AI Requirements Driving Liquid Cooling Design

Author: Brandon Peterson, SVP Business Development, CoolIT Systems Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

CoolIT Systems presents the technical requirements driving liquid cooling design evolution, from cold plate thermal management to CDU scaling. The AI data center capacity market is projected at 65% CAGR from 2023-2028, with liquid cooling growing at 45% CAGR. By 2028, one-third of all data center capacity will be liquid cooled. CoolIT provides end-to-end liquid cooling from chip (cold plates) to CDU.

The presentation makes a key distinction: it is not rack density or even chip TDP that drives liquid cooling requirements, but rather the hotspots on the chip. Lower-density racks with fewer servers may still require liquid cooling because the chip demands it. Even some peripheral components around the chip edge are driving challenging thermal requirements. Flow rate per kilowatt is transitioning from 1.0 LPM/kW to 1.5 LPM/kW, with significant implications for quick disconnect sizing (transitioning from 1-inch to 2-inch QDs, with larger sizes needed for megawatt racks).

The residual air cooling challenge is highlighted: even capturing 90% of heat in a megawatt rack still leaves 100+ kW of heat to dissipate via air, driving interest in full-coverage cold plates for lower-density board components. CDU evolution maps from rack-based (~100 kW per 2U, HPC era 2015-2020) to row-based (1-2 MW, AI deployments 2024-2026) to facility CDUs (multi-MW, AI deployments 2026+) to building-level CDUs (future, for GW facilities). The expected progression is toward 3-6 MW facility CDUs in 2026, then 10 MW+ beyond.