Layer 1 · How the heat leaves
Single-phase immersion
Uniform high-density fleets built for it from day one.
The shape of it
- Capital cost
- High, and it changes the whole facility design
- Maturity
- proven
- Density supported
- 100+ kW per rack equivalent
- PUE
- 1.02-1.10
Verified 2026-09-06.
The catch
Direct-to-chip is the 2026 production method; immersion is not. Serviceability, hardware warranties and the fact that you cannot mix it into a normal hall are the reasons, not thermal performance — where it wins.
What it actually decides
Best PUE available. Chosen by people building a purpose-built fleet, not by people upgrading a datacentre.
What this is powering
A GB200 NVL72 rack draws 120–132 kW and a GB300 pushes 135–200 kW, against a 2026 average rack of about 27 kW. The silicon decision and the power decision are the same decision, made eighteen months apart.
Layer 2 — Silicon · GB200 NVL72 · The 150 W inference option
Weighed against
- Direct-to-chip liquid cooling vs Single-phase immersion — The production method against the thermally superior one. Serviceability decides it, not physics.
- Two-phase immersion vs Single-phase immersion — Marginal thermal gain against a live PFAS regulatory question.
Others in how the heat leaves
- Air cooling — Up to ~30 kW per rack
- Rear-door heat exchangers — ~30-50 kW per rack
- Direct-to-chip liquid cooling — 50-200+ kW per rack
- Two-phase immersion — Highest available
- Free and evaporative cooling — Depends on the paired system