A short lesson sequence. Work the notes in numbered order; each builds on the last.
Part I: The Domain. Every megawatt of compute becomes a megawatt of heat. Cooling is the second half of the power equation — and the constraint that breaks high-density plans.
Assumes: Hub 01. You can't cool what you can't quantify — the heat load comes from the power chain.
Learning path
02.10 Air Cooling: CRAH, Chillers, and Heat Rejection
The traditional cooling stack — CRAH/CRAC units, chiller plants, cooling towers and dry coolers — how each removes heat and where each sits in procurement (chillers 45-65 weeks, CRAH 35-50 weeks; Trane, Carrier, Daikin, Vertiv, STULZ). The hard limit: air cooling breaks down past roughly 30 kW per rack, which is why the rest of this hub exists. Builds on: 01.20 — every MW of compute is a MW of heat to reject.
02.20 Direct-to-Chip Liquid Cooling
Cold plates, coolant distribution units (CDUs), manifolds, quick-disconnects — the technology that took over 40% of the liquid-cooling market as a drop-in fit for standard servers. Procurement angle: liquid cooling splits off from HVAC into its own buying track (CDUs by the hundred per project), introduces new specification concerns (fluid chemistry, flow rates, leak detection), and is now the default configuration in AI server RFPs. Builds on: 02.10 — the air-cooling limit that forces the move to liquid.
02.30 Immersion Cooling: Single-Phase vs Two-Phase
Full-server immersion baths: single-phase (pumped dielectric fluid) vs two-phase (boiling fluorocarbons), the near-1.0 PUE claims, and the real costs — fluids at hundreds of dollars per gallon and PFAS regulatory exposure. When immersion becomes near-mandatory (racks heading toward 600 kW-1 MW) and how to evaluate an immature vendor landscape without getting stranded on a dead-end platform. Builds on: 02.20 — where direct-to-chip runs out of headroom.
02.40 Water: WUE and the Water-Energy Tradeoff
Water Usage Effectiveness (liters per kWh of IT energy), benchmarks from near-zero to 2.5 L/kWh, and the classic tradeoff: evaporative cooling buys PUE with water, air cooling buys water with PUE — while closed-loop liquid breaks the tradeoff entirely. Water is becoming a siting and permitting constraint, so cooling technology must be evaluated against local water risk, not just thermal performance — including how to diligence water risk during site selection and write it into cooling specifications. Builds on: 02.10 (the evaporative-vs-air tradeoff), 02.20 (closed-loop liquid breaking it).
02.50 Cooling Procurement in the AI Era
Synthesizing the hub into buying practice: how the shift to liquid changes RFQs (separate liquid-cooling tracks, CDU counts, vendor qualification on GPU-platform compatibility), the vendor consolidation wave, and why specifications must be written for the next generation's rack densities — a cooling system bought for today's density is obsolete on arrival. Builds on: 02.10-02.40 — the full hub, applied to buying.
Key sources: Mordor Intelligence (liquid cooling market); Schneider Electric warm-water cooling analysis; The Green Grid WUE methodology; DataCenterKnowledge on the liquid cooling vendor landscape.