Cloud Architecture
AI Data Center Liquid Cooling: Direct-to-Chip, Immersion, and the Thermal Architecture Decisions Every GPU Cluster Gets Wrong
The GB200 NVL72 draws 120-130 kW per rack. Air cooling is dead for serious AI clusters. Here is a practitioner's guide to direct-to-chip cooling, single-phase and two-phase immersion, facility requirements, and how to pick the right thermal architecture before you pour concrete.