Cisco and NVIDIA are proposing a new pluggable-optics format designed to push AI-network links from 200Gb/s to 400Gb/s per electrical lane without abandoning the removable transceiver model data centers already know how to operate. The Ultra-Dense Small Form Factor Pluggable, or USFP, was announced at the 2026 Open Compute Project Global Summit.
The first planned design uses 16 full-duplex lanes in roughly an OSFP-sized footprint. At 400Gb/s per lane, that works out to 6.4Tb/s of aggregate capacity per module. Cisco and NVIDIA say the connector can scale to 32 lanes, creating a path to 12.8Tb/s. The companies also claim roughly 4× the bandwidth density and about 1.3× the power efficiency of 200Gb/s-per-lane pluggable optics. Those figures are vendor targets, not independent production benchmarks.
Early industry discussion is focusing on whether USFP can preserve field-replaceable optics while increasing switch-face density.
The real bottleneck is the front panel
AI fabrics do not scale only by making switch silicon faster. Each switch also needs enough physical ports, electrical lanes and optical bandwidth to move traffic between accelerators. Once front-panel density becomes the limit, operators may have to add another switch tier, which adds optics, power draw, latency and cost.
USFP attacks that physical constraint directly. Cisco says the first implementation doubles both the number of lanes and the signaling rate per lane compared with current 1.6T-class pluggables. The new connector uses a two-dimensional contact array and a two-stage insertion mechanism intended to reduce contact wear and preserve signal integrity at 400Gb/s electrical speeds.
Cisco Silicon One architect Martin Lund explains why 1.6T optics, switch density and Ethernet scaling have become central constraints in large AI clusters.
Power is part of the optics battle
The design target is also unusual because USFP supports up to 100 watts of module heat dissipation without requiring a dedicated embedded liquid-cooling loop inside the transceiver. Cisco and NVIDIA argue that simplifying the thermal design could preserve the serviceability advantages of pluggable optics even as per-module bandwidth rises sharply.
NVIDIA’s Spectrum-X Ethernet Photonics platform shows the parallel push to reduce optical power and improve reliability as AI networks move toward gigawatt-scale deployments.
The broader market signal is significant. BitcoinVersus.Tech has already tracked Credo’s 1.6T optical links, Lumentum’s AI-optics capacity being sold out into 2029, and Ciena’s move toward 6.4T co-packaged optics. USFP suggests the industry is still trying to extend pluggable optics much further before operators are forced to move entirely to fixed or co-packaged optical architectures.
The bullish interpretation is that AI-network bandwidth demand is rising fast enough to justify a new physical transceiver standard. The caution is that USFP is still a specification effort: Cisco and NVIDIA have not announced a firm commercial-availability date, and the 4× density and 1.3× efficiency claims still need qualification in shipping systems. TE Connectivity and Amphenol are participating in the hardware ecosystem, and the companies say they plan to develop USFP as an open multi-source specification.
BitcoinVersus.Tech Editor’s Note:
We volunteer daily to ensure the credibility of the information on this platform is Verifiably True. If you would like to support to help further secure the integrity of our research initiatives, please donate using the QR code provided
For best QR-code scanning results, please scan this donation code from a desktop or larger screen. Mobile screens may prevent camera recognition.
Leave a Reply