Field Guide · NVIDIA Rubin_
NVIDIA Rubin: The Deployment Guide_
Everyone is talking about Rubin's specs. Fewer people are talking about what it takes to actually deploy it. Vera Rubin NVL144 — the platform after Blackwell — pairs the Rubin GPU with NVIDIA's Vera CPU, HBM4, and NVLink 6, and pushes rack power toward levels that break today's assumptions. This guide covers what changes for power, cooling, and the cable plant, and how to make your build Rubin-ready now.
What Changes for Deployment_
800V DC distribution becomes baseline. Racks climb from ~120 kW (GB300) toward ~600 kW at Rubin Ultra — busway, breaker, and PDU capacity must be planned a generation ahead.
Direct-to-chip liquid cooling is mandatory and must scale. CDU capacity and facility water loops need headroom for the higher heat density; air-assist is no longer relevant at the rack.
NVLink 6 (~3.6 TB/s/GPU) and ConnectX-9 / 1.6T networking. Higher-radix, higher-bandwidth topologies mean a denser, more demanding cable plant and tray routing.
Heavier, denser trays and a larger NVLink domain (NVL144 → NVL576) raise floor-loading, structural, and blind-mate alignment requirements.
Specifications reflect NVIDIA's public roadmap disclosures and will firm up as launch nears. The deployment implications, however, are already clear: plan for more power, more cooling, and a denser fabric than Blackwell.
Rubin Readiness Checklist_
- 01Design power for 800V DC and reserve busway capacity well above current 120 kW racks.
- 02Size CDUs and facility water loops with headroom for Rubin-class heat density — not just today's load.
- 03Plan the back-end fabric for 1.6T (ConnectX-9) and higher-radix spine-leaf topologies.
- 04Verify floor loading and structural pathways for heavier NVL-class racks.
- 05Stage cable-plant standards (TIA-606 labeling, IEC 61300-3-35 fiber inspection) before density climbs.
- 06Engage your physical-integration partner during design — not after the hardware ships.
NVIDIA Rubin FAQ_
When is NVIDIA Rubin available?
NVIDIA's public roadmap targets the Vera Rubin NVL144 platform for the second half of 2026, with Rubin Ultra (NVL576) following in 2027. Operators planning new builds today should design for Rubin-class power and cooling now, because retrofitting later is far costlier.
How is Rubin different from Blackwell (GB200/GB300)?
Rubin pairs the new Rubin GPU with NVIDIA's custom Arm-based Vera CPU, moves to HBM4 memory, and steps up to NVLink 6 (~3.6 TB/s per GPU) and ConnectX-9 / 1.6T networking. Per-rack power rises sharply — Rubin Ultra trends toward ~600 kW racks versus ~120 kW for GB200/GB300 — making 800V DC distribution and high-capacity liquid cooling the baseline rather than the exception.
What power and cooling does NVIDIA Rubin require?
Rubin-class racks exceed today's 120 kW envelope and push toward ~600 kW at the Rubin Ultra tier. That means direct-to-chip liquid cooling is mandatory, CDU capacity and facility water loops need real headroom, and power should be planned around 800V DC distribution and higher busway capacity. Under-provisioned cooling or power becomes a hard ceiling on what you can deploy.
What is Vera Rubin NVL144?
Vera Rubin NVL144 is the flagship Rubin rack system — 144 Rubin GPU dies in a single NVLink domain, combined with Vera CPUs, HBM4 memory, and NVLink 6. It is the successor to the GB300 NVL72 and the reference platform most large AI operators will deploy first.
Who will deploy NVIDIA Rubin infrastructure?
The teams already executing Blackwell (GB200/GB300) deployments today are the ones positioned to deploy Rubin first — the physical-integration skills carry over and intensify. Leviathan Systems performs GPU rack assembly, structured cabling, InfiniBand/NVLink and back-end fabric, liquid-cooling integration, and commissioning for AI-scale data centers across the United States, and is actively planning power, cooling, and fabric for Rubin-class deployments now.
Ready to Deploy Your GPU Infrastructure?_
Tell us about your project. Book a call and we’ll discuss scope, timeline, and the best approach for your deployment.
Book a Call