NVIDIA and Cisco unveil USFP optics for AI factories

NVIDIA and Cisco are pitching USFP as a way to keep pluggable optics viable as AI factories grow. Today's 200Gb/s-per-lane modules force operators to add switch layers and optical links when a switch fills up, which raises latency, power consumption and costs. USFP is designed to double both the lane count and the bandwidth per lane, so AI fabrics can connect more 1.6Tb/s endpoints without adding network tiers. The first planned transceivers pack 16 full-duplex lanes in an OSFP-sized footprint and can expand to 32 or more lanes later. A new two-dimensional array connector engages high-speed contacts only after insertion, reducing wear and making insertion force more predictable. USFP also supports up to 100 watts of power dissipation without a dedicated embedded cooling circuit inside the transceiver, which is intended to simplify switch design and improve reliability. Beyond the technical specs, NVIDIA and Cisco are trying to build a multi-source hardware ecosystem with TE Connectivity and Amphenol, and they plan to create an open industry specification through an upcoming community-led project. That approach could make it easier for suppliers to develop compatible high-density pluggable infrastructure for 400Gb/s-per-lane fabrics, though the plan depends on additional partners joining.