NVIDIA opens Vera Rubin designs and plans denser AI optics
At the Open Compute Project Global Summit on October 12, 2026, NVIDIA said it had contributed Vera Rubin rack and compute tray designs through its MGX architecture. It also outlined work with Cisco on a new pluggable optical format.
Shared rack designs reach OCP
The company’s OCP update describes MGX 1.2 as supporting racks up to 250 kilowatts. Changes include standardized power and cooling interfaces, configurable input and output bays, fanless compute trays and a liquid cooled 5,000 amp busbar. These are design specifications, rather than measurements of a particular operator’s installed fleet.
For system builders, common interfaces offer a clearer target for integrating components. Deployment and maintenance benefits will still depend on the products built around them.
USFP raises the connection ceiling
Cisco’s technical explanation makes an important capacity distinction. The initial USFP design has 16 lanes operating at 400 gigabits per second, giving 6.4 terabits per second of aggregate capacity. Reaching 12.8 terabits requires a future 32 lane configuration. The larger figure should not be read as the initial module’s capacity.
The claimed fourfold density increase compares that 6.4 terabit design with today’s 1.6 terabit pluggables. Cisco also describes an opportunity for roughly 1.3 times greater power efficiency. These are vendor projections for the design, not independently measured application throughput or a guarantee of lower power consumption across an entire data center.
Cisco says USFP is intended to support systems using both 200G and 400G serializer and deserializer technology. Preserving modules that technicians can replace independently is part of its argument for extending pluggable optics as other optical architectures develop.
A new connector and a future specification
The joint announcement describes USFP as Ultra Dense Small Form Factor Pluggable. It uses an OSFP sized footprint with a new array connector whose contacts engage after insertion to reduce wear. A stated ability to dissipate up to 100 watts without a dedicated cooling circuit inside the transceiver is a thermal design claim. It does not remove the need to cool the surrounding system.
NVIDIA and Cisco are working with TE Connectivity and Amphenol on cages, connectors and modules. They plan an open specification through a future community project. The announcement gives no USFP shipping date, price or customer deployment results.
Products follow different timelines
Cisco’s separate OCP release illustrates why those distinctions matter. It says its 102.4 terabit Silicon One G300 is available for customer sampling and that the first systems using it are shipping. The newly announced G302 targets switches at the top of server racks, with 51.2 terabits of capacity and shipments planned before the end of 2026.
Those are separate switching product timelines. The G300 systems shipping now and the planned G302 shipments cannot be used to date USFP availability. Cisco’s release explicitly allows for changes to delivery schedules for products still in development.
Our coverage of LG and AIR’s cooling agreement separates planned data center capacity from completed installations.
