The Future of Single-Phase Direct Liquid Cooling

Author: Kamal Mostafavi, CTO, CoolIT Systems Collection: 2026-OCP-EMEA-Summit-Executive-Sessions

Summary

Kamal Mostafavi, CTO of CoolIT Systems (celebrating 25 years), presented a comprehensive data-driven forecast for liquid cooling adoption in AI data centers, arguing that single-phase Direct Liquid Cooling (DLC) will remain the mainstream thermal management solution for the foreseeable future, even as rack densities scale toward 250 kW and pod-level deployments reach 9–10 MW. The talk draws on CoolIT's own market models (validated with SemiAnalysis) to present a layered thermal management picture spanning processor, server, rack, and pod.

The thermal trend data is stark: processor TDP is on an exponential growth curve, projected to reach approximately 7,000W by 2029 (from ~2,400W in the current market). Heat flux is between 100–200 W/cm² today, rising but not at the same rate as TDP. Peripheral component cooling (SFPs, CX9, DIMMs, SSDs, power supplies) is expanding the scope of liquid cooling loops from the current 50–70% coverage toward 100% heat capture — a requirement for racks above 250 kW. CoolIT's own cold plate portfolio now supports up to 4 kW per cold plate, enabling up to 15 kW of cooling per server loop, 500 kW+ per rack, and 30 MW+ grouped CDUs at pod level.

The CDU market dynamics show an impending shift: liquid-to-air CDUs currently dominate installed capacity but are being displaced by liquid-to-liquid CDUs for new data centers. Critically, liquid-to-air CDUs represent only 20% of GW installed but 45% of CDU TAM by dollar value because they cost 4x more to deploy than liquid-to-liquid. CoolIT argues single-phase DLC wins across all cooling hierarchy levels (die, chip, server, rack, pod) vs. both two-phase DLC and immersion, primarily because of: effective heat capture at all levels from ultra-high heat flux to long-distance transport; a multi-vendor interoperable ecosystem; and simple, scalable deployment already proven at GW scale.