NVIDIA Vera Rubin Enters Full-Production Ramp for Rack-Scale AI Factories

As of Sunday, June 21, 2026, NVIDIA Vera Rubin is no longer just a future platform on a keynote slide. NVIDIA said on May 31, 2026 that Vera Rubin is ramping into full production, with production shipments set to begin this fall. That timing matters because the platform is aimed at the next wave of large AI deployments: not single servers, but POD-scale AI factories built from multiple specialized racks working together. NVIDIA positions Vera Rubin as a major step beyond Grace Blackwell for agentic AI workloads, claiming 10x agent throughput at scale compared with the previous-generation Grace Blackwell platform. (nvidianews.nvidia.com)

The most interesting part of Vera Rubin is that NVIDIA is not describing it as a standalone GPU launch. The platform combines Vera Rubin NVL72 systems, Vera CPU racks, NVIDIA Groq 3 LPX inference accelerator racks, Vera BlueField-4 STX storage racks, and Spectrum-6 SPX Ethernet racks into one integrated design. In NVIDIA’s language, the AI factory is becoming a product-level system where compute, networking, storage, security, power, cooling, and management all need to be planned together. That shift is important for agentic AI because these workloads can involve long context, tool use, retrieval, code execution, reinforcement learning environments, and high-volume inference running at the same time. (nvidianews.nvidia.com)

View NVIDIA DGX Spark Personal AI Desktop Supercomputer on partner website

For readers tracking the hardware details, the Vera Rubin NVL72 rack integrates 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs, along with ConnectX-9 SuperNICs and BlueField-4 DPUs across 18 compute trays and 9 NVLink switch trays. NVIDIA says sixth-generation NVLink provides 3.6 TB/s of bandwidth per GPU and 260 TB/s of scale-up bandwidth per rack. The Vera CPU itself uses 88 custom NVIDIA Olympus cores, supports NVIDIA Spatial Multithreading, and is paired with GPUs through NVLink-C2C at 1.8 TB/s of coherent bandwidth. On the inference side, NVIDIA Groq 3 LPX is a rack-scale accelerator with 256 LPU accelerators per rack, 315 PFLOPS of FP8 compute, 128 GB total SRAM, and 640 TB/s scale-up bandwidth. (developer.nvidia.com)

See ASUS GeForce RTX 5090 price

The “million-GPU AI factory” phrase is not only about placing more accelerators in data centers. At that scale, networking becomes one of the main limits. Vera Rubin introduces Spectrum-X Ethernet Photonics, which NVIDIA describes as a co-packaged-optics switching technology with 200Gb/s SerDes, now in production. NVIDIA says this approach improves power efficiency and uptime compared with traditional transceiver-based networks, while freeing more power for compute. The platform also integrates BlueField-4 DPUs with software-defined networking up to 800Gb/s, multi-tenant isolation, confidential computing support, and security features designed for shared AI infrastructure. (nvidianews.nvidia.com)

Buy SANDISK Optimus GX PRO 8100 PCIe 5 SSD here

A platform this large only becomes relevant if system builders can actually produce and deploy it. NVIDIA says hundreds of ecosystem partners across more than 350 factories and 30 countries are involved in the Vera Rubin ramp, including major names such as Dell Technologies, HPE, Lenovo, Supermicro, ASUS, Foxconn, GIGABYTE, QCT, Wistron, and Wiwynn. That broad manufacturing base is key to Vera Rubin’s role in hyperscale and cloud AI infrastructure because rack-scale systems require more than chip availability; they need validated mechanical designs, liquid cooling, storage integration, networking, firmware, and serviceability. The takeaway is simple: Vera Rubin turns the AI hardware conversation from “how fast is the GPU?” into “how efficiently can the whole factory produce tokens?” (nvidianews.nvidia.com)

View Lenovo Legion Pro 7i laptop with Core Ultra 9 and NVIDIA RTX 5090 24GB on partner website