Liquid Cooling for AI Data Centers

Overview

Liquid cooling for AI data centers refers to direct liquid cooling (DLC) architectures that remove heat from GPU racks via coolant distribution units (CDUs) and cold plates rather than forced air. The primary economic driver is power efficiency: air-cooled data centers spend approximately 37% of total facility power on HVAC/air conditioning (per Schneider Electric), while liquid-cooled facilities reduce that overhead to approximately 2% — a 35-percentage-point differential that translates directly into additional GPU compute capacity in power-constrained facilities.

GPU power density growth from ~300 W to 1,000 W+ per GPU has outpaced the ability of legacy air-cooling solutions — cooling doors, power shelves, cooling towers — to keep pace with data-center thermal loads, making liquid cooling mandatory for the highest-density configurations. NVIDIA's GB200/300 NVL72 racks run at ~140 kW and require four data-hall busways, a configuration only available in purpose-built facilities; Supermicro's air-cooled B300 SuperCluster (576 GPUs, 20 racks) targets the majority of existing facilities with standard two-busway power.

In a 1 gigawatt data center, choosing liquid over air cooling for a B300 cluster yields approximately 100,000 additional B200-equivalent GPUs from the reclaimed cooling power — valued at roughly $1.95 billion in GPU rental terms based on Microsoft's cited offer. Supermicro's liquid-cooled CDUs use dual-redundant pumps with redundant power supplies, pressure and temperature telemetry via IPMI, and automated response via Super Cloud Composer software. The 4U Liquid-Cooled B300 SuperCluster packs 576 B300 GPUs into 11 racks (vs 20 for the air-cooled equivalent) using nine 250 kW CDUs.