NVIDIA and Cisco Plan USFP Format for Denser AI Ethernet Networks

NVIDIA and Cisco presented plans for the USFP format for pluggable optical modules in Ethernet networks designed for AI infrastructure. The specification is not yet complete or commercially available.

NVIDIA Cisco USFP is a planned new format for pluggable optical transceivers for Ethernet networks in AI data centers. The companies announced it on October 12, 2026, as a proposed Ultra-Dense Small Form Factor Pluggable format intended to combine higher interconnect density with the serviceability of pluggable optical modules.

According to both companies, the format is intended to support electrical lanes at 400 Gb/s per lane. NVIDIA and Cisco are working with TE Connectivity and Amphenol on connectors, cages, and modules. The goal is then to create an open industry specification through a future community multi-source project.

USFP Is Intended to Remain Pluggable

USFP is proposed as a pluggable solution, meaning its optical modules can be replaced in the device. This approach differs from architectures in which the optical interconnect is more tightly integrated directly into the network hardware.

The announcement focuses on Ethernet AI networks. In large GPU clusters, the network layer is becoming more important because interconnects carry communication between compute nodes and switches. If adopted, the new format could affect the design of future switches, network cards, and optical component offerings.

In its accompanying materials, Cisco also states a goal of supporting systems with 200G and 400G SerDes. However, the format’s basic publicly communicated parameter is an electrical lane operating at 400 Gb/s.

NVIDIA Cisco USFP: Stated Capacity Is Not Yet Final

The companies’ materials refer to a maximum aggregate capacity of up to 12.8 Tb/s. However, the detailed description of the number of lanes and resulting capacity differs between the accompanying materials. The exact final USFP configuration has therefore not yet been finalized.

The manufacturers also claim the possibility of four times the throughput density and approximately 1.3 times the energy efficiency. These are company projections, not independently verified results. No interoperability tests or independent measurements of power consumption and reliability have been published so far.

Specification and Availability Are Still to Come

NVIDIA and Cisco have not announced a USFP availability date. It is therefore not a finished standard or commercially supplied product, but a plan whose next step is intended to be an open specification and a multi-source agreement involving the broader community.

For the market, it will be important whether additional manufacturers of optics, connectors, switches, and network cards join the project. The timing of the first samples, interoperability conditions, and the actual technical parameters of the final specification will also be important.

If the format gains traction, it could offer data centers a way to increase the density of Ethernet interconnects for AI infrastructure without an immediate transition to other optical architectures. For now, however, it remains a stated proposal and a set of declared goals from its authors.

Sources

  • NVIDIA Newsroom – Confirms the joint NVIDIA and Cisco announcement, the planned USFP specification, and the forward-looking nature of the announcement.
  • Cisco Newsroom – Lists USFP’s technical goals, collaboration with TE Connectivity and Amphenol, and the planned community multi-source project.
  • Cisco Blogs – Clarifies that this is a proposed plan and describes the stated goal of supporting both 200G and 400G SerDes systems.

Verified and updated: October 13, 2026 06:27

Sharing