Outdated data centers not ready for AI requirements

When it comes to AI, many organizations think of chatbots and language applications, but Boutsen sees the rise of agentic AI as a significant development for the IT infrastructure sector. AI agents are systems that perform tasks independently, process information, and collaborate with each other. This places major pressure on the underlying digital infrastructure, warns Boutsen.

“Scalable and affordable support for hundreds of thousands of concurrent AI users and AI agents is a major challenge for modern data centers and server rooms striving for a fast Time to First Token (TTFT). It requires large, logically coupled GPU clusters in which processors (CPU & GPU) continuously exchange data with each other, with the lowest possible latency.”

Power density explodes

The consequences of the rapidly growing demand for computing power are immediately noticeable on the shop floor. Traditional enterprise racks often have a power density of a few kilowatts to approximately 20 kW, depending on the configuration and utilization rate. AI platforms require much higher power densities. “Some AI racks can already reach power levels of 100 to 400 kW today. This is ten to several tens of times higher than with traditional enterprise racks,” says Boutsen.

The latest generation of AI infrastructure makes the rapid developments visible. “For example, NVIDIA brings together 72 GPUs in a single NVLink domain in its liquid-cooled NVL72 platform. The latest generation, the Vera Rubin platform, has been ready to enter mass production since the summer and already sets the limit at 250kW per IT rack.”

Boutsen continues: “HPE started developing a fanless liquid cooling architecture capable of dissipating up to 400 kW of heat per rack as early as 2018, which resulted in our Cray EX4000 series. Today, HPE Labs works closely with all industry stakeholders to achieve the step to 600 kW and ultimately 1 MW per IT cabinet.”

Liquid cooling becoming increasingly important

But the question is not only how much computing power is needed, but above all how organizations can accommodate that capacity sustainably, affordably, and reliably.

One of the major challenges in the AI era is cooling equipment and spaces. Boutsen explains: “Traditional air cooling is no longer sufficient at the highest power densities. For this reason, more and more companies are switching to direct liquid cooling (liquid cooling) in large-scale AI and HPC environments. By dissipating heat directly at the processor, you can cool rooms more efficiently and require less energy-intensive cooling systems.”

Boutsen expects water-glycol-based liquid cooling to play an increasingly important role in AI platforms for data centers. “With this form of cooling, it is possible to work with higher water temperatures than with traditional systems. This reduces the need for mechanical cooling and creates new opportunities for heat reuse.”

He continues: “I advise companies to design data centers and IT rooms as part of a broader energy system. In this system, for example, waste heat can be utilized for district heating networks, greenhouses, or industrial processes. This is both efficient and sustainable.”

Grid capacity becomes a strategic factor

In addition to power density and cooling, AI creates a third challenge for the sector: energy management. Waiting times for new or expanded high-volume consumer connections can increase to several years due to grid congestion.

Boutsen: “If companies do not anticipate this, future AI initiatives will be thwarted by a shortage of utilities, not a shortage of IT resources.” The data center specialist advises organizations not to wait until AI projects are fully rolled out. “As soon as AI plays a business-critical role, the infrastructure must be ready. Therefore, I call on companies to map out their future IT landscape now, including capacity requirements and scalability.”

Sometimes the solution is literally around the corner. If you place the equipment close to energy production sites, you do not first have to transport the required power over an overloaded electricity grid. “However, fiber optic connections, connectivity, permits, and physical space remain indispensable preconditions,” says Boutsen.

Why modular data centers are gaining ground

An additional problem is that the development of AI is progressing so rapidly that the construction cycle of a data center cannot keep up. AI hardware is renewed every few years, but new data center projects have a lead time of several years. In practice, this sometimes means that a data center is already outdated upon delivery.

According to Boutsen, the solution to this dilemma lies in modular construction. “By building IT capacity, power supply, and cooling in a modular way, more flexibility is created to support future AI workloads. The moment a new generation of hardware becomes available, you do not have to redesign the facility. You can simply replace the individual IT modules and expand the cooling and power supply if necessary.”

Sustainability broader than PUE

Meanwhile, sustainability is also still high on the political agenda.

The European Commission is working on a common assessment system for data centers that provides insight into, among other things, energy use, water consumption, waste heat, and the contribution to the electricity grid. The first sustainability labels are from 2027 expected.

Don't wait until the question presents itself

Boutsen succinctly summarizes the common thread of his keynote. “AI is not just a new workload which you simply add to the existing infrastructure. It fundamentally changes the way data centers are designed, cooled, powered, and made sustainable.”

The IT Infra event takes place on Wednesday, November 25, 2026, at Congrescentrum 1931 in 's-Hertogenbosch. Register free of charge. via the website.