NVIDIA launched the Rubin platform at CES on 5 January 2026, with Vera Rubin NVL72 as its rack-scale system: 72 Rubin GPUs and 36 Vera CPUs, connected by NVLink 6 switches, with ConnectX-9 SuperNICs and BlueField-4 processors. It was shown as “Vera Rubin NVL144” at GTC in March 2025; the new name counts GPU packages rather than dies. NVIDIA has not stated the rename itself; trade press reported the reason.

Per GPU, NVIDIA’s product page lists 288 GB of HBM4. Each Vera CPU has 88 custom NVIDIA Olympus Arm cores (176 threads) and up to 1.5 TB of LPDDR5X memory.

Each rack generation moves the limits that decide what a cluster can do: how large a model fits without splitting, how many users’ context fits in memory, and how fast GPUs exchange data during training and inference. Vera Rubin raises all three over Blackwell Ultra, at a higher rack power.

NVIDIA claims “up to 10x lower cost per token” and training mixture-of-experts models “with 4x fewer GPUs” against Blackwell. These are NVIDIA’s measurements on chosen workloads.

Figures are from NVIDIA’s product pages as read on 2 November 2026; the Vera Rubin page’s NVLink figure is lower than the January launch release. The Compute Rack page compares the two racks row by row with sources.

NVLink 6 is NVIDIA’s chip-to-chip interconnect: the NVLink switch trays in the middle of the rack let all 72 GPUs read each other’s memory directly, so the rack behaves like one large accelerator for a model split across it. NVIDIA’s January release gave 3.6 TB/s per GPU; the current product page gives 3 TB/s (216 TB/s for the rack).

HBM4 is the fourth generation of high-bandwidth memory, stacked beside the GPU die. More capacity per GPU means larger models or more users’ context per rack; more bandwidth means faster token generation, which is usually limited by how fast weights and cache can be read.

For the facility, Vera Rubin is fully liquid-cooled with 45 °C inlet water, which allows heat rejection without chillers in many climates, and it runs at a higher rack power than Blackwell (NVIDIA’s power-software blog cites 136 kW for a rack before power capping).

Construction

Halls built for air-cooled or hybrid racks may need rework for liquid distribution and floor loading.

Power & Cooling

Higher rack power and 100% liquid cooling: the facility needs liquid-cooling capacity per rack position and power distribution sized above Blackwell-era designs.

IT Infrastructure

A new rack generation to receive, qualify and integrate alongside existing Blackwell fleets, with new firmware and software baselines.

Networking

ConnectX-9 at 1.6 Tb/s per GPU doubles scale-out port speed, which drives switch and optics choices (see USFP and Spectrum-X Ethernet Photonics).

1. AnnouncedDescribed publicly; no date for customers to buy it.
2. PlannedA stated availability window that has not arrived yet.
3. Shipping (current stage)In production and reaching early customers or partners.
4. Generally availableOrderable by any customer through normal channels.
5. DeployedRunning in customers’ production environments, confirmed by them or the manufacturer.

NVIDIA said in January 2026 that Rubin was “in full production” and “available from partners the second half of 2026”; on 31 May that production shipments were “set to begin starting this fall”. Its product page now says “Available Now” and that partners are shipping Vera Rubin-based systems. First customer deliveries are reported in trade press but were not confirmed at the server makers for this entry.

• The Compute RackThe rack drawn tray by tray, and compared with GB300, GB200 and AMD Helios.
• Rack-scale systemsNVL72, NVL36x2, HGX and how rack-scale systems differ.
• The evolution of AI infrastructureHopper to Blackwell to Rubin on one timeline.

1. NVIDIA Kicks Off the Next Generation of AI With Rubin(opens in a new tab)NVIDIA Newsroom · 5 January 2026 · Manufacturer
2. NVIDIA Vera Rubin NVL72 (product page, read 2 November 2026)(opens in a new tab)NVIDIA · Technical documentation
3. NVIDIA Vera Rubin Ramps Into Full Production(opens in a new tab)NVIDIA Newsroom · 31 May 2026 · Manufacturer
4. NVIDIA GB300 NVL72 (product page, read 2 November 2026)(opens in a new tab)NVIDIA · Technical documentation

Figures are as the manufacturer states them unless marked otherwise. None has been independently tested by BSS/OSS Academy.