As Artificial Intelligence and high-performance computing drive demand for higher-density infrastructure, data centre operators must rethink traditional cooling strategies to manage increasingly concentrated heat loads while maintaining efficiency and operational resilience. Klaus Dafinger, Cooling Marketing Manager at Legrand, tells us why targeted rack-level cooling and hybrid cooling architectures are becoming essential to supporting AI workloads, improving energy efficiency and enabling data centres to expand capacity without extensive infrastructure redesign.
Why localised heat is reshaping data centre cooling
Artificial Intelligence (AI) workloads are exposing a fundamental weakness in traditional data center cooling: heat is no longer evenly distributed across the room. Historically, cooling strategies were designed around relatively consistent rack densities, where optimising room-level airflow delivered predictable thermal performance. Today’s AI deployments look very different.
Operators are increasingly managing mixed environments in which legacy enterprise applications sit alongside GPU clusters ranging from 80–150 kW per rack. These high-density deployments create highly localised thermal hotspots that challenge conventional room-based cooling models and are forcing operators to rethink where cooling is applied.
As a result, the cooling conversation is moving beyond how to circulate more air through the data hall and towards how to remove heat precisely where it is created.
The limits of room-level cooling
Fan wall architectures have become an increasingly popular approach for improving room-level airflow management. By using multiple smaller fans arranged within a dedicated wall, they provide more uniform airflow, improved pressure control, built-in redundancy and greater energy efficiency than many traditional cooling approaches.
These systems are particularly effective in environments where thermal loads are relatively consistent across the white space. However, modern AI deployments are changing those assumptions.
Unlike traditional enterprise environments, AI infrastructure rarely produces evenly distributed heat. High-density GPU clusters are often concentrated within specific areas of the data hall, creating localised thermal hotspots that may not be efficiently addressed by room-level airflow optimisation alone.
Fan wall architectures remain an excellent solution for many facilities, but their effectiveness is constrained when cooling demand is concentrated in specific racks while the wider room operates at lower densities.
Common considerations include:
- Large infrastructure footprint requiring dedicated wall space.
- Performance is optimised when thermal loads are relatively uniform.
- Limited ability to target airflow at individual high-density racks.
- Successful deployment depends on detailed airflow modeling and room design.
- Typically requires chilled water infrastructure and associated control systems.
AI is changing the cooling challenge
The rapid growth of AI and high-performance computing is shifting cooling design from a room-temperature problem to a heat-source problem.
Rather than distributing similar workloads throughout the facility, operators are increasingly deploying clusters of high-density AI servers alongside conventional enterprise infrastructure. These mixed-density environments place very different demands on cooling systems.
Cooling strategies are therefore being driven less by airflow alone and increasingly by the characteristics of the IT load itself.
This is why operators are increasingly looking for cooling architectures that can scale selectively, targeting the racks and rows where thermal risk is highest without forcing unnecessary change across the entire facility.
Why rack-level cooling is becoming more relevant
This shift is making rack-level technologies more relevant, particularly for operators that need to support AI growth without committing immediately to a full-scale cooling redesign.
Rear Door Heat Exchangers (RDHx) capture heat directly from the rack exhaust using a liquid-cooled heat exchanger integrated into the cabinet’s rear door. As warm exhaust air leaves the servers, thermal energy is transferred into the liquid circuit before the conditioned air is returned to the data hall.
Rather than relying solely on room-level airflow, this approach removes heat much closer to its source and gives operators a practical bridge between conventional air cooling and more intensive liquid cooling architectures.
For data centres navigating mixed densities, this enables operators to:
- Increase rack density where required, without upgrading cooling capacity across the entire data hall.
- Align cooling capacity with actual heat generation to improve efficiency by targeting only the highest-density racks.
- Deploy cooling incrementally, allowing investment to follow business growth, AI adoption and server refresh cycles.
- Minimise infrastructure disruption, as RDHx can often be integrated within existing rack footprints without major structural modifications.
- Support mixed-density environments, cooling AI racks independently while allowing conventional workloads to continue using existing air-cooled infrastructure.
- Improve operational resilience by localising cooling and reducing the impact of maintenance or faults compared with more centralised cooling architectures.
Cooling strategies are becoming hybrid
Perhaps the biggest change occurring across the industry is that cooling is no longer viewed as a choice between one technology and another.
Instead, many operators are adopting hybrid cooling strategies that combine room-level airflow management, rack-level cooling, and direct liquid cooling according to workload requirements.
Traditional enterprise applications may continue to be cooled efficiently using established air-cooling methods. High-density AI deployments may justify the use of Direct-to-Chip (D2C) liquid cooling. Between these extremes lies a growing number of mixed-density environments where rack-level cooling technologies such as RDHx can increase capacity without requiring wholesale infrastructure redesign.
Rather than replacing existing cooling systems, hybrid architectures enable operators to modernise progressively, balancing thermal performance, operational risk and capital investment while preparing facilities for future AI growth.
Conclusion
The future of data centre cooling will not be defined by a single technology, but by how precisely operators can match cooling capacity to workload demand.
Instead, successful operators will increasingly combine room-level airflow optimisation, rack-level cooling, and direct liquid cooling to meet the specific requirements of each workload.
As AI infrastructure continues to evolve, the competitive advantage will come from cooling strategies that can address concentrated heat without overhauling the entire data hall.
For many modern data centres, targeted, workload-driven cooling is becoming a defining characteristic of AI-ready infrastructure and a practical route to scaling capacity with control.
To learn more about how RDHx technologies can support high-density AI environments and phased liquid cooling adoption, visit the Legrand website.
