen explains:

“There are some new players coming in – Oklo, for instance – as well as SMRs more broadly. Nuclear is a way of scaling power, and while there’s a long runway due to a lack of full understanding, it’s actually a very mature technology. From a commercial and technical readiness perspective, it’s far further along than many people realize.”

McGregor reinforces the point by stressing that resilient power strategies have always depended on diversification:

“Nobody is planning a data center around a single power source. Over the past two years, it’s become clear that relying solely on the grid brings real challenges and trade-offs. That’s why layered approaches – using batteries, supercapacitors, on-site generation, and potentially SMRs – are gaining momentum and hold significant potential.”

In many ways, he argues, the industry is returning to an earlier model in which the grid was never intended to be the primary source of power but a backup. Regulatory approvals may take time, but the potential is significant.

Labor

Labor is emerging as another critical constraint, as the sheer number of people required to keep gigawatt-scale AI data centers operational is staggering. Olsen points to estimates from Crusoe co-founder Chase Lochmiller, suggesting that as many as 7,000 people may be working on-site at every single facility. These personnel must have a highly specialized skillset, particularly in electrical and mechanical disciplines.

From the context of scalability, deploying such large teams across multiple geographies, climates, and regulatory environments adds another layer of logistical complexity.

This is where Olsen points to prefabrication as a strategic enabler:

“The prefabricated model allows you to move much faster because you can apply a standardized process,” he says, indicating the potential to reduce both labor intensity and deployment timelines.

Together, these developments underscore that scaling AI is no longer a question of infrastructure alone. It is a full systems challenge, requiring an integrated, full-stack approach to keep pace with the next phase of AI growth.

Watch the full Trends and Outlooks broadcast here.

To learn more about Vertiv OneCore, visit vertiv.com.