Unboxing the AI Stack
Chip · L08 · Chips & Memory

NVIDIA Rubin

Rubin is NVIDIA’s data-center GPU after Blackwell, launched at CES on January 5, 2026 as the GPU of the Vera Rubin platform, which NVIDIA named for the astronomer Vera Florence Cooper Rubin and pairs with its own Vera CPU.1 NVIDIA’s technical blog gives the Rubin GPU 336 billion transistors across two compute dies, against 208 billion for Blackwell, with 224 streaming multiprocessors and a third-generation Transformer Engine that uses “hardware-accelerated adaptive compression” to raise 4-bit throughput.2 Each GPU carries 288 GB of HBM4, which puts Rubin at #1 (tied) in AI Accelerators With the Most HBM per Chip (October 2026).3

NVIDIA’s datasheet rates one Rubin GPU at 50 PFLOPS of NVFP4 for inference, a sparse figure, and 35 PFLOPS dense, with 17.5 PFLOPS of dense FP8. Its memory bandwidth is 19.2 TB/s in the 72-GPU Vera Rubin NVL72 rack, while NVIDIA’s HGX page lists 22 TB/s, the figure NVIDIA gave at launch.342 Rubin ships in three main forms: the Vera Rubin NVL72 rack (72 Rubin GPUs and 36 Vera CPUs), the eight-GPU HGX Rubin NVL8 board and the four-GPU Vera Rubin NVL4.3

On May 31, 2026, NVIDIA said Vera Rubin was ramping into full production, with production shipments set to begin in the fall.5 CoreWeave received its first Vera Rubin NVL72 production racks in September 2026 and made the system available on its cloud on September 30, according to NVIDIA.6 The first third-party-verified results came in MLPerf Inference v6.1 in September 2026, where Nebius’s 36-GPU preview system served 591,368 DeepSeek-R1 tokens per second in the server scenario.7 OpenAI’s 2025 letter of intent with NVIDIA targets its first gigawatt of NVIDIA systems on Vera Rubin in the second half of 2026.8 The platform’s specs, NVIDIA’s cost-per-token claims and the first results are set out in NVIDIA Vera Rubin by the Numbers.

Illustration: the ten layers of the AI stack, with Chips & Memory, the Rubin GPU’s layer, pulled out.

Key facts

MakerNVIDIA1
LaunchedJanuary 5, 2026 (CES)1
Transistors336 billion per GPU, two compute dies Vendor-stated2
Memory per GPU288 GB HBM4; 19.2 TB/s in Vera Rubin NVL72, 22 TB/s on the HGX page Vendor-stated34
NVFP4 per GPU50 PFLOPS sparse; 35 PFLOPS dense Vendor-stated3
FP8 / FP6 per GPU17.5 PFLOPS dense Vendor-stated3
NVLink 6 per GPU3 TB/s (datasheet); 3.6 TB/s at launch Vendor-stated31
MLPerf Inference v6.1, DeepSeek-R1 server591,368 tokens/s on 36 GPUs (Nebius, preview) Measured7
First cloud availabilitySeptember 30, 2026 (CoreWeave, Vera Rubin NVL72)6

As of . NVIDIA marks its NVFP4 inference figure as sparse and its training figures as dense; the two aren’t comparable. The MLPerf result was verified by MLCommons and run by the submitter. How figures are labeled.

Appears in

Sources

  1. NVIDIA Newsroom, “NVIDIA Kicks Off the Next Generation of AI With Rubin — Six New Chips, One Incredible AI Supercomputer” (January 5, 2026), retrieved 2026-10-19.
  2. NVIDIA Technical Blog, “Inside the NVIDIA Vera Rubin Platform: Six New Chips, One AI Supercomputer” (January 5, 2026), retrieved 2026-10-19.
  3. NVIDIA, NVIDIA Vera Rubin Datasheet (PDF), retrieved 2026-10-19.
  4. NVIDIA, NVIDIA HGX Platform (HGX Rubin NVL8 specifications), retrieved 2026-10-19.
  5. NVIDIA Newsroom, “NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide” (May 31, 2026), retrieved 2026-10-19.
  6. NVIDIA Blog, “From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI” (September 30, 2026), retrieved 2026-10-19.
  7. Nebius, “MLPerf Inference v6.1: Previewing NVIDIA Vera Rubin NVL72 and leading full-rack Nebius system with NVIDIA GB300 NVL72 results” (September 16, 2026), retrieved 2026-10-19.
  8. NVIDIA Newsroom, “OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems” (September 22, 2025), retrieved 2026-10-19.