NVIDIA expands open architecture push at OCP Global Summit to scale AI factories

NVIDIA used the Open Compute Project’s Global Summit in San Jose to argue that the next phase of AI growth will depend on more than faster chips. As “AI factories” scale from megawatts to gigawatts, the company said, builders need open standards that reduce integration friction across the entire stack, from the facility design to power, cooling, rack internals and software operations.

At the center of NVIDIA’s pitch was its Vera Rubin rack and compute tray design work, delivered through OCP’s open MGX approach. NVIDIA said it contributed key Vera Rubin rack and compute tray designs to OCP via the open NVIDIA MGX architecture, positioning the effort as a common foundation for next-generation AI factory infrastructure that multiple system builders can adopt.

NVIDIA described an updated MGX 1.2 specification for Vera Rubin AI factories that extends the open architecture with rack-scale blind-mate liquid cooling and interoperable rack-scale interconnect infrastructure. The company said MGX 1.2 also includes a liquid-cooled 5,000A busbar intended for higher-power AI systems, adds native 19-inch equipment support, and enables higher-density rack designs.

Beyond cooling and power distribution, NVIDIA said MGX 1.2 expands modularity with configurable I/O bays, fanless liquid-cooled compute trays, standardized power and cooling interfaces, and a common infrastructure that is meant to help ecosystem partners build interoperable accelerated-computing platforms. NVIDIA also tied the specification to operational goals, including simplifying assembly, test and service to streamline manufacturing and field support, while supporting rack deployments up to 250kW.

NVIDIA also addressed the power and facility sides of scaling AI factories through standards and reference implementations. The company said it joined industry leaders in contributing to the third-generation MGX AI data center reference design through OCP’s Open DC Initiative, and it linked the open rack work to liquid-cooling technologies described as MQD and MQDB.

In software, NVIDIA said its DSX AI factory platform uses “extreme codesign” to unify computing, networking, power, cooling, facilities and software so the facility can operate as a coordinated, optimized system. At OCP, NVIDIA said it expanded DSX OS, its open operating layer, with open-source components for configuring networks, validating clusters, managing racks and provisioning bare metal, drawing on experience from operating NVIDIA DGX Cloud. NVIDIA also said developers can visit a DSX AI Factory GitHub page to explore and contribute to the new and existing open software components.

NVIDIA’s architecture pitch leaned heavily on open specifications that define how the broader infrastructure should be built. The company said that, at OCP, NVIDIA joined partners AMD, Google and Meta to publish the Open Data Center Specification v0.7, described as a unified blueprint for next-generation AI factory facility design spanning power architecture, cooling infrastructure and physical layout. NVIDIA also cited a separate OCP grid resilience specification published by Google, Meta, Microsoft and NVIDIA, which it described as aiming for large-scale AI infrastructure to stay connected, respond automatically and restore to near-full load in less than two seconds during utility grid disturbances.

Several partner announcements at the summit were framed as evidence that open interfaces can support both performance and scale. Meta said it is adopting NVIDIA Spectrum-6 102.4Tb/s Ethernet switches to build larger, higher-performance Vera Rubin AI factories on open networking. In the announcement, Meta described using Spectrum-6 in its N-IceCube switch that slots into IcePack, Meta’s liquid-cooled rack, and running FBOSS, Meta’s open source network operating system. NVIDIA said Spectrum-6, paired with Spectrum-X Multiplane technology enabled by Spectrum-6 and ConnectX-9 SuperNICs, is designed to distribute GPU traffic across up to eight independent network planes to scale beyond 100,000 GPUs with fewer switch tiers and lower infrastructure costs.

Separately, NVIDIA and Cisco said they plan USFP, a new optical pluggable transceiver module form factor intended to help AI factory builders increase bandwidth density and power efficiency. NVIDIA said USFP is designed to fit a similar footprint to the widely adopted 200Gb/s-per-lane OSFP form factor, with a path to 400Gb/s-per-lane networking, and it described configurations that pack 16 full-duplex lanes with plans to scale to 32 lanes to raise bandwidth density compared with traditional 200Gb/s-per-lane pluggable optics.

It is not entirely clear from NVIDIA’s OCP update how quickly MGX 1.2 and the related open software components will convert into widely available, multi-vendor systems, or what specific performance deltas customers should expect versus prior generations. The company also did not provide quantified cost or time-to-deploy outcomes in this announcement, instead focusing on standards, interoperability, and design goals for manufacturing and service. For investors and industry operators, the near-term watch is whether OCP standardization leads to broader adoption by data center operators, equipment makers and cloud providers building Vera Rubin-class infrastructure.

In practical terms, NVIDIA’s message is that “AI factories” are becoming system integration projects, not just hardware purchases. If open facility and rack blueprints gain traction, the industry could see faster replication of proven designs across geographies, with less reliance on bespoke customer-specific specifications. The OCP summit moment also suggests NVIDIA is positioning itself as a standards contributor as much as a chip supplier, betting that the business of AI infrastructure will increasingly reward compatibility and repeatability.