The pace of Artificial Intelligence (AI) innovation is forcing major change inside the world’s data centres. From high-density hardware to hybrid cooling strategies and real-time telemetry, today’s facilities must be both resilient and agile – while maintaining security at every layer.
We spoke with Dominik Dziarczykowski, Market Development Manager, High Density & Liquid Cooling, at Vertiv, about what it takes to build infrastructure that’s ready for AI now and in the future and why operational intelligence depends on integrated thinking.
How is AI changing the physical demands placed on data centre infrastructure?
AI is creating concentrated, high-power computing environments that differ significantly from traditional enterprise or cloud loads. Where five years ago a rack might average 8kW–10kW, today we’re seeing AI racks drawing 120kW or more. And many operators are preparing for 150kW and upwards in AI-specific zones, with forecasts indicating that this may increase to 300-600kW and possibly 1MW by 2030.
This level of density doesn’t just affect cooling, it changes the structural load on raised floors, increases the importance of cable routing and requires new approaches to airflow and power delivery.
The impact is physical, electrical and procedural. Facilities are having to plan for increased heat rejection, faster provisioning and tighter fault tolerance. These changes are exposing bottlenecks in legacy designs and forcing operators to re-evaluate how infrastructure is deployed and maintained.
What are the biggest cooling challenges AI infrastructure creates?
Established air cooling remains effective for a wide range of equipment, but it starts to hit its limits in high-density environments. AI accelerators, graphics processing units (GPUs) and clustered compute nodes tend to concentrate heat in very small footprints and that requires precision cooling.
Liquid cooling, whether direct-to-chip or rear-door heat exchanger, is becoming a necessary tool in the thermal arsenal. But the move to liquid is not binary. Most data centres are evolving towards hybrid environments where air and liquid systems operate together. That introduces complexity. There are new fluid networks to manage, different maintenance routines and greater reliance on monitoring to maintain balance and safety.
One of the key challenges is designing for flexibility. AI environments change fast and cooling needs to support that, not constrain it.
With such rapid change, how are operators adapting build and deployment models?
Speed is becoming as important as capacity. AI cycles move quickly, but traditional data centre builds do not. That’s why there’s a real push toward prefabricated, modular and tightly integrated systems. These approaches allow operators to get capacity online faster, with fewer dependencies between trades and fewer variables on-site.
We’re also seeing more integrated planning. Cooling, power, monitoring and enclosure design are being specified in parallel, not sequentially. That’s important because these systems need to work together from day one. The days of bolting on a cooling unit as an afterthought are over. Everything from pipework to thermal zoning is part of the design conversation earlier than ever.
What role is physical security playing as density and value per square metre increases?
AI systems often carry not only sensitive data but also valuable intellectual property. The compute infrastructure itself – accelerators, custom chips, networking – represents a substantial investment. As a result, physical security is becoming more layered and more dynamic.
It’s no longer just about locked cages or perimeter access. Operators are introducing compartmentalised access, role-based controls and multi-factor authentication for physical entry. Smart cabinet locks, surveillance with AI-enabled monitoring and even real-time asset tracking are becoming standard in some environments.
In multi-tenant and colocation facilities, these challenges are even more acute. Tenants want assurance that their infrastructure is physically isolated and actively monitored. For operators, this means investing in systems that integrate physical and logical access logs and that can flag unusual patterns across both.
Physical security also has to scale with automation. In lights-out or low-touch environments, systems need to detect, alert and sometimes respond autonomously – whether that’s denying access, logging environmental anomalies or triggering remote inspection protocols.
How is the power landscape shifting in AI-scale deployments?
We’re seeing several important shifts. First, the rise in density means higher draw per rack, which puts more strain on power distribution systems. Second, the nature of AI workloads creates spikier demand curves. This variability must be matched by more responsive power infrastructure.
At the same time, we’re seeing a redefinition of the role of backup. Many operators are starting to view their uninterruptible power supply (UPS) systems not simply as emergency failovers, but as assets that can contribute to energy strategy. That includes grid support, peak shaving and participating in demand response programmes. These models are already live in parts of Europe and interest is growing in Asia and North America as well.
Integration is key here. Power systems, cooling and monitoring platforms need to work in concert. Without that alignment, operators risk oversizing, underutilising or mismanaging their infrastructure.
What are the blind spots emerging in next-generation infrastructure?
One of the most overlooked areas is the impact of dual cooling systems on operations and maintenance. Introducing liquid loops creates new dependencies, but not all facilities are updating their maintenance protocols to match. Issues like fluid leakage, flow rate fluctuations or partial blockages are manageable, but only if the right monitoring and response systems are in place.
Another blind spot is around interoperability. When equipment is deployed piecemeal, especially under time pressure, there’s a risk of mismatch between systems. That includes data formats between building management system (BMS) platforms, compatibility of cooling controls or even inconsistent commissioning standards. Intelligent data centres need shared visibility to maintain operational consistency.
Are customers treating sustainability as part of the infrastructure brief, or a separate concern?
Sustainability practices are now part of the core planning brief for most major operators. That includes not only power usage effectiveness (PUE) and water usage effectiveness (WUE), but also material and energy sourcing, lifecycle impact, waste heat reuse and end-of-life recycling plans. In some regions, this is driven by regulation, in others, by competitive or investor pressure.
The challenge is to balance speed, performance and footprint. AI infrastructure needs to be energy-intensive at the point of compute, but the systems around it must work efficiently. That’s where precision cooling, smart power usage and automation can make a real difference. Efficiency isn’t just an environmental concern – it is a cost, risk and performance consideration too.
What’s next for intelligent infrastructure?
The next phase is orchestration. We’re moving into an era where power, cooling, security and compute need to be managed dynamically, based on live workload data. The systems that succeed will be the ones that respond to change – whether that’s a new rack coming online, a surge in compute activity or a cooling unit operating outside expected parameters.
This means investing in better telemetry, smarter controls and platforms that bring disparate systems into a unified operational view. AI is accelerating this requirement.
