NVIDIA and Cisco propose denser optical connections for larger AI networks
The planned USFP modules would preserve removable optics while increasing bandwidth density. Hardware partners are involved, but an open specification is still to come.
Loading page…
The planned USFP modules would preserve removable optics while increasing bandwidth density. Hardware partners are involved, but an open specification is still to come.
Listen to this story
NVIDIA and Cisco’s proposed USFP optical-module format is intended to increase front-panel network capacity in AI clusters while keeping modules removable and avoiding another switch layer. Its first design targets 12.8 Tb/s in an OSFP-sized footprint, but the performance figures are company expectations: USFP is not yet a completed standard, and no commercial availability date has been announced.
The first planned modules use 16 full-duplex lanes at 400 Gb/s per lane and are intended for switches and SuperNICs.
The companies claim four times the front-panel bandwidth density and about 1.3 times the power efficiency of 200-Gb/s-per-lane pluggable modules.
USFP is designed to dissipate up to 100 watts without a dedicated cooling circuit inside the transceiver; surrounding equipment still requires cooling.
Larger AI clusters would gain more network capacity without adding switch layers under a design NVIDIA and Cisco introduced on October 12, 2026. The companies announced plans for USFP, a denser format for removable optical modules, aimed at moving more data through the same physical footprint while preserving a familiar deployment model.
The collaboration was presented at the Open Compute Project Global Summit. USFP stands for Ultra-Dense Small Form Factor Pluggable: a proposed physical format for optical transceivers, the modules that carry network traffic over optical connections.
When a switch reaches its connection capacity, expanding the network can require additional switches and optical links, increasing latency, power consumption and cost. NVIDIA and Cisco’s proposed remedy is to double both the number of data lanes in each module and the bandwidth carried by each lane.
USFP is designed for electrical signals of 400 gigabits per second per lane. The first planned transceivers would pack 16 full-duplex lanes—carrying data in both directions—into an OSFP-sized footprint. They are intended for switches and SuperNIC network interface cards.
USFP uses a two-dimensional array of contacts and a two-step insertion process. High-speed contacts engage after the module has been inserted. The companies say that sequence reduces contact wear, improves signal integrity and makes insertion force more predictable.
The connector is intended to allow later expansion to 32 full-duplex lanes and beyond. The companies say its 400Gb/s design also improves performance at 200Gb/s per lane.
USFP is designed to dissipate up to 100 watts without a separate, dedicated cooling circuit embedded inside the transceiver. NVIDIA and Cisco say this would reduce switch-design complexity and improve reliability. It does not mean the surrounding equipment needs no cooling.
Cisco’s optical-systems chief Bill Gartner emphasized serviceability and deployment flexibility. The companies are aiming to retain a removable-module operating model as connection speeds rise, rather than asking customers to accept a more complex one.
TE Connectivity and Amphenol are working with NVIDIA and Cisco on USFP cages, connectors and modules. Additional partners are expected to join, giving suppliers a shared physical format for developing higher-density equipment.
An open industry specification is planned through an upcoming community-led, multi-source project. It is not yet a completed standard, and the announcement gives no commercial availability date for the first transceivers.
Loading discussion...
Join the conversation
Explain which priority you would put first, and why.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.