Home Blog GreenTech and Sustainable IT Liquid cooling for AI data centers: direct liquid cooli...
🌱 GreenTech and Sustainable IT March 17, 2026 12 min read

Liquid cooling for AI data centers: direct liquid cooling guide

GreenTech and Sustainable IT Enterprise Guide 2026 SCALE D2C D2C Technology GreenTech and Sustainable IT Enterprise Guide 2026

Liquid cooling for AI data centres has moved from "nice to have" to "technically required" for dense GPU clusters. A single NVIDIA DGX H100 system dissipates 10.2kW — far exceeding what air cooling can remove from a standard data centre rack. Direct Liquid Cooling (DLC) with cold plates on CPUs and GPUs is now the mandated cooling solution for H100 and H200 deployments, while full immersion cooling is emerging for maximum density. This guide covers DLC architecture, vendor options, deployment requirements, and the operational changes liquid cooling demands from enterprise data centre teams.

Why Air Cooling Fails for AI GPU Clusters

AI GPU Heat Density vs Air Cooling Limits
Air cooling in data centres moves heat using forced airflow — CRAC/CRAH units cool server inlet air; hot exhaust exits the rear. Maximum practical air-cooled rack density: 15–20kW per rack with hot aisle containment. NVIDIA H100 SXM5: 700W TDP per GPU; 8 GPUs per DGX H100 = 5.6kW GPU power alone; full DGX H100 system (including networking): 10.2kW. A rack of DGX H100 units exceeds 40kW — requiring DLC regardless of data centre cooling capacity. Attempting to air-cool H100 clusters produces thermal throttling, reliability failures, and voided warranties.

DLC Implementation Approaches

ApproachWhat Gets Liquid-CooledResidual Air RequiredPUE AchievableRetrofit Difficulty
Cold Plate DLC (CPU/GPU only)CPU and GPU die — highest heat generatorsYes — for VRMs, memory, other components1.15–1.25Medium — requires liquid distribution units (CDU)
Full Server DLCCPU, GPU, VRM, memory, NVMe via liquid railsMinimal — small residual air for storage1.05–1.15High — custom server design required
Rear Door Heat ExchangerAll server heat captured at rack exhaustYes — servers still air-cooled internally1.20–1.35Low — attaches to existing rack rear
Single-Phase ImmersionEntire server submerged in dielectric fluidNone1.03–1.05Very high — bespoke tank infrastructure
40kW
Minimum rack density for a single DGX H100 rack — requiring DLC regardless of data centre air cooling capacity. NVIDIA mandates DLC for DGX H100/H200 deployments and will void warranty without it
25%
Higher GPU boost clock frequencies achievable with DLC vs air cooling — lower die temperatures enable sustained performance beyond thermal throttling limits, recovering GPU performance in addition to reducing PUE
50°C
Maximum coolant supply temperature for DLC systems on most NVIDIA GPU designs — enabling "free cooling" (no chiller required) when ambient water temperature is below this threshold, dramatically reducing cooling energy in temperate climates
01
DLC Infrastructure
Coolant Distribution Unit (CDU) Design

The CDU is the heart of a DLC system: pumps coolant to server cold plates, controls flow rate and pressure, monitors supply/return temperatures and flow. Size CDU for the rack's maximum thermal load plus 20% headroom. For DGX H100 deployments: 60kW CDU minimum per rack (10.2kW DGX × 4 racks, plus headroom). Water quality management (deionised water or glycol mixture) is critical — poor water quality causes galvanic corrosion in cold plates. Engage your facilities team and the CDU vendor on water chemistry specifications before procurement.

CDU sizingWater chemistry20% headroom
02
Infrastructure
Facility Water Loop Integration

CDUs connect to the facility chilled water loop or dry cooler loop. Design for free-cooling operation: if your chilled water loop operates at 40–50°C supply temperature, no chiller is required for DLC — major energy saving. Work with your mechanical engineer on: secondary loop isolation (prevent data centre water from entering building HVAC water), leak detection under raised floor or at every CDU, and emergency drain-down procedures. Facility water loop modifications typically require 3–6 months lead time — plan ahead of GPU procurement.

Free-cooling thresholdLoop isolationLeak detection
AI Data Centre Liquid Cooling

Our DevOps and digital transformation teams advise on AI GPU infrastructure including DLC requirements, CDU specification, and data centre readiness assessment for H100/H200 deployments. Book a free advisory session.

Frequently Asked Questions

End-to-end GreenTech and Sustainable IT strategy, implementation, and optimisation for enterprise and D2C brands. Contact us for a free consultation.

Strategy projects: 4–8 weeks. Full implementation: 3–12 months. ROI typically within 12–18 months.

Yes — D2C brands to enterprise. View our pricing.

GREENTECH AN

Ready to Implement GreenTech and Sustainable IT?

Our specialist team delivers measurable ROI for enterprise and D2C brands.

Free Audit