As AI training and inference workloads push rack power densities beyond the practical limits of air cooling, liquid cooling technologies have moved from a specialist niche to a mainstream design requirement for AI-ready data center facilities.
Why Air Cooling Runs Out of Headroom
Air has fundamental thermodynamic limitations as a cooling medium. Beyond roughly 30–40kW per rack, the airflow volumes and velocities required to remove heat efficiently become impractical — driving noise, energy consumption, and physical space requirements that undermine the facility's overall efficiency.
The Liquid Cooling Spectrum
Liquid cooling is not a single technology but a spectrum of approaches, each suited to different density and retrofit scenarios:
- Rear-door heat exchangers — capture heat at the rack exhaust, offering a relatively low-disruption upgrade path for existing facilities
- Direct liquid cooling — delivers coolant directly to cold plates on compute components, supporting the highest sustained densities
- Immersion cooling — submerges hardware directly in dielectric fluid, offering exceptional density and efficiency at the cost of greater operational change
The choice between liquid cooling technologies is rarely about which is "best" in absolute terms — it is about matching technology to density requirements, retrofit constraints, and operational maturity.
Design Implications Beyond the Rack
Adopting liquid cooling reshapes decisions well beyond the rack itself — influencing facility water strategy, mechanical plant sizing, structural floor loading, and even building services routing. Facilities designed without liquid cooling pathways from the outset face significant retrofit costs and operational disruption later.
Designing for an Uncertain Future
Given the pace of change in AI hardware, the most resilient facility designs build in flexibility — supporting a transition from air to hybrid to fully liquid-cooled configurations as workload requirements evolve, without requiring disruptive structural intervention at each stage.
