Liquid Cooling for AI Data Centers: Direct Liquid Cooling vs Rear-Door Heat Exchangers

Modern AI accelerators have made the thermal problem unavoidable. A single NVIDIA H100 SXM5 dissipates 700W; the Blackwell B200 pushes past 1,000W. Air cooling, which sustained the data center industry for decades, cannot remove that much heat from a standard rack footprint at acceptable cost or reliability. Liquid cooling is now the default design assumption for any serious AI infrastructure build, and the choice between direct liquid cooling and rear-door heat exchangers carries consequences that ripple through facility design, operations, and long-term flexibility.

Direct liquid cooling: how it works and where it excels

Direct liquid cooling (DLC), also called direct-to-chip cooling, routes chilled water or a dielectric fluid directly to cold plates mounted on CPUs, GPUs, and memory modules. Heat is transferred into the fluid at the source, eliminating the air-to-liquid conversion step entirely. The result is dramatically lower PUE — DLC systems routinely achieve PUE below 1.15, compared to 1.4–1.6 for conventional air-cooled facilities. The tradeoff is complexity: DLC requires wet infrastructure inside the server rack, specialized server models with integrated cold plates, and facility plumbing that reaches every rack. Leak detection, fluid chemistry management, and connector reliability become operational disciplines.

Rear-door heat exchangers: the retrofit path

Rear-door heat exchangers (RDHx) mount at the back of a standard server rack and use a liquid-cooled coil to absorb heat from server exhaust air before it enters the hot aisle. Standard servers ship unchanged — no cold plates, no modified plumbing inside the chassis. This makes RDHx the practical choice for enterprises upgrading existing facilities that were never designed for in-rack liquid cooling. A well-designed RDHx system can handle 30–50 kW per rack effectively, which covers current-generation single-GPU servers. The limitation becomes apparent with the densest AI configurations: at 80–120 kW per rack, residual heat that bypasses the door coil still requires supplemental cooling, and the air-to-liquid transfer efficiency is inherently lower than DLC.

• DLC: best for new builds or major renovations targeting 60 kW+ per rack density
• RDHx: best for retrofitting existing data centers hosting current-generation GPU servers
• DLC requires server models with factory-integrated cold plates — not all GPU servers support it
• RDHx works with any standard server, preserving hardware flexibility
• Hybrid approaches are common: DLC for dense GPU rows, RDHx for compute and storage rows
• Coolant temperature matters — warmer coolant (40–45°C) enables free cooling in temperate climates

The facility infrastructure implications

Both technologies require a facility-side cooling distribution unit (CDU) that connects to building chilled water or a dedicated dry cooler loop. Sizing the CDU correctly — with margin for future expansion — is one of the most consequential decisions in the design phase. Under-sizing creates a hard ceiling on deployable GPU density. Over-sizing increases upfront capital cost without proportional benefit. The best designs include modular CDU capacity that can scale in increments aligned with rack deployments, avoiding large speculative investments in cooling infrastructure before the workload demands it.

The facilities conversation has to happen before the server conversation. The biggest deployment delays we see come from teams that finalized hardware specs before they understood what their building could support thermally.

How Nexus Compute helps

Nexus Compute supplies GPU servers pre-configured for both DLC and RDHx deployment, working with leading CDU vendors to provide integrated cooling solutions. Our engineering team performs facility thermal assessments and produces cooling architecture recommendations before purchase commitments are made. Whether you are retrofitting an existing data center or designing a new AI-native facility, we can model the full thermal envelope and identify the lowest-TCO cooling path for your specific power and density targets.