NVIDIA Vera Rubin Enters Production: The New Generation of AI Factories for the Era of Intelligent Agents

NVIDIA has announced the start of volume production for the Vera Rubin platform, an architecture intended to serve as the foundation for future AI factories designed to run large-scale agent workloads. The company states that the first systems will be available starting this fall, while the global supply chain is already engaged in creating the new infrastructures.

The Vera Rubin platform represents the third generation of rack-scale systems developed by NVIDIA based on the MGX architecture. According to the company, over 150 partners in Taiwan and more than 350 facilities distributed across 30 countries are involved in the production of the new systems, supported by manufacturers such as Dell Technologies, HPE, Lenovo, Supermicro, ASUS, Foxconn, GIGABYTE, Pegatron, and Quanta Cloud Technology.

The declared goal is to support the growth of so-called AI agents, systems capable of executing complex processes that include reasoning, information retrieval, the use of external tools, and response generation. For this type of application, NVIDIA claims that Vera Rubin can offer processing power up to ten times greater than the current Grace Blackwell platform when employed on a large scale.

The architecture is organized around a POD made up of five specialized racks that operate as a single supercomputer. Inside are the Vera Rubin NVL72 systems, Vera processors, storage infrastructure based on BlueField-4 STX, and Spectrum-6 SPX Ethernet racks, all integrated into a single platform designed to maximize performance in distributed AI workloads.

Among the key technological innovations is the introduction of NVIDIA Spectrum-X Ethernet Photonics, a new generation of switches based on co-packaged optics (CPO) technology. The solution combines Spectrum-X switching electronics with optical components integrated directly into the package, eliminating some of the typical complexities of traditional transceivers.

According to NVIDIA, this architecture allows for a fivefold improvement in network energy efficiency, a fivefold increase in AI system operational time, and a 30% acceleration in deployment times compared to conventional Ethernet infrastructures. The technology is already in production and is proposed as a fundamental element for creating AI factories capable of scaling up to one million GPUs.

On the connectivity front, Vera Rubin also integrates BlueField-4 DPUs, capable of handling software-defined networking up to 800 Gb/s and providing advanced tenant isolation mechanisms. In this way, NVIDIA aims to simplify the management of extremely large clusters while improving security and operational control.

Security is indeed one of the central aspects of the new platform. AI factories are increasingly used to process proprietary data, regulated content, and critical business models, necessitating an approach that protects the entire infrastructure.

To this end, Vera Rubin adopts an extended NVIDIA Confidential Computing model across the entire hardware stack. The NVL72 systems integrate data encryption capabilities on high-speed interconnections and hardware attestation mechanisms that allow for verifying platform integrity and detecting any tampering.

At the software level, the NVIDIA DOCA platform comes into play, enabling the implementation of zero-trust security policies, multi-tenant isolation, real-time threat detection, and end-to-end encryption up to 800 Gb/s without burdening host CPUs. The goal is to ensure the protection of data, models, operational context, and inferential processes even in shared cloud environments.

To accelerate the build-out of new infrastructures, NVIDIA also offers DSX, a platform that combines reference designs, simulation tools, infrastructural software, and operational management technologies. DSX has been specifically developed for the Vera Rubin architecture and aims to reduce the implementation times of AI factories while optimizing energy consumption and operational costs.

Among the first operators to announce the adoption of the technologies related to Vera Rubin are several cloud providers and AI specialists, including CoreWeave, Lambda, Oracle Cloud Infrastructure, Microsoft Azure, IBM Cloud, and Nebius.