Nvidia has published details of the rack behind its next AI platform, Vera Rubin NVL72. Two changes stand out for anyone who builds, cools or services data center equipment: the compute trays have no internal cables, hoses or fans, and the whole rack is designed to be cooled with warm water. The figures below are Nvidia's own, as of Oct 24, 2026. Nvidia says the rack is in full production and on track to ship in the second half of 2026.
The short version: fewer cables and fans inside the rack means faster assembly and service, and warm-water cooling means many sites can reject heat with outside air and dry coolers instead of running compressor chillers.
What is in the rack
According to Nvidia, one Vera Rubin NVL72 rack holds 72 Rubin GPUs and 36 Vera CPUs across 18 compute trays, plus 9 NVLink switch trays. The trays connect to a copper "spine" at the back of the rack. Nvidia puts the rack's weight at about 4,000 lbs.
No cables, hoses or fans in the trays
- Parts plug into a circuit-board midplane instead of being wired together with internal cables.
- Nvidia says assembling a compute tray drops from about 1.5 to 2 hours on its previous Blackwell design to about 5 minutes.
- Nvidia says servicing is up to 18 times faster, and switch trays can be swapped while the rack keeps running.
How the warm-water cooling works
Nvidia says the racks can be 100% liquid cooled and accept coolant arriving at up to 45°C (113°F). Coolant flows through the trays, picks up heat, and returns to a coolant distribution unit (CDU). The CDU hands the heat to a facility loop, which a dry cooler (a closed-loop radiator with fans) releases to outside air. Because the water can be that warm, Nvidia says data centers in many climates can skip most compressor-based chilling, and the power saved could run up to 10% more racks in the same power budget. The coolant, typically deionized water or a propylene-glycol mix, stays sealed in the loop.
Warmer coolant does not reduce the heat the chips produce. It reduces the extra energy spent removing it.
What this means for facilities
- Rooms planned for these racks need liquid cooling infrastructure (CDUs, piping, leak containment), not only air handling.
- Fewer internal cables does not mean less cabling overall: power, network and facility connections still need planning, labeling and testing.
- Floor loading matters at roughly 4,000 lbs per rack.
Video: Nvidia's official overview, "The NVIDIA Rubin Platform: Six New Chips, One AI Supercomputer," shows the tray with no cables, hoses or fans: watch on YouTube (NVIDIA channel).
Sources
- NVIDIA Technical Blog: NVIDIA Vera Rubin POD (Mar 16, 2026)
- NVIDIA Technical Blog: Inside the NVIDIA Rubin Platform (Jan 5, 2026)
- NVIDIA Newsroom: Rubin platform announcement (Jan 5, 2026)
All figures are as published by Nvidia and checked Oct 24, 2026. Diagram drawn by XOtavo from Nvidia's description; not to scale.
Planning a room for liquid-cooled equipment, or the power, network and cabling around it? Call 713.980.2000 or send us a note.
Let's Talk →