NVIDIA has unveiled a series of open hardware designs, software tools and industry standards created to support the massive computing and power requirements of next-generation AI data centers.

Unveiled at the Open Compute Project (OCP) Global Summit, the initiative includes contributions to OCP from NVIDIA’s Vera Rubin architecture, along with collaborations with Google, Meta, Microsoft and Cisco.

The wide-ranging announcements support NVIDIA’s efforts to expand its role beyond supplying GPUs to providing the infrastructure required to run large AI data centers. Additionally, the company’s push to standardize the infrastructure surrounding its chips could give equipment suppliers and data center operators more flexibility while reducing integration work.

Vera Rubin Rack

A key announcement is NVIDIA’s contribution of Vera Rubin rack and compute tray designs to OCP through its modular MGX architecture. The updated MGX 1.2 specification supports rack deployments of up to 250 kilowatts, incorporating liquid cooling and a liquid-cooled 5,000-amp busbar to distribute electricity.

The architecture also supports standard 19-inch equipment, configurable input/output bays and liquid-cooled compute trays without fans. These designs are intended to allow hardware manufacturers to build compatible systems while reducing the need for customized infrastructure.

NVIDIA is extending this architecture across two distinct computing platforms. Its Vera Rubin NVL72 is designed for large-scale AI training and inference, while the Vera CPU rack supports agentic AI workloads, including software orchestration and interactions with external applications.

This matters because AI agents require computing capabilities tailored to tasks like coordinating workflows and interacting with external applications. By supporting both workloads within a common hardware architecture, NVIDIA is positioning MGX as a platform for multiple types of AI infrastructure.

However, the company has not provided independently verified figures showing how much the new architecture reduces deployment costs or improves operating efficiency. Those results will depend on industry adoption and real-world implementations.

Open Standards for Power and Cooling

NVIDIA also collaborated with AMD, Google and Meta on Open Data Center Specification v0.7, which establishes common technical specs for facility design, including cooling systems, electrical infrastructure and physical layouts.

Separately, NVIDIA joined Google, Microsoft and Meta to develop the OCP Grid Resilience Specification. The standard requires AI facilities to respond automatically to electrical grid disruptions and recover to nearly full operating load in less than two seconds.

This rapid recovery is a design requirement, not a demonstrated performance result. Still, the spec addresses the effect of large AI facilities on electrical grid stability, a growing issue as facilities need more power.

NVIDIA also expanded its DSX AI Factory platform with four open-source software components. These include Config Manager for network configuration, Cluster Readiness Engine for testing GPU systems, Rack Management Service for hardware administration and Cluster API Provider for provisioning physical servers.

NVIDIA claims the provisioning software can prepare production clusters in hours rather than days. The tools draw on the company’s experience running DGX Cloud and are built for integration into cloud providers’ and software partners’ platforms.

In networking news, Meta plans to adopt NVIDIA Spectrum-6 networking, while NVIDIA and Google are collaborating on Falcon Transport support for future ConnectX-10 SuperNICs and BlueField processors.

NVIDIA and Cisco introduced plans for Ultra-Dense Small Form Factor Pluggable (USFP), an optical networking module supporting 400 gigabits per second per lane. The companies say USFP could provide up to 12.8 terabits per second of capacity, with four times the bandwidth density and approximately 1.3 times the power efficiency of existing 200Gb/s-per-lane modules. But the technology remains under development.

“The blocking limit in AI networking has moved to the front panel of the switch. NVIDIA and Cisco are betting pluggable optics can hit 400Gb/s per lane and still let operators swap a failed module in minutes,” said Mitch Ashley, VP and Practice Lead, at The Futurum Group.

“For buyers, the open spec and promised Spectrum-X to Silicon One interoperability is even more important than the speeds. A network fabric is a multi-year commitment. Watch whether independent module makers build to the spec. That decides whether USFP widens choice or formalizes an NVIDIA-Cisco lane.”