NVIDIA contributed its Vera Rubin rack and compute tray designs to the Open Compute Project (OCP) at the OCP Global Summit in San Jose, anchoring a wave of open-architecture disclosures spanning compute, networking, power and software. The third-generation MGX 1.2 specification supports rack deployments up to 250 kW and introduces a liquid-cooled 5,000 A busbar, configurable I/O bays and fanless liquid-cooled compute trays.

What changes in the MGX 1.2 reference design?

MGX 1.2 extends NVIDIA's open rack architecture with rack-scale blind-mate liquid cooling, interoperable interconnect infrastructure and native 19-inch equipment support. NVIDIA said the update reduces assembly, test and field-service times and supports rack densities up to 250 kW. A separate Vera CPU rack specification extends the same modular chassis principles to tool calling, orchestration and sandboxed execution for agentic workloads.

How open are the new software components?

NVIDIA released four open-source components for the DSX AI Factory platform, all published to GitHub:

• NVIDIA Config Manager: automates switch configuration across AI factory fabrics.
• Cluster Readiness Engine: stress-tests GPU clusters against real workload simulations before training jobs run.
• Rack Management Service: exposes racks as first-class units for firmware operations at scale.
• Cluster API Provider for NVIDIA Infrastructure Controller: cuts bare-metal deployment and tenant onboarding from days to hours.

The components derive from NVIDIA's experience running DGX Cloud, the company said.

What do the new OCP specifications cover?

NVIDIA, AMD, Google and Meta jointly published the Open Data Center Specification v0.7, a unified blueprint covering facility layout, power architecture and cooling. Separately, Google, Meta, Microsoft and NVIDIA published the OCP Grid Resilience Specification, which requires large-scale AI infrastructure to stay connected during grid disturbances and restore to near-full load in under two seconds. NVIDIA said both documents are now available for manufacturers and colocation providers to target without proprietary customer specifications.

How does Meta fit into the networking stack?

Meta confirmed it is adopting NVIDIA Spectrum-6 102.4 Tb/s Ethernet switches to build Vera Rubin AI factories on open networking. Spectrum-6 delivers twice the switching capacity and per-lane bandwidth of Spectrum-4, with 200G-per-lane lossless Ethernet. Meta's Spectrum-6-based switch, N-IceCube, runs FBOSS on the Switch Abstraction Interface and slots into IcePack, Meta's 102.4 T liquid-cooled rack.

Spectrum-X Multiplane technology, enabled by Spectrum-6 and ConnectX-9 SuperNICs, distributes GPU traffic across up to eight independent network planes. NVIDIA said the architecture supports agentic AI factories exceeding 100,000 GPUs and that traffic can shift between planes on component failure.

"We are excited about our close partnership with NVIDIA in making multiplane fabrics operationally scalable and performant," said Omar Baldonado, senior director of data center and AI networking at Meta. "Extending FBOSS to Spectrum ASICs means we can run that same software stack across our fabrics."

What is the USFP optical form factor?

NVIDIA and Cisco disclosed plans for the Ultra-Dense Small Form Factor Pluggable (USFP) module, sized near today's 200 Gb/s-per-lane OSFP but engineered for 400 Gb/s-per-lane fabrics. USFP packs 16 full-duplex lanes with a path to 32, yielding four times the bandwidth density of 200 Gb/s-per-lane pluggables and roughly 1.3x higher power efficiency. TE Connectivity and Amphenol are collaborating on cages, connectors and modules under a community-led, multi-source model.

Why does 800 VDC matter for Vera Rubin?

Visionbay introduced K-1, an AI factory platform built around an 800 VDC architecture and future Vera Rubin deployments. The design centralizes AC-to-DC conversion at the cluster level and phases 800 VDC distribution into existing facilities, combining NVIDIA's 800 VDC reference architecture with power and thermal systems from Vertiv. NVIDIA and Google separately confirmed Falcon Transport support on future ConnectX-10 SuperNICs and BlueField DPUs, adding multipath load balancing, programmable congestion control and hardware loss recovery to Ethernet AI fabrics.