GPU rack power density and cooling requirements - Air cooling ceiling and liquid cooling requirement

DERIVED ANALYSIS from existing repo records (pwr2-vantage-power-density, pwr2-vantage-cooling, gb200-nvl72-rack-120kw-liquid-required, gb300-nvl72-rack-1400w-per-gpu, liquid-cooling-retrofit-cdu-density-ceiling, colo-vantage-liquid-to-rack-not-to-chip). Power density is a first-order deployment constraint. Key data points: (1) GB200 NVL72 rack: 120kW, requires liquid cooling. (2) GB300 NVL72 rack: 1,400W per GPU, 72 GPUs = ~100kW per rack. (3) H100 SXM: ~700W per GPU, 8 GPUs = ~5.6kW per node, air-coolable. (4) H200 SXM: ~700W per GPU, 8 GPUs = ~5.6kW per node, air-coolable. (5) B200: ~1,000W per GPU, 8 GPUs = ~8kW per node, may require liquid cooling depending on chassis. (6) Air cooling ceiling: approximately 15-20kW per rack for standard datacenter air cooling. (7) Liquid cooling required above ~20kW per rack. (8) Vantage power density tiers: 8.3-20 kW/rack (Tier 1), 20-50 kW/rack (Tier 2), 50-100+ kW/rack (Tier 3, liquid cooling required). The power density constraint means that deploying GB200 NVL72 or GB300 NVL72 racks requires a liquid-cooled datacenter, which limits deployment options and increases colocation costs.