NVIDIA and Cisco have jointly introduced a new-generation ultra-high-density pluggable optical module standard called USFP, with the first product supporting 400Gb/s electrical signaling per lane, delivering four times the bandwidth density of existing solutions and marking a new phase in AI data center interconnect technology.
The two companies announced the plan on October 12 and are partnering with TE Connectivity and Amphenol to build a hardware ecosystem covering the development and production of cages, connectors, and modules. Meanwhile, NVIDIA and Cisco plan to advance USFP as an open industry specification through a community-led multi-source project.
The release of USFP directly addresses the interconnect bottleneck created by the scaling of AI factories. As demand for AI model training and inference continues to expand, insufficient switch port bandwidth will force operators to stack more network layers, driving up latency, power consumption, and construction costs. If the new standard lands on schedule and achieves broad industry adoption, it will substantially reshape the optical module supply chain landscape.
Performance Leap: Four Times Bandwidth Density with Energy Efficiency Gains
The core breakthrough of USFP lies in raising per-lane electrical signaling rates from 200Gb/s to 400Gb/s. The first planned product packages 16 full-duplex lanes within the same module size as existing OSFP, achieving 6.4Tb/s aggregate capacity and reserving an upgrade path to 32 lanes with a total capacity of 12.8Tb/s.
Compared with 200Gb/s per lane pluggable modules of the same size, USFP achieves four times the bandwidth and front-panel bandwidth density improvement, along with approximately 1.3 times better power efficiency. Gilad Shainer, Senior Vice President at NVIDIA (NVDA), said, "Building the world's largest AI factories begins with rethinking the smallest connections."
In terms of thermal design, USFP supports up to 100 watts of power consumption without requiring an independent cooling loop integrated within the transceiver, effectively reducing switch design complexity and improving overall reliability.
Connector Innovation: Two-Dimensional Array Structure Breaks Through Signal Integrity Bottleneck
USFP adopts an entirely new two-dimensional array connector paired with a two-step insertion design, where high-speed contacts engage only after the module completes physical insertion. This mechanism effectively reduces contact wear, improves signal integrity, and makes insertion force more predictable.
The connector architecture is purpose-built for 400Gb/s performance while also delivering superior signal performance under 200Gb/s per lane conditions, providing a compatible foundation for smooth transitions from existing systems. Bill Gartner, Senior Vice President and General Manager of the Optical Systems and Optical Components business at Cisco (CSCO), noted that scaling pluggable optical modules to 400Gb/s per lane requires not merely incremental extensions of existing module designs, but full-stack collaborative optimization spanning chips, systems, optical components, and network operating systems.
USFP also supports multiple optical technologies and connection distances, giving system designers the flexibility to address different AI factory deployment scenarios.
Ecosystem Building: TE Connectivity and Amphenol Join to Reshape the Supply Chain
NVIDIA and Cisco bringing in TE Connectivity and Amphenol to participate in the development of USFP cages, connectors, and modules is the most directly supply-chain-relevant information in this announcement. Both are major suppliers in the connector space, and their confirmed participation means USFP has already formed a preliminary multi-source supply landscape at the hardware foundation level, with more partners expected to join subsequently.
A unified module form factor provides suppliers with a common development foundation, conducive to forming a scalable and competitive supply system around 400Gb/s per lane interconnect. This model continues the tradition of the pluggable optical module ecosystem emphasizing interoperability and supplier diversification, offering differentiated advantages in field replaceability and supply chain maturity compared with approaches such as co-packaged optics (CPO).
This collaboration is also built on the existing synergy between NVIDIA's Spectrum-X Ethernet networking and Cisco's Silicon One Ethernet networking, aimed at simplifying AI infrastructure deployment for enterprises, neocloud service providers, and sovereign cloud customers.
Open Standard Path: Multi-Source Specification Project Lowers Deployment Barriers
NVIDIA and Cisco plan to promote USFP as an open industry specification through an upcoming community-led multi-source project. The establishment of an open standard will lower the barrier for a broader range of vendors to participate in the ecosystem and provide data center operators with greater flexibility in procurement and replacement decisions.
For AI factory builders, the value proposition of USFP manifests in two dimensions: first, scaling compute without adding network layers, avoiding the latency and cost increases associated with stacking switching layers; second, by increasing data transmission per watt, compressing the share of network interconnect in the overall AI factory power budget.
It should be noted that NVIDIA and Cisco have not yet disclosed a specific product launch or mass production timeline, and USFP remains in the design and planning stage. Its ultimate market impact will depend on the pace of advancing the open specification, the speed of ecosystem partner expansion, and the industry's actual choices among different optical interconnect paths such as pluggable, near-packaged, and co-packaged solutions.