Layer 1 · How the heat leaves
Direct-to-chip liquid cooling
Anyone deploying GB200-class hardware. NVIDIA has confirmed liquid cooling is a mandatory architectural requirement for GB200 NVL72, not a preference.
The shape of it
- Capital cost
- Higher than air; lower than immersion
- Maturity
- proven
- Density supported
- 50-200+ kW per rack
- PUE
- 1.1-1.2
Verified 2026-09-09.
The catch
Cold plates capture most of the heat; the rest still leaves as air, so you keep a hybrid air path. It is not a full replacement for the existing system.
What it actually decides
The 2026 production method, full stop. It holds 47% of the AI datacentre liquid cooling market on the strength of standard rack compatibility and a straightforward retrofit path.
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.
- Air cooling vs Direct-to-chip liquid cooling — Where the crossover actually sits: about 30 kW a rack on cost, 45-50 kW on thermals.
Others in how the heat leaves
- Air cooling — Up to ~30 kW per rack
- Rear-door heat exchangers — ~30-50 kW per rack
- Single-phase immersion — 100+ kW per rack equivalent
- Two-phase immersion — Highest available
- Free and evaporative cooling — Depends on the paired system
Sources
Want this whole layer as one table?
The power and cooling options on these pages as a single CSV — lead time, capex, carbon, maturity, density supported, PUE, the catch and what each one decides — ready to sort, filter or load into a notebook.
Get the energy table — free for research and evaluation
Four short lines and the download starts at once (CC-BY-NC-4.0, built from these pages the moment you ask). What you type is used only to answer the request and is never sold. Prefer Hugging Face? Request it there.
Using it inside a product, a client report or anything you sell? That needs a commercial licence.