NVIDIA and Cisco announced plans on 12 October for a new optical transceiver format called Ultra-Dense Small Form Factor Pluggable (USFP), unveiled at the OCP Global Summit in San Francisco. The module is designed to break the networking bottleneck that limits how many GPUs can work together inside an AI data centre.
The USFP format supports 400 gigabits per second per lane and can carry up to 12.8 terabits per second of aggregate bidirectional capacity. According to the companies' joint announcement, that is four times the bandwidth density and roughly 1.3 times the power efficiency of the 200 Gb/s per lane optical solutions in widespread use today. Crucially, the design keeps transceivers removable and replaceable, preserving the operational model that data centre teams already know.
The urgency behind the announcement stems from a structural shift in AI workloads. Cisco President and Chief Product Officer Jeetu Patel has said publicly that AI agents consume roughly five times more network bandwidth than a human performing the same task. With agent token consumption already surpassing human consumption, that gap is expected to widen sharply as agentic AI moves from experiment to production at scale.
When a switch reaches its capacity limit, operators currently have to add extra switch layers and optical links, which raises latency, power consumption and cost. USFP is intended to let builders scale their compute without adding those additional network tiers, keeping infrastructure costs and complexity in check.
NVIDIA and Cisco said they are working with hardware partners TE Connectivity and Amphenol to build a multi-source ecosystem around the format. An open industry specification is also planned, though no commercial availability date has been announced and the standard is not yet finalised.
The announcement extends a broader partnership between the two companies. In August, Cisco expanded its Secure AI Factory with NVIDIA programme to include rack-scale, liquid-cooled compute systems built with Supermicro, targeting enterprise, neocloud and sovereign cloud customers.
Source : NVIDIA and Cisco joint press release
