Liquid Cooling for Networking

Overview

Liquid cooling for networking equipment has crossed the threshold from optional to mandatory. Tomahawk 6-class switch ASICs dissipate nearly 2kW per chiplet — approaching GPU-level power — making traditional air cooling insufficient for the thermal load. The industry entered its second generation of liquid cooling for networking around 2025 and is transitioning to a third generation based on OV3 rack standards in 2026.

Google deployed liquid cooling in custom network racks approximately five years before OCP 2025, making it an early pioneer. However, the broader industry is now at the inflection point, with Celestica equipping all factories for OV3-standard liquid cooling in 2026. This liquid cooling requirement is also driving engineering complexity: Tomahawk 6 requires 50-layer exotic PCBs (with only a few worldwide suppliers capable), demanding supply chain qualification at manufacturing scale.

The killer application driving liquid cooling adoption in networking is the same GPU-scale power density that already forced liquid cooling into AI compute racks — once networking ASICs cross the ~1kW threshold, conventional air cooling can no longer maintain safe operating temperatures at rack density.