Every switch has a ceiling, and AI builders keep hitting it. Nvidia and Cisco used this month’s OCP summit to float a module format meant to push that ceiling back: Ultra-Dense Small Form Factor Pluggable, or USFP, which is specified for 400Gb/s on each lane instead of 200.
The first parts in the plan put 16 full-duplex lanes into a body the size of today’s octal module, roughly doubling lane count and lane speed at once. Nvidia says that adds up to 4x the bandwidth and front-panel density of a comparable existing transceiver and about 1.3x better power efficiency, letting one switch reach 12.8Tb/s of aggregate capacity. Future versions are meant to reach 32 lanes or more.
The hardware change is the connector. USFP uses a two-dimensional array with a staged insertion sequence, so high-speed contacts close only after the module is seated. That trims wear, cleans up signal integrity and makes insertion force predictable enough to scale.
The economics are the pitch. When a switch tops out, operators add another layer of switches and optical links, and pay for it in latency and power. Nvidia network chief Gilad Shainer frames USFP as work on the smallest parts of the largest systems.
Cisco’s Bill Gartner says incremental tweaks to today’s modules were not enough to double lane speed.
