NVIDIA Rubin

NVIDIA Rubin is the successor to NVIDIA's Blackwell data-centre platform, named after the American astronomer Vera Rubin, whose observations of galaxy rotation provided early evidence for dark matter. Rubin systems combine a new Rubin GPU architecture with the NVIDIA Vera CPU and, for the first time, an inference tier licensed from the AI chip company Groq, and are designed to power the large "AI factories" that train and serve frontier artificial intelligence models.^[1][2]

Background

NVIDIA describes Vera Rubin as the third generation of its MGX rack-scale system family and its first platform built around POD-scale computing, in which multiple purpose-built racks work together as one coherent supercomputer. The platform was detailed across NVIDIA's 2026 GTC events: at GTC in March 2026 the company announced seven new chips in full production, and at GTC Taipei in May 2026 it announced the platform was ramping into full production, with production shipments beginning in the second half of the year.^[1][3]

Technology

The core of the platform is the Vera Rubin NVL72 rack, which integrates 72 Rubin GPUs and 36 Vera CPUs connected by NVIDIA's NVLink 6 interconnect. NVIDIA says the system can train large mixture-of-experts models with one quarter of the GPUs required by the Blackwell generation, and delivers up to 10 times higher inference throughput per watt at one tenth the cost per token. Independent breakdowns of published specifications list the Rubin GPU at 336 billion transistors, fabricated on TSMC's N3 process with 288 GB of HBM4 memory and 22 terabytes per second of bandwidth.^[1][4]

Two companion racks extend the platform. A Vera CPU rack packs 256 liquid-cooled Vera processors for the reinforcement-learning environments and agent workflows that surround model training.^[1] The NVIDIA Groq 3 LPX, built under licence from Groq, places 256 LPU inference accelerators with 128 GB of on-chip SRAM in a single rack; deployed alongside the GPU racks it is intended to serve trillion-parameter models at long context with up to 35 times higher inference throughput per megawatt. Groq's founder, Jonathan Ross, joined NVIDIA as part of the licensing arrangement.^[1][5][6]

Networking and storage complete the system: Spectrum-X Ethernet with co-packaged optics for million-GPU scale-out, BlueField-4 DPUs for multi-tenant isolation and confidential computing, and BlueField-4 STX storage racks that offload the key-value cache that long-running AI agents generate. NVIDIA states that more than 200 data centre infrastructure partners are supporting the generation.^[1][3]

Availability and impact

Vera Rubin products are scheduled to ship from the second half of 2026 through cloud providers including AWS, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure, alongside NVIDIA Cloud Partners such as CoreWeave, Crusoe, Lambda, Nebius, Nscale and Together AI, and systems from Dell, HPE, Lenovo, Supermicro and Taiwanese manufacturers. AI laboratories including Anthropic, Meta, Mistral AI and OpenAI have indicated they intend to use the platform for training and serving large models.^[1]

The generation arrives as the industry's focus shifts from single models toward fleets of autonomous agents, a workload that demands both extreme throughput for long tasks and low-latency interactivity for user-facing responses - the reason the platform pairs GPUs with a dedicated LPU inference tier. As with earlier NVIDIA generations, the most advanced parts fall under United States export-control rules that restrict sales to certain countries.^[1][3]

>See Also

• NVIDIA
• NVIDIA Blackwell
• NVIDIA H100
• Groq
• High Bandwidth Memory
• AI Chip Export Controls