The Pulse
NVIDIA Opens Vera Rubin Designs Through OCP
At the October 12, 2026, OCP Global Summit, NVIDIA contributed Vera Rubin rack and compute-tray designs to an open industry architecture. The company also announced open-source data-center software and networking work with Meta, Cisco and G
AI.info Team · Written with AI ·
More than 100,000 GPUs are the scale NVIDIA and Meta say their networking work is designed to support. At the October 12, 2026, OCP Global Summit, NVIDIA contributed Vera Rubin rack and compute-tray designs to the Open Compute Project, while announcing open-source data-center software and new networking collaborations with Meta, Cisco and Google. The announcements extend the company’s hardware platform into shared designs and operating tools that other infrastructure builders can use.
MGX designs reach beyond NVIDIA systems
NVIDIA says its Vera Rubin rack and compute-tray designs enter OCP through MGX, its modular reference architecture. The third-generation MGX specification adds rack-scale blind-mate liquid cooling, interoperable rack interconnects and a liquid-cooled 5,000-amp busbar. It also supports racks up to 250 kilowatts, along with configurable I/O bays and fanless liquid-cooled compute trays.
The contribution is more specific than opening every detail of Vera Rubin. NVIDIA is sharing design elements and rack specifications through an OCP framework, giving system builders a common basis for developing compatible infrastructure. The company also described a Vera CPU rack specification for agent workloads such as tool calling, orchestration and sandboxed execution, alongside the Vera Rubin NVL72 system for large-scale training and inference.
The open approach targets the physical work of building and servicing high-density systems, not just the chips inside them. NVIDIA says standardized power and cooling interfaces, plus modular rack designs, can simplify integration, manufacturing and field service. The company describes the resulting architecture as supporting faster deployment and reuse of infrastructure.
Open software targets setup and cluster checks
NVIDIA also released open-source components for its DSX AI Factory platform, which coordinates computing, networking, power, cooling and software. The tools cover switch configuration, GPU-cluster readiness testing, rack management and bare-metal provisioning. NVIDIA says the components draw on its experience running DGX Cloud and are intended for cloud providers and software partners building multi-tenant platforms.
The readiness tool stress-tests clusters with workload simulations before training begins; the provisioning component is intended to prepare bare-metal systems for production use. Those tasks address the gap between acquiring servers and operating them as a coordinated service. NVIDIA has made the components available through GitHub, according to its announcement.
Meta’s network plans target 100,000-GPU systems
Meta plans to use NVIDIA Spectrum-6 Ethernet switches in Vera Rubin AI factories, the companies said. Spectrum-6 offers 102.4 terabits per second of switching capacity, and Meta’s N-IceCube switch is designed for the company’s liquid-cooled IcePack rack. The companies say their Spectrum-X Multiplane networking can distribute GPU traffic across as many as eight network planes, allowing systems to scale beyond 100,000 GPUs with fewer switch tiers.
Meta is also extending its FBOSS open-source network operating system to Spectrum ASICs. “We are excited about our close partnership with NVIDIA in making multiplane fabrics operationally scalable and performant, a critical requirement both for our own infrastructure builds and for our joint cloud partners,” said Omar Baldonado, senior director of data center and AI networking at Meta.
Power standards and optics fill out the announcement
The companies also highlighted shared standards for data-center design and power-grid response. NVIDIA’s post describes the Open Data Center Specification as a joint effort involving AMD, Google, Meta and NVIDIA, covering facility design from power architecture to cooling and layout.
NVIDIA and Cisco announced USFP, a proposed pluggable optical transceiver format designed for 400-gigabit-per-second lanes. The companies say its 16 full-duplex lanes can scale to 32, with four times the bandwidth density of traditional 200-gigabit-per-second-per-lane pluggable optics and roughly 1.3 times greater power efficiency. They plan to work with TE Connectivity and Amphenol on components and advance the format through a multi-source effort.
Google and NVIDIA also announced support for Google’s Falcon Transport protocol on future ConnectX-10 SuperNICs and BlueField DPUs. The companies describe Falcon as a way to manage congestion, recover dropped data and distribute traffic across Ethernet paths. Together, the announcements put rack designs, cluster software, power standards and network components into the OCP discussion; the open-source software components are available on GitHub.