USFP: NVIDIA and Cisco announce a 400G-per-lane optics module for AI factories

NVIDIA and Cisco announced plans for USFP, Ultra-Dense Small Form Factor Pluggable, a new optical transceiver module designed for 400 Gb/s per lane. It is the first pluggable module aimed at 400G-per-lane electrical signals, with a stated goal of 12.8 Tb/s of total capacity per module. In an OSFP-sized footprint, both companies claim 4x the bandwidth density and roughly 1.3x the power efficiency of current 200 Gb/s-per-lane optics.

Why the network is the part that fills up first

The problem the announcement targets is a familiar one in AI factory design. Models grow, inference demand grows with them, and the fabric has to move more data between accelerators. When a switch runs out of capacity, expanding further usually means adding switch tiers and more optical links, which adds latency, power draw and cost. The network becomes the layer that fills up before the compute does.

USFP's approach is to make each physical connection carry more, so the same switch fabric reaches more accelerators. The first transceivers pack 16 full-duplex lanes in an OSFP-sized footprint and fit both switches and SuperNICs, with the connector designed to scale to 32 lanes and beyond. Doubling both the lane count and the bandwidth per lane is how the companies say you connect more 1.6 Tb/s endpoints without adding network tiers.

There are two engineering details worth noting. The module uses a new two-dimensional array connector with a two-step insertion, which the companies say reduces contact wear and makes insertion force more predictable. And USFP is designed to dissipate up to 100 watts without a dedicated embedded cooling circuit in the transceiver itself, which simplifies switch design. No third party has validated either number yet; they come from the press release.

An ecosystem play, not a product drop

What ships is a specification, not a product. TE Connectivity and Amphenol are the first partners building USFP cages, connectors and modules, with more expected to join, and the two companies plan to publish an open industry specification through a community-led multi-source project. The work builds on their existing collaboration to make NVIDIA Spectrum-X Ethernet and Cisco Silicon One Ethernet networks interoperate, which is the practical reason this announcement carries weight: it comes from the two vendors whose Ethernet fabrics most AI factories already standardize on.

NVIDIA SVP Gilad Shainer framed it as reinventing pluggable optics for the largest AI factories. Cisco SVP Bill Gartner emphasized that the goal was scaling to 400G-per-lane without forcing operators into a more complex operating model, keeping pluggable serviceability and deployment flexibility.

What to take from it

Three things. First, direction: 400G-per-lane Ethernet is now a stated target from the two main Ethernet fabric vendors, which tells you where AI networking standards are heading over the next generation. Second, power: the pitch is explicitly about shrinking networking's share of the data center power budget so the same footprint carries more compute. If the 1.3x efficiency figure holds in deployed systems, that is real headroom for operators. Third, timing: this is a form factor and an ecosystem. No product, no pricing, no ship date in the announcement. If you are planning fabric upgrades on a cycle longer than a year, it is a design point worth tracking. If you rent your compute, it changes nothing in your bill for the time being.