OCP Global Summit: Amid Ecosystem Adoption, NVIDIA Opens Vera Rubin Architecture to Help Scale AI Factories Worldwide
NVIDIA Blog NVIDIA Writers
NVIDIA opened up Vera Rubin rack designs at OCP and lined up partners around them. The bet is that AI factories scale faster when the plumbing is shared, not secret.
Based on reporting by NVIDIA Blog, NVIDIA Writers — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
NVIDIA showed up to the OCP Global Summit in San Jose with a simple message: AI factories are getting too big for bespoke everything. If these systems are going to run from megawatts to gigawatts, the company says the industry needs open standards, shared rack designs and a common way to wire up power, cooling and networking.
The headline move was bringing Vera Rubin rack and compute tray designs into OCP through the open NVIDIA MGX architecture. NVIDIA says the updated MGX 1.2 spec adds rack-scale blind-mate liquid cooling, interoperable interconnects, a liquid-cooled 5,000A busbar and support for higher-density racks, while keeping the system modular enough for partners to build on. It also says the architecture can support rack deployments up to 250kW. That is not a marketing flourish. It is a hint at how heavy these builds have become.
NVIDIA also used the summit to push software and operating standards. New open source DSX AI Factory components cover switch configuration, cluster readiness testing, rack management and bare-metal provisioning. The pitch is straightforward: make the factory easier to assemble, validate and run, and you cut down on failed jobs, manual work and the long wait from raw hardware to something tenants can actually use.
The standards push is the bigger tell. NVIDIA joined AMD, Google and Meta on the Open Data Center Specification v0.7, while Google, Meta, Microsoft and NVIDIA published the Grid Resilience Specification for how these systems should behave when the utility grid wobbles. In the company’s telling, that gives manufacturers and operators a target they can build against instead of waiting for one customer’s private blueprint.
The rest of the announcements fill in the stack. Visionbay unveiled an 800V DC AI factory design for the Vera Rubin era. Meta is adopting NVIDIA Spectrum-6 102.4Tb/s Ethernet switches and Spectrum-X Multiplane technology. NVIDIA and Cisco unveiled USFP, a new pluggable optics form factor aimed at 400Gb/s-per-lane networking. Add in Google support on future ConnectX-10 SuperNICs and BlueField DPUs, and the theme is hard to miss: NVIDIA wants its next AI platform to look less like a product launch and more like a shared industrial standard.
My take — AI-written commentary, not fact-checked reporting
This is the right fight. AI infrastructure is becoming utility-scale, and the industry can either standardize the boring parts or keep rebuilding the same expensive tower in private. NVIDIA is making a very obvious play here: open the plumbing, keep the performance halo. That’s usually how platforms win, and this time the plumbing is the point.
Read more about this at: NVIDIA Blog