Liquid Cooling for High-Density AI Compute

Overview

As AI workload rack densities approach and exceed 1 MW per rack, air cooling becomes physically infeasible. The thermal load from GPU compute at these densities cannot be removed by forced-air convection alone, requiring direct liquid cooling via CDUs (Coolant Distribution Units) and cold plates as the only viable option.

The architectural implications are significant. In the 2027 consolidated data center design presented by Flex at OCP Global Summit 2025, the entire data hall is liquid-cooled and fed by a 4,000 A double-mod busway at up to 800 V DC — shrinking the used floor footprint from 160 sq ft (48 V air-cooled) to 32 sq ft, an 80% reduction. The liquid cooling infrastructure (CDUs scaled to 600 kW, 1,200 kW, or 1,800 kW modules) integrates directly with the power delivery architecture, requiring co-design rather than sequential handoffs between power, cooling, and IT teams.