At the OCP Global Summit, NVIDIA and Cisco announced their plans for the Ultra High Density Small Form Factor Pluggable (USFP) optical module, a form factor specifically designed for 400Gb/s per-lane Ethernet in AI architectures.
▌What Makes USFP's "Pluggability" Different?
Recently, many products have highlighted "pluggability" as a major selling point, with Avicena also promoting its Micro LED pluggable optical interconnects. However, current products marketed for their pluggability actually fall into two categories: front panel pluggable for switches, and near-chip/on-board pluggable inside the chassis/tray. The replaceable components vary, which could be the module itself, an on-board optical engine, or a fiber optic bundle.
OSFP, QSFP-DD, XPO, and the newly introduced USFP are all front panel pluggable. OSFP and QSFP-DD are traditional standard cage modules, while USFP and XPO utilize dedicated panel cages. In contrast, NPO and Avicena Micro LED LightBundle are near-chip/tray-internal pluggable.
NPO optical engines can be deployed on switch or GPU carrier boards, in close proximity to the ASIC. Avicena's pluggable interface is located near the chip inside the GPU tray, where only the fiber bundle ferrules are plugged and unplugged.
Based on this classification, looking at the USFP introduced by NVIDIA and Cisco, and comparing it with NPO and XPO, which also emphasize pluggability, what differences has USFP achieved?
▌Comparison with NPO and XPO
NVIDIA and Cisco state that front panel density is a limiting factor for scaling AI architectures: once the switch ports are fully utilized, the next step requires adding another switch layer with more optical modules, which increases latency, power consumption, and cost. The USFP module retains a pluggable design, allowing modules to be replaced and maintained on-site.
The first batch of planned transceivers integrates 16 full-duplex lanes within a package similar in size to the OSFP module, achieving a transmission rate of up to 6.4Tb/s per module. It can also scale to 32 lanes and a transmission rate of 12.8Tb/s. The front panel bandwidth density of USFP reaches 4 times that of existing 200Gb/s per-lane pluggable modules, with per-bit energy efficiency improved by approximately 1.3 times compared to the latter.
In terms of end-user demand, as the bandwidth of AI switches reaches the scale of 102.4T and 200T, optical interconnects must be both highly dense and replaceable.
In the past, the conventional wisdom was that when single-module bandwidth exceeds 1.6T, achieving ultra-high-density interconnects could only rely on NPO or CPO solutions, as the front panel space and power consumption of switches could hardly support it. USFP achieves 6.4T in a size similar to OSFP and is scalable to 12.8T, proving that the front panel pluggable form factor can match the target bandwidth levels of NPO in terms of single-module bandwidth metrics.
However, the system power consumption bottlenecks of the two are completely different. USFP merely benchmarks bandwidth metrics, which does not equate to gaining the system energy efficiency advantages of NPO. Moreover, as a switch front panel pluggable module, USFP still requires long-distance PCB routing for electrical signals. NPO places the optical engine near the ASIC, relying on extremely short electrical links to reduce SerDes power consumption, addressing a completely different system bottleneck.
The direct competitor corresponding to USFP should be the XPO technology. Both XPO and USFP are hot-pluggable high-density optical modules for the switch front panel, and their cages are physically incompatible and cannot be mixed.
In March 2026, Arista led the formation of the XPO MSA, bringing together 45+ vendors. The XPO single-lane baseline is 200G/lane PAM4, with a future roadmap upgradable to 400G/lane. The module width is approximately 2.7 times that of OSFP, whereas the USFP size is close to OSFP.
Both can reach an upper limit of 12.8T, but the implementation paths differ: XPO uses 200G rate stacked lanes (64 × 200G = 12.8T), while USFP directly increases the single-lane rate (6.4T / 32 × 400G = 12.8T). There is a significant difference in thermal design: XPO modules have built-in cold plates, supporting up to 400W power consumption, natively targeting liquid-cooled data centers; USFP has a power consumption limit of 100W, primarily relying on air cooling, and does not feature module-level liquid cooling blind-mate connectors.
▌USFP May Occupy a New Ecological Niche
Currently, NVIDIA has invested in multiple technologies. On the CPO route, Spectrum-X / Quantum-X Photonics have been shipped to CoreWeave, Lambda, Meta, and Oracle. On the NPO route, Rubin Ultra NVL576 adopts the NPO architecture within Scale-up clusters, no longer using traditional switch front panel pluggable transceiver modules (the NPO optical engine is an on-board socket-style pluggable, allowing the optical engine to be replaced individually).
The release of USFP this time aims to occupy a new ecological niche. NVIDIA and Cisco plan to transform this design into an open industry standard through a community-led multi-party collaboration agreement. NVIDIA and Cisco are working with TE Connectivity and Amphenol to develop cage structures, connectors, and modules. It is expected that more partners will join in the future, further expanding the interoperability efforts between NVIDIA Spectrum-X and Cisco Silicon One Ethernet.
It should be noted that USFP is designed for 400G/lane electrical signals, but currently, neither NVIDIA nor Cisco has launched a commercialized native 400G/lane SerDes switch chip. USFP is an open form factor specification defined in advance, with the optical module and connector ecosystems being developed first, while the switch ASIC is still on the product roadmap. This is also a key constraint for the implementation of the USFP ecosystem.