What happened
NVIDIA and Cisco say they are developing Ultra-Dense Small Form Factor Pluggable, or USFP, a 400Gb/s-per-lane optical module for AI-factory Ethernet networks. The first planned modules would use an OSFP-sized footprint with 16 full-duplex lanes, support switches and SuperNICs, and provide up to 12.8Tb/s of aggregate capacity. The companies also plan an open industry specification and are working with TE Connectivity and Amphenol on related hardware.
The proposal describes a pluggable optical form factor using 16 full-duplex lanes at 400Gb/s per lane, with an OSFP-sized footprint and support for switches and SuperNICs. NVIDIA and Cisco say the resulting modules could provide up to 12.8Tb/s of aggregate capacity while they work with TE Connectivity and Amphenol on related hardware and an open industry specification.
Source details: nvidianews.nvidia.com ↗
Why it matters
If the claimed specifications are achieved in deployed systems, USFP could let operators connect more computing capacity within existing network footprints while reducing networking’s share of power consumption. That matters as AI infrastructure expands, because adding bandwidth through extra switch tiers and optical links can increase , power use, cost, and operational complexity. The announcement is a design and ecosystem plan, however, not evidence of tested customer deployments or general availability.
The source says USFP is designed to double both lane count and bandwidth per lane relative to the described 200Gb/s-per-lane modules, enabling more 1.6Tb/s endpoints without adding network tiers. This could improve infrastructure utilization for large AI clusters, but the supplied source does not provide independent benchmarks, deployment results, or customer commitments.
The proposed form factor is relevant to AI infrastructure because it targets the interconnects that link accelerators, switches, and SuperNICs. NVIDIA and Cisco say the design preserves pluggable optics’ serviceability and deployment flexibility while supporting higher bandwidth density. Those are company claims and have not been independently established in the source.
The companies’ plan for a community-led, multi-source specification could reduce dependence on a single supplier if it results in interoperable hardware. The source does not identify a completion date, participating standards body, final specification, or commercial terms.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
The key questions are when USFP hardware will ship, which vendors will support the specification, and whether actual AI-factory operators achieve the claimed power, , density, and serviceability benefits.
The source gives no launch date, customer availability date, pricing, system qualification schedule, or evidence that the first planned transceivers are commercially available. Access for enterprises, neoclouds, and sovereign clouds therefore remains unknown.
Future disclosures should establish whether the stated 12.8Tb/s capacity and approximately 1.3x power-efficiency comparison hold under representative workloads and complete system configurations, rather than only at the component level.
The proposed ecosystem includes TE Connectivity and Amphenol for cages, connectors, and modules, with additional partners expected. The breadth of actual supplier adoption and interoperability will determine whether USFP becomes a practical industry standard or remains a limited collaboration.
The source also claims compatibility work between NVIDIA Spectrum-X Ethernet and Cisco Silicon One-based networks. It does not specify supported products, software versions, certification requirements, or deployment constraints.