NVIDIA and Cisco are trying to solve one of the least glamorous but most binding limits in AI data centers: the front panel of the switch. At the Open Compute Project Global Summit in Santa Clara on October 12, the two companies announced plans for Ultra-Dense Small Form Factor Pluggable, or USFP, a new optical module designed for 400 gigabits per second per lane.
The problem USFP targets is straightforward. As AI models grow and inference demand rises, operators need more bandwidth per connection and more connections per switch. When a switch runs out of capacity, the usual answer is another network tier — more switches, more optical links, and with them more latency, power draw, and cost. Denser optics let builders connect more systems inside the same network footprint instead.
On the specifications the companies published, the first planned USFP transceivers will pack 16 full-duplex lanes into a footprint about the size of today's OSFP module, serving both switches and SuperNICs. The connector is designed to scale to 32 lanes and beyond, for up to 12.8 terabits per second of overall capacity. Compared with 200-gigabit-per-lane pluggables, NVIDIA and Cisco claim four times the bandwidth and front-panel density at roughly 1.3 times the power efficiency — in other words, more bits moved for less power per bit.
Getting to 400 gigabits per lane required more than shrinking an existing module. USFP uses a new two-dimensional array connector with a two-step insertion design, in which high-speed contacts engage only after the module is seated. The companies say that reduces contact wear, improves signal integrity, and makes insertion force more predictable. The design also supports power dissipation of up to 100 watts without a dedicated cooling circuit embedded in the transceiver, which is meant to simplify switch design and improve reliability. All USFP configurations are intended to support optical designs that simplify signal processing to cut latency and power.
Just as notable is what USFP is not: it is not a move away from pluggables. Even as parts of the industry experiment with co-packaged optics, NVIDIA and Cisco are explicitly preserving the field-replaceable model operators know — modules that can be swapped, serviced, and sourced from multiple vendors. The two companies are working with TE Connectivity and Amphenol on cages, connectors, and modules, and they plan to turn USFP into an open industry specification through a community-led multi-source project, with more partners expected to join. The effort also builds on their existing work pairing NVIDIA Spectrum-X Ethernet with Cisco Silicon One-based systems for enterprises, neoclouds, and sovereign clouds.
For enterprise and cloud network teams, the practical takeaway is timing and optionality, not an immediate purchase. This is an announced plan, not a shipping product, and the announcement gave no sampling or availability date. What it signals is where high-end Ethernet AI fabrics are headed: 400-gigabit lanes, much denser switch faces, and an open, multi-source supply chain if the specification gains traction. Operators planning the next refresh cycle should watch for the draft specification and first vendor roadmaps before assuming USFP fits their timelines.