NVIDIA H100 SXM was the last GPU generation comfortably air-cooled at rack density. H200, B100, B200, and GB200 NVL72 push thermal density beyond what air can manage in standard UK colocation facilities. This is the practitioner's guide to the three cooling architectures + when each fits.
Air cooling — where it still works
Supermicro SYS-821GE-TNHR (8U 8× H100/H200 SXM) — air-cooled at ~10.5 kW per server. Fits standard 42U rack with 6-8 servers and thermal headroom in most UK colos.
NVIDIA HGX H100 SXM5 (700W per GPU) — manageable in air-cooled 8U chassis.
Reality: most UK enterprise AI clusters in 2026 are still air-cooled. Sufficient for inference + small-scale training.
Direct-to-chip (D2C) liquid cooling — where it fits now
B200 SXM (1000W per GPU) — requires direct-to-chip liquid cooling at rack density. Supermicro SYS-421GE-TNHR2-LCC 4U is the typical platform.
NVIDIA DGX B200 — D2C cooling. Requires facility CDU + manifold infrastructure.
Per-rack density: 25-35 kW achievable. Many UK colocation facilities now support D2C — confirm before procurement.
Full immersion cooling — frontier scale
NVIDIA GB200 NVL72 rack-scale superchip — ~120 kW per rack. Requires purpose-designed immersion-cooled facility.
Not all UK colos support this density. Limited to specialist AI-focused facilities (some hyperscaler-adjacent).
Frontier model training only — most UK enterprises don't need this.
UK colo reality 2026
Standard 4-8 kW/rack — most legacy UK colo.
15-25 kW/rack — modern colo supports for AI workloads.
30-60 kW/rack — newer AI-focused colo + some Tier 1 international.
60-120 kW/rack — specialist immersion / liquid-cooled AI facilities. Limited UK options.
What Servnet does
Servnet pre-surveys UK colocation power + cooling before quoting AI infrastructure. We have established relationships with UK colos supporting 25-60+ kW/rack — see our colo procurement guide.