NVIDIA Opens Vera Rubin Architecture to Accelerate the Global Expansion of AI Factories
NVIDIA is expanding its commitment to open infrastructure standards as the global demand for artificial intelligence (AI) computing continues to increase. At the Open Compute Project (OCP) Global Summit in San Jose, California, the company announced new partnerships, hardware design contributions, and open-source software initiatives aimed at helping businesses build and operate next-generation AI factories more efficiently.
As AI infrastructure expands from megawatt-scale facilities to systems requiring gigawatts of power, companies face growing challenges involving electricity distribution, cooling, networking, system reliability, and deployment speed. NVIDIA’s latest initiatives aim to address these challenges by establishing common standards that equipment manufacturers, cloud providers, and data center operators can adopt across their infrastructure projects.
A central announcement was NVIDIA’s contribution of Vera Rubin rack and compute tray designs to OCP through its open NVIDIA MGX architecture. The contribution provides system builders with a shared foundation for developing advanced AI infrastructure and scaling deployments across different applications.
The announcements also included open standards for AI factory facilities and grid resilience, new open-source software for the NVIDIA DSX AI Factory platform, networking advancements developed with Meta, and new power distribution designs intended for future high-density computing environments.
Together, these initiatives demonstrate NVIDIA’s strategy of combining tightly integrated computing, networking, power, cooling, and software with open standards to improve infrastructure efficiency and accelerate the global deployment of AI factories.
NVIDIA Expands the Open MGX Architecture for Vera Rubin
NVIDIA has contributed its third-generation MGX AI data center reference designs to OCP’s Open DC Initiative, extending its modular architecture to support future AI factory deployments based on the NVIDIA Vera Rubin platform.
The updated MGX 1.2 specification introduces several infrastructure improvements designed to support increasingly powerful AI systems. These include rack-scale blind-mate liquid cooling, interoperable rack-scale interconnects, and a liquid-cooled 5,000-amp busbar designed for high-power computing environments. The specification also supports native 19-inch equipment integration and higher-density rack configurations.
Additional improvements include configurable input/output bays, fanless liquid-cooled compute trays, and standardized power and cooling interfaces. These features are intended to make it easier for ecosystem partners to develop compatible systems while maintaining flexibility in hardware selection and infrastructure design.
The architecture supports rack deployments of up to 250 kilowatts and emphasizes improved manufacturing, assembly, testing, maintenance, and field servicing. By standardizing key interfaces, NVIDIA aims to help operators deploy AI infrastructure faster while simplifying upgrades and equipment replacement.
Open liquid-cooling technologies, including the MQD and MQDB approaches, complement these efforts by helping establish common methods for managing the substantial heat generated by high-performance AI systems.
NVIDIA is also extending the MGX architecture beyond accelerated computing with a dedicated Vera CPU rack specification. While Vera Rubin NVL72 is designed to support large-scale AI training and inference, the Vera CPU rack targets workloads associated with agentic AI, including tool execution, orchestration, and sandboxed computing.
These workloads allow AI agents to interact with applications, data, and digital services in real time. Using shared design principles for computing, networking, power, and cooling, the Vera CPU rack is intended to help infrastructure providers deploy specialized systems within a common, modular ecosystem.
New Open-Source Software Simplifies AI Factory Operations
Alongside its hardware contributions, NVIDIA introduced new open-source software components for its DSX AI Factory platform. The platform applies a system-wide design approach to coordinate computing, networking, electricity, cooling, facilities, and software.
At the center of this architecture is DSX OS, an operating layer intended to help AI factories function as coordinated systems rather than collections of independent hardware components.
The newly announced software tools focus on network configuration, cluster validation, rack management, and bare-metal infrastructure provisioning.
NVIDIA Config Manager automates network switch configuration, helping operators deploy and maintain consistent AI networking environments with less manual intervention. This can simplify large-scale infrastructure management and reduce the likelihood of configuration inconsistencies.
The NVIDIA Cluster Readiness Engine uses simulations of real workloads to stress-test GPU clusters before production training begins. By identifying hardware problems in advance, the tool aims to reduce costly job failures and improve infrastructure readiness.
The NVIDIA Rack Management Service provides a hardware abstraction and operational framework for managing racks as standardized units. This approach is designed to simplify firmware administration and improve consistency across large AI deployments.
Meanwhile, the Cluster API Provider for NVIDIA Infrastructure Controller is intended to accelerate bare-metal deployment and tenant onboarding, helping organizations turn newly installed hardware into production-ready computing clusters more quickly.
Developed using experience gained from operating NVIDIA DGX Cloud, these open-source components are designed to help cloud providers and software partners integrate AI factory capabilities into multi-tenant platforms.
Developers can explore and contribute to the software through NVIDIA’s DSX AI Factory GitHub resources.
NVIDIA and Hyperscalers Establish Common Infrastructure Standards
NVIDIA also joined AMD, Google, and Meta in publishing the Open Data Center Specification v0.7, which establishes a common framework for next-generation AI factory facilities.
The specification addresses essential elements of data center design, including electrical power architecture, cooling infrastructure, and physical facility layouts. By creating a shared technical reference, the initiative aims to help manufacturers, colocation providers, and infrastructure operators develop compatible systems without relying entirely on individual customers’ proprietary requirements.
A separate OCP Grid Resilience Specification, published by Google, Meta, Microsoft, and NVIDIA, focuses on how large AI facilities should respond to disturbances in the utility power grid.
The specification establishes expectations for remaining connected during grid disturbances, responding automatically to changing conditions, and restoring operations to near-full load in less than two seconds.
These capabilities are increasingly important as AI facilities consume more electricity and place greater demands on local power infrastructure. Coordinated responses to grid disturbances can help improve operational stability and reduce disruptions to computing workloads.
NVIDIA’s DSX reference designs are intended to provide a practical implementation path for these standards, connecting facility planning and power architecture with software capable of monitoring and responding to grid signals.
By making these specifications openly available, the companies aim to give the broader industry a common foundation for designing, manufacturing, and deploying the infrastructure required by future AI factories.
Source Link: https://blogs.nvidia.com/