NVIDIA is advancing the development of AI factories by opening its Vera Rubin architecture, providing a common foundation for next-generation AI infrastructure. This initiative, highlighted at the Open Compute Project (OCP) Global Summit, aims to help scale AI factories globally through open standards and new partnerships. The focus is on improving deployment, power efficiency, and tokens-per-watt performance across the entire AI ecosystem, from the factory itself to the utility grid and rack-scale architecture.
Open Standards for AI Infrastructure
At the OCP Global Summit, NVIDIA, alongside partners like AMD, Google, and Meta, published new open specifications to standardize AI factory facility design. The Open Data Center Specification v0.7 outlines a unified blueprint covering power architecture, cooling infrastructure, and physical layout. Additionally, the OCP Grid Resilience Specification, also published by Google, Meta, Microsoft, and NVIDIA, details how large-scale AI infrastructure should behave during grid disturbances, ensuring systems remain connected, respond automatically, and restore to near-full load within two seconds.
These open standards provide manufacturers, colocation providers, and data center operators with a common target for building and delivering the infrastructure layer that powers AI factories. This approach allows for the design and fabrication of necessary equipment without relying on proprietary specifications from individual customers, streamlining the global construction of AI factories. For operators building Vera Rubin AI factories, this creates a direct path from open community standards to a fully operational system.
Expanding MGX Architecture for AI Factories
NVIDIA contributed key Vera Rubin rack and compute tray designs to OCP through its open NVIDIA MGX architecture. This move provides system builders with a common foundation for developing next-generation AI factory infrastructure. The updated MGX 1.2 specification extends this open architecture for AI factories to include rack-scale blind-mate liquid cooling, interoperable rack-scale interconnect infrastructure, and a liquid-cooled 5,000A busbar designed for higher-power AI systems. It also adds native 19-inch equipment support and higher-density rack designs, supporting deployments up to 250kW.
The MGX 1.2 specification also enhances modularity with configurable I/O bays, fanless liquid-cooled compute trays, standardized power and cooling interfaces, and a common infrastructure that facilitates interoperable accelerated computing platforms. This modularity is intended to improve assembly, test, and service times, streamlining manufacturing and field service to enable faster and more efficient infrastructure buildout.
Supporting the Agentic AI Era
Beyond large-scale training and inference, the Vera CPU rack design specification extends the NVIDIA MGX open architecture to support the growing era of agentic AI. While Vera Rubin NVL72 is foundational for large-scale training and inference, the NVIDIA Vera CPU rack is specifically designed for tool calling, orchestration, and sandboxed execution. This design helps AI agents interact with applications, data, and services in real time. It shares the same rack-scale design principles that unify compute, networking, power, and cooling into a common modular platform.
This approach enables partners to deploy specialized infrastructure for agentic AI while benefiting from a shared ecosystem, simplified integration, and a consistent foundation for next-generation AI factories. This focus on agentic AI reflects a broader industry trend where AI models are evolving to perform more complex, multi-step tasks by interacting with external tools and environments.
New Open Source Software for DSX AI Factory
NVIDIA also launched new open source software components for its DSX AI Factory platform, which unifies computing, networking, power, cooling, facilities, and software to maximize AI output. These components, developed from NVIDIA’s experience running NVIDIA DGX Cloud, include NVIDIA Config Manager for automating switch configurations, NVIDIA Cluster Readiness Engine for stress-testing GPU clusters, NVIDIA Rack Management Service for hardware abstraction, and Cluster API Provider for NVIDIA Infrastructure Controller for accelerating bare-metal deployment.
These open source components enable cloud providers and software partners to integrate AI factory capabilities directly into their multi-tenant platforms, allowing developers to explore and contribute to the DSX AI Factory platform on GitHub. This initiative aims to make the operational aspects of managing large-scale AI infrastructure more accessible and efficient.
Practical Uses and Limitations
The opening of NVIDIA's Vera Rubin architecture and the introduction of new open standards and software components are significant steps towards creating a more standardized and scalable infrastructure for AI development. For developers and AI enthusiasts, this means a more robust and efficient environment for training and deploying advanced AI models, including those designed for agentic AI tasks. The standardized designs and open-source tools could accelerate innovation by reducing the complexity of building and maintaining AI factories.
However, while open standards promote interoperability and efficiency, the actual implementation and adoption across diverse hardware and software ecosystems may still present challenges. The transition to higher-power architectures, such as Visionbay's 800 VDC AI factory design for Vera Rubin, addresses power delivery issues, but large-scale infrastructure changes require significant investment and coordination. The benefits are primarily realized by organizations operating at scale, though the underlying advancements will trickle down to broader AI development through cloud services and specialized platforms.