NVIDIA and Cisco Unveil USFP Technology for AI Networks

NVIDIA and Cisco are developing USFP technology to deliver up to 12.8Tb/s capacity and improve bandwidth density for next-generation AI networks.

The companies announced plans for Ultra-Dense Small Form Factor Pluggable (USFP), a high-density optical module designed to support 400Gb/s-per-lane connections in next-generation AI networks. The architecture delivers up to 12.8Tb/s of overall capacity, giving AI infrastructure operators a way to increase network bandwidth within a compact, OSFP-sized footprint.

What is USFP Technology?

Announced at the Open Compute Project (OCP) Global Summit on October 12, USFP targets a growing challenge in AI infrastructure: connecting increasingly powerful computing systems while managing network complexity, energy consumption, and operating costs.

Also Read: Schneider Electric and NVIDIA Advance AI Data Center Infrastructure 

As AI models become larger and inference workloads increase, data centres need faster connections between graphics processing units (GPUs), switches, storage systems, and other computing components. USFP is designed to address that demand by increasing the amount of data that can move through each switch's connection points.

The Role of Optical Transceivers

Optical transceivers play an important role in modern data centres. They convert electrical signals from networking equipment into optical signals that travel over fibre-optic cables, allowing systems to exchange large amounts of data across a facility.

USFP is designed to increase the bandwidth these connections can support.

The first planned transceivers will use 16 full-duplex lanes within a footprint comparable to an Octal Small Form Factor Pluggable (OSFP) module. The architecture can expand to 32 lanes and beyond, providing a path to higher network capacity as AI infrastructure requirements grow.

Performance and Power Efficiency

According to NVIDIA and Cisco, USFP will deliver four times the bandwidth and front-panel bandwidth density, along with about 1.3 times the power efficiency of existing pluggable modules using 200Gb/s-per-lane technology.

The design also supports up to 100 watts of power dissipation without requiring a separate embedded cooling circuit within the transceiver. This could simplify switch design while helping operators manage the thermal demands of high-density AI infrastructure.

For AI data centre operators, the potential benefit is straightforward: connect more computing capacity without continually adding network layers, equipment, and optical links.

Building an Industry Ecosystem

The development goes beyond the two companies. NVIDIA and Cisco are working with connectivity manufacturers TE Connectivity and Amphenol to develop USFP cages, connectors, and modules. Additional partners are expected to join the initiative.

The shared form factor will give hardware manufacturers a common design foundation for developing compatible components. This matters because a new optical technology needs support across switches, networking equipment, connectors, and transceivers before data centre operators can deploy it at scale.

NVIDIA and Cisco also plan to develop an open industry specification through a community-led multi-source project. The approach is intended to encourage participation from multiple suppliers rather than limit the technology to a single manufacturer's hardware ecosystem.

This could give AI infrastructure builders more flexibility when sourcing networking components and designing their systems.

Scaling AI Infrastructure

USFP extends the existing collaboration between NVIDIA and Cisco on Ethernet networking for AI infrastructure. The companies are working to ensure NVIDIA Spectrum-X Ethernet networking and Cisco Silicon One-based Ethernet networks operate together.

Their broader approach targets enterprises, neocloud providers, and sovereign cloud operators building infrastructure for demanding AI workloads. For these operators, networking is becoming an increasingly important part of AI infrastructure planning. Adding more GPUs does not automatically improve performance if the network connecting them cannot move data fast enough.

Higher bandwidth and improved power efficiency can help reduce these bottlenecks while limiting the need for additional networking equipment.

The companies' emphasis on maintaining the serviceability and deployment flexibility of pluggable optics is also significant. Rather than requiring operators to adopt an entirely different networking model, USFP is designed to extend a familiar approach to higher performance levels.

Future Deployment and Availability

The announcement reflects a broader shift in data centre design as AI infrastructure moves towards larger, more interconnected computing clusters. For operators, the challenge is no longer just acquiring more computing power.

They must also provide the networking capacity, cooling, and power infrastructure needed to keep those systems working efficiently. USFP could become part of that infrastructure as AI factories scale.

Its higher bandwidth density could help operators connect more computing resources within existing network footprints, while its power-efficiency targets address some of the energy demands associated with moving data.

The development may also create opportunities for suppliers of optical components, connectors, and high-speed networking equipment as the ecosystem expands. However, its commercial impact will depend on industry adoption, hardware compatibility, and manufacturers' ability to bring the technology to market.

Despite the announcement, NVIDIA and Cisco have not made USFP commercially available. The companies are still building the hardware ecosystem and plan to establish an open industry specification through a multi-source initiative.

No confirmed product availability or shipping timeline has been announced. Next milestones include developing compatible components, expanding industry participation, and eventually integrating into networking equipment used by AI infrastructure operators.

For now, USFP offers a proposed path towards higher-capacity AI networks. Its success will depend on whether the companies can turn the design into an open, widely supported technology that data centre builders can deploy at scale.