Nvidia and Cisco have unveiled a new optical networking interface aimed at moving data around AI data centres faster and more efficiently. The proposed USFP standard targets up to 12.8Tb/s of total bandwidth, a substantial ambition for infrastructure that can otherwise leave costly accelerators waiting on data.
Cisco Watch analysis
What happened
TechSpot reports that USFP is designed to support 400Gb/s per lane, with an initial version using 16 full-duplex lanes and connecting to both switches and Nvidia SuperNIC network adapters. The companies say the interface offers four times the capacity of optical interfaces using 200Gb/s per lane, alongside a 1.3-fold improvement in power efficiency.
The full 12.8Tb/s capacity is a target, not a finished specification already in customers’ racks. The report says the design is expected to expand to 32 or more lanes in a later phase. Nvidia and Cisco are working with component partners TE Connectivity and Amphenol; the system is part of a wider mix of optical approaches, rather than a replacement for every other design.
Why it matters
AI workloads move enormous volumes of data between processors. Faster links could help data-centre operators connect more capacity without adding network layers, while improved power efficiency matters in facilities where both electricity and cooling are becoming stubborn line items.
But the useful numbers are still design claims and future targets. The next test is whether the promised bandwidth, efficiency and deployment model survive the trip from a standard on paper to working equipment at scale.
Our read
This is a meaningful infrastructure development, not just a shinier plug: the aim is to ease a real networking constraint in AI systems. The ambition is impressive; the sensible response is to keep one eye on measured performance and availability before treating the headline figures as delivered capacity.
What to watch
- When the first USFP transceivers and compatible equipment become available.
- Whether the planned move to 32 or more lanes delivers the targeted 12.8Tb/s.
- How measured power efficiency compares with existing optical interfaces.
Discussion spark: Would you prioritise higher bandwidth or lower power use when choosing the next generation of data-centre networking gear?
Sources and evidence
- Source update (15 October 2026, 19:08 UTC)
not affiliated with or endorsed by Cisco