Enabling Liquid Cooling on a Global Scale

Author: Andy Regimal, Director of Infrastructure Architecture, Azure Engineering Group, Microsoft Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Microsoft presents its strategy for deploying liquid cooling globally across Azure's 70+ regions by developing a next-generation Heat Exchange Unit (HXU) that can be deployed in existing air-cooled data centers without facility modifications. GPU TDPs have jumped from 400W to 1,200-1,300W today, with 2,300-2,500W parts expected soon, and devices above ~700W require liquid cooling. Current rack elevations exceed 100-130 kW heading toward 200+ kW, with 80% of rack components today requiring liquid cooling, heading to 100%.

The fundamental challenge is that facility water data centers represent only a tiny fraction of global hyperscale capacity. A world map shows North America at 90% air / 10% liquid, Europe at 95% air / 5% liquid, Asia-Pacific at 92% air / 8% liquid, and Latin America & MEA at 98% air / 2% liquid. The next-generation HXU targets 10x the cooling capacity of HXUs from four years ago and 2x the capacity of the version shared at OCP 2024, in the same footprint with the same airflow per kilowatt and approach temperature.

The HXU design prioritizes four pillars: performance (latest fan, pump, heat-exchanger technology), scalability (global deployment in air-cooled data centers using modular, multi-source components), rapid deployment (no construction or facility modification needed, deployable in IT rows by existing data center personnel), and availability at scale (five-nines/99.999% cooling uptime via component redundancy, hot-swappable assemblies, and real-time telemetry). AI cluster metrics require 95%+ available nodes, 90%+ cluster availability, and 2-4 hour MTBF. Microsoft plans to donate the NextGen HXU specification to OCP in late 2026.