NVIDIA and Cisco have proposed a new pluggable optical-module format designed for the traffic density of large AI clusters. Ultra-Dense Small Form Factor Pluggable, or USFP, is intended to carry electrical signals at 400 gigabits per second per lane and provide a path to 12.8 terabits per second of capacity in a module footprint comparable with today’s OSFP hardware.
The companies announced the design plans on 12 October 2026 at the Open Compute Project Global Summit. “Plans” is the operative word. USFP is not yet a finished open specification, a broadly available product or an independently benchmarked deployment.
AI clusters are running into front-panel limits
Training and serving large models requires accelerators to exchange gradients, parameters and inference data at high speed. Switch capacity can rise while the physical front panel remains fixed. Once a switch runs out of high-bandwidth ports, operators may need another switching tier, adding optics, cost, power consumption and latency.
USFP attempts to increase both the data rate of each electrical lane and the number of lanes in one pluggable module. The first planned modules use 16 full-duplex lanes in an OSFP-sized footprint. Cisco’s technical explanation describes 16 lanes at 400 Gb/s as 6.4 Tb/s in one direction and a future 32-lane path to 12.8 Tb/s. Marketing material also cites 12.8 Tb/s of overall bidirectional capacity for the initial full-duplex arrangement. The distinction matters when comparing module specifications.
A new connector carries the density burden
The design uses a two-dimensional array connector with a staged insertion mechanism. High-speed contacts engage after the module is partly inserted, an approach intended to control insertion force, reduce wear and preserve signal integrity at 400 Gb/s per lane. The format is also designed to dissipate as much as 100 watts without a separate embedded cooling loop inside the transceiver.
NVIDIA and Cisco say USFP can deliver four times the bandwidth density and about 1.3 times the power efficiency of 200-Gb/s-per-lane pluggable optics. Those figures are vendor design targets, not results from production networks. Actual efficiency will depend on optical reach, modulation, digital signal processing, cooling and how much switching hardware a deployment avoids.
Pluggability competes with tighter integration
At extreme bandwidth, some vendors move optics closer to the switch chip through co-packaged optics, shortening high-speed electrical traces and potentially reducing power. Pluggable modules take another route: they remain removable from the switch front panel, preserving familiar replacement and inventory practices. USFP tries to keep that operational model while increasing density.
That serviceability has value because optical modules fail on a different schedule from switching silicon. Yet a 100-watt module also concentrates heat at the front panel, and the new connector means USFP is not simply a faster part for an unchanged socket. Switch, cage, thermal and manufacturing designs must evolve together.
An ecosystem has been announced, not completed
TE Connectivity and Amphenol are working with the two companies on cages, connectors and modules. Cisco and NVIDIA say they plan an open, community-led multi-source specification; Cisco’s broader OCP announcement describes that specification as forthcoming. Interoperability will depend on the eventual document, compliance testing and participation beyond the founding vendors.
USFP addresses a real constraint: moving more data through a limited network faceplate without multiplying switch tiers. Its 400-Gb/s lanes and dense connector offer a plausible route. The harder proof will come when shipping transceivers demonstrate reach, reliability, thermal behaviour and watts per transported bit in working AI fabrics under sustained production loads and across diverse optical reaches.
Featured image: Data-centre network equipment. Contextual image, not USFP hardware. Photo: Taylor Vick/Unsplash.
Leave a Reply