Nvidia contributes key Vera Rubin rack designs to the Open Compute Project: what is in them

Nvidia said on Monday 12 October, at the Open Compute Project (OCP) Global Summit in San Jose, that it had contributed "key NVIDIA Vera Rubin rack and compute tray designs to OCP through the open NVIDIA MGX architecture". Nvidia said that at the summit it joined AMD and several hyperscalers, the companies that run the largest data centres, in publishing two data-centre standards, and Nvidia and Cisco announced plans for a new optical module. This morning's Daily Update carried a short item; this piece looks at what is in the documents.

What the Open Compute Project is

OCP says it was initiated in 2011 by Facebook, now Meta, after the company designed its own data centre in Prineville, Oregon, from the ground up. By OCP's account that building was "38% more energy efficient to build and 24% less expensive to run" than Facebook's earlier sites. The foundation's stated mission is to apply "the benefits of open source and open collaboration to hardware": companies publish designs for servers, racks, power and cooling so that any manufacturer can build to them.

That matters for cost and supply. When a design is open, a data-centre owner can buy from several suppliers rather than one, and suppliers can build stock to a known target. OCP said on 12 October that a new IDC impact study forecasts that, for 2030, "over $890B will change hands between buyer and seller on IT equipment that is based upon OCP recognized designs", up from a forecast of $295B for 2029 a year ago.

What Nvidia contributed

Nvidia says it joined other companies in contributing to the third-generation MGX reference design, through OCP's Open DC Initiative. According to Nvidia's blog, the updated MGX 1.2 specification:

• supports rack deployments "up to 250kW";
• adds rack-scale blind-mate liquid cooling and "a liquid-cooled 5,000A busbar designed for higher-power AI systems";
• adds "fanless liquid-cooled compute trays", configurable I/O bays and native 19-inch equipment support.

Nvidia also described a Vera CPU rack design specification on the same MGX base, which it says is designed for "tool calling, orchestration and sandboxed execution" for AI agents.

The two hyperscaler standards

Open Data Center Specification v0.7. Nvidia says it joined AMD, Google and Meta to publish it. The 18-page document, whose cover dates it as effective from 27 August, sets design targets for AI data halls on a 2027–2028 timeline: racks 100 inches tall weighing up to 8,000lb (3,629kg), an 18ft minimum ceiling for two-busway layouts, and a cluster size of "12 MW Minimum, 24 MW preferred". Earlier versions were Google's alone. The authors say one aim of opening it is "to collect end-to-end feedback".

Grid resilience. The OCP specification published by Google, Meta, Microsoft and Nvidia sets out how large AI sites should behave when the electricity grid has a fault. It cites a North American event in August 2024 that caused "a simultaneous load loss of 1,500 MW", and a larger one in July 2026. The core rule: during a voltage dip the computing load stays on the grid rather than switching to backup for set times, then must "restore to 90% of pre-fault load within 2 seconds". It covers the computing load only, which the document puts at "~70-90% of total data center load", and adopts the convention proposed by the US grid operator MISO while closely mirroring a proposal from ERCOT.

What is still a plan

Several announcements describe future products rather than finished ones:

• USFP optics. Nvidia and Cisco "announced plans for" a pluggable optical module running at 400Gb/s per lane, which they say is designed to be "the first module to support 400Gb/s-per-lane electrical signals". They claim it gives 4x the bandwidth density and about 1.3x the power efficiency of 200Gb/s-per-lane optics. TE Connectivity and Amphenol are working on cages, connectors and modules, "with additional partners expected to join", and the two companies "plan to create an open industry specification" through "an upcoming community-led multi-source project".
• Falcon. Nvidia and Google announced Falcon Transport support on "future" ConnectX-10 network cards and BlueField processors.
• Visionbay's K-1. Visionbay "unveiled plans" for an 800-volt DC AI factory design built with "future NVIDIA Vera Rubin deployments in mind".

Nvidia also said Meta is adopting its Spectrum-6 Ethernet switches to build larger Vera Rubin AI factories.

For context, we have previously reported that Korea's planned national AI factory is expected to use Vera Rubin and Blackwell systems, and that, according to IEEE Spectrum, HBM4 memory will be used in the Vera Rubin GPU.

Why it matters

Nvidia's argument is that with these standards open, "manufacturers around the world have a common target to build for", and can start "without waiting for proprietary specifications from individual customers". The grid document makes a similar case for utilities: one baseline instead of "redundant, site-by-site engineering analysis".

On our reading, there are two sides to this. Open designs can widen the supplier base and shorten build times for anyone building AI data centres. But the rack designs are built on Nvidia's own MGX architecture, so the open standard and Nvidia's commercial interest point the same way. The optics and Falcon items are announcements of intent; how open the USFP specification turns out to be depends on a multi-source project the two companies describe as upcoming.