From Fans to Fluids: The Critical Need For Liquid Cooling in AI Data Centers

Author: Bobby Kintol, Hive Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Bobby Kintol, a self-described "liquid cooling super fan" with experience dating to 2006 (half-megawatt containerized data centers), presents the case for liquid cooling as the standard for AI data centers. The presentation traces the speaker's career from early containerized liquid cooling (2006) through consumer PC liquid cooling to the current AI-driven data center adoption. Rack power has jumped 3-10x almost overnight, from 30 kW air-cooled racks to 120 kW common today, with 300 kW expected in 2026 and megawatt-plus racks projected by 2030.

The presentation makes several practical engineering arguments for liquid cooling beyond thermal management. Design benefits: liquid cooling enables packing 40 GPUs where air cooling fits 8, and returns triple-slot PCIe GPUs to single-slot form factors. Performance benefits: liquid cooling typically runs dies at ~65C versus 85-90C with air, yielding 4% power reduction per 10C decrease and enabling sustained maximum turbo frequency (versus bouncing in and out with air cooling), with some CPUs offering 50% turbo overhead. Packaging innovation: removing the front-to-back airflow constraint allows placing temperature-sensitive components (HDDs, memory, optical transceivers) behind CPUs, enabling entirely new server architectures analogous to how LED technology freed lighting from the Edison bulb form factor.

Industry growth projections show liquid cooling increasing by 7x in 2026 versus 2025 (from 5% to 38% of shipments), with ~21% CAGR through 2030. The speaker emphasizes that liquid cooling is not exotic technology -- the PC gaming industry has shipped over 1 million CPU liquid coolers annually since 2013 (2 million+ in the 2020s), and the electric vehicle industry (Tesla octa-valve example) has extensive liquid thermal management experience.