Overview of NVIDIA's OCP Contributions

NVIDIA is significantly expanding its contributions to the Open Compute Project (OCP) to standardize next-generation AI data center infrastructure. Announced at the OCP Global Summit in San Jose, the initiatives introduce the Vera Rubin MGX 1.2 specification, integrating 800V DC power architectures and advanced open rack designs. Collaborating with industry giants including Cisco, Google, Meta, AMD, and Microsoft, NVIDIA aims to establish common hardware and software architectures for AI factories scaling toward gigawatt-class campuses. These strategic moves aim to mitigate the escalating power and thermal demands of modern AI factories. A separate Vera CPU rack design is also targeted at CPU-intensive agentic AI operations.

Technical Details: MGX 1.2 and 800V DC Power

At the core of the hardware announcements is the NVIDIA MGX 1.2, an updated rack architecture for Vera Rubin systems. This third-generation design supports rack deployments up to 250 kW, featuring blind-mate liquid cooling, a liquid-cooled 5,000A busbar, modular I/O bays, and fanless liquid-cooled compute trays. The architecture also includes native 19-inch equipment support and standardized interfaces for broader compatibility. To address escalating power demands, Visionbay unveiled the K-1 platform, based on NVIDIA’s 800V DC reference architecture and Vertiv power technology. This design centralizes AC-to-DC conversion at the cluster level, distributing higher-voltage DC closer to computing resources to optimize liquid-cooled, high-density systems.

Networking Advancements: Spectrum-6 and USFP Optics

Networking developments are pivotal for scaling AI fabrics. Meta plans to deploy NVIDIA Spectrum-6 102.4 Tbps Ethernet switches within its liquid-cooled IcePack networking rack, running Meta’s FBOSS software via the OCP Switch Abstraction Interface. Utilizing Spectrum-X Multiplane, Meta can distribute GPU traffic across up to eight network planes, supporting fabrics exceeding 100,000 GPUs. Furthermore, NVIDIA and Cisco proposed the Ultra-Dense Small Form Factor Pluggable (USFP) optics for 400 Gb/s-per-lane networking. Targeting the footprint of conventional 200G-per-lane modules, USFP aims to deliver four times the bandwidth density and 1.3 times the power efficiency, with a roadmap extending to 32 full-duplex lanes. Collaborating with TE Connectivity and Amphenol, the companies are working to overcome traditional pluggable limitations.

Infrastructure Management and Grid Resilience

Beyond hardware, NVIDIA introduced DSX AI Factory, adding open-source infrastructure management components like the Config Manager, Cluster Readiness Engine, Rack Management Service, and Cluster API Provider. These components draw heavily on NVIDIA DGX Cloud operations to significantly reduce deployment and validation effort. Concurrently, addressing facility-level challenges, NVIDIA joined AMD, Google, and Meta in publishing the Open Data Center Specification v0.7. Additionally, a joint effort with Google, Meta, and Microsoft produced the OCP Grid Resilience Specification. This targets automated response and restoration to near-full load in less than two seconds after grid disturbances, ensuring continuous operation for critical AI workloads.

Implications for AI Data Centers

These initiatives mark a critical shift in OCP standardization, moving from individual servers and switches to comprehensive AI factory ecosystems. Facility power and thermal engineering increasingly constrain the useful output of high-density AI systems. By introducing 800V DC power distribution, standardized liquid cooling, and modular rack mechanics, NVIDIA and its partners are addressing the physical limitations of modern data centers. The integration of Falcon Transport for future ConnectX-10 SuperNICs and BlueField DPUs further ensures that network transport can keep pace with compute density through multipath load balancing and hardware loss recovery.

Future Outlook and Industry Impact

While open reference designs provide a crucial blueprint, they do not automatically establish interchangeable products. The announced performance targets for USFP optics and grid resilience require rigorous production validation and multivendor interoperability demonstrations. Ultimately, hyperscale operators will evaluate these architectures based on deployment time, serviceability, energy use, and sustained workload throughput. As AI factories continue to scale toward gigawatt-class capacities, the success of these open standards will dictate the pace of next-generation AI infrastructure deployment across the global technology landscape.