NVIDIA and Cisco are developing a new high-density optical transceiver design that could help AI data centers move more data through their networks without requiring additional networking equipment or consuming as much power. Announced today, October 12, 2026, the new Ultra-Dense Small Form Factor Pluggable (USFP) aims to address a growing challenge for AI infrastructure: scaling network bandwidth as AI computing systems become increasingly powerful.

The companies say the new form factor is designed to support electrical signaling at 400 gigabits per second (Gbps) per lane, with aggregate bandwidth reaching up to 12.8 terabits per second (Tbps). It is intended to help operators build larger Ethernet networks for AI workloads while retaining the serviceability and deployment flexibility of existing pluggable transceivers.

The announcement comes as AI factories — large-scale data center environments built to train and run AI models — require increasingly fast connections between GPUs, networking equipment, storage systems and other computing resources. Improving the performance of those connections is becoming just as important as increasing the processing power available inside the data center.

USFP aims to deliver more bandwidth in the same space

The new USFP design focuses on increasing the amount of data that networking equipment can handle within a familiar physical footprint. NVIDIA and Cisco plan to introduce initial transceivers with 16 full-duplex lanes in a module footprint comparable in size to an Octal Small Form Factor Pluggable (OSFP) module.

Each lane is designed to support 400Gbps signaling, allowing the planned 16-lane configuration to reach 12.8Tbps of aggregate capacity. The companies also envision scaling the design to 32 or more full-duplex lanes in future configurations.

That higher bandwidth density could allow data center operators to connect more computing systems without adding extra layers of network switches. Additional network tiers can increase complexity, latency, power consumption and deployment costs, making it harder to scale large AI clusters efficiently.

NVIDIA and Cisco say USFP will deliver four times the bandwidth density of existing 200-Gbps-per-lane pluggable modules in a comparable footprint. They also project approximately 1.3 times the power efficiency, helping reduce the networking portion of an AI factory’s overall power budget. These are company-stated comparisons rather than independently verified performance results.

A new connector designed for 400-Gbps-per-lane signaling

USFP is not simply a higher-bandwidth version of an existing optical module. The design includes a new two-dimensional array connector and a two-step insertion mechanism intended to improve signal integrity and reduce wear on electrical contacts.

The connector engages its high-speed contacts after the module has been inserted, which the companies say helps make insertion force more predictable. The design is also intended to support 200-Gbps-per-lane operation, giving system designers another option when building networking equipment around different performance requirements.

Thermal management is another important consideration. USFP is designed to support power dissipation of up to 100 watts without requiring a separate, dedicated embedded cooling circuit inside the transceiver. According to NVIDIA and Cisco, this could simplify switch design and help improve reliability as networking hardware handles higher bandwidth.

The companies also say the form factor supports different optical technologies and connection distances. Its optical designs are intended to reduce latency and power consumption by simplifying signal processing, although actual results will depend on the implementation and deployment requirements.

NVIDIA and Cisco want to build an open hardware ecosystem

The collaboration extends beyond the two companies. NVIDIA and Cisco are working with TE Connectivity and Amphenol on USFP cages, connectors and modules, with additional participants expected to join the ecosystem.

The companies also plan to develop an open industry specification through a community-led, multi-source project. That approach could give multiple suppliers a common design foundation instead of leaving the next generation of high-density pluggable optics dependent on a single manufacturer.

For data center operators, a broader supplier ecosystem could offer more flexibility when sourcing networking components and designing systems for different AI workloads. However, the announcement describes plans for developing the ecosystem and specification, rather than a finished, widely available product lineup.

The effort builds on NVIDIA and Cisco’s existing work to make NVIDIA Spectrum-X Ethernet networking and Cisco Silicon One-based Ethernet networks work together. The companies say the collaboration is intended to simplify AI infrastructure deployments for enterprises, neocloud providers and sovereign cloud operators.

Why optical networking matters for AI infrastructure

As AI clusters grow, the connections between computing systems become a major engineering consideration. Faster processors and accelerators can generate and exchange enormous amounts of data, putting pressure on the networks that connect them. Increasing bandwidth without proportionally expanding networking equipment can help operators use existing space and infrastructure more effectively.

Optical transceivers convert electrical signals into optical signals for transmission over fiber, then convert incoming optical signals back into electrical form. They are an important part of high-speed data center networks, particularly where optical links can provide the reach and bandwidth required to connect equipment across racks and facilities.

USFP is designed to extend the familiar pluggable-transceiver model into higher-bandwidth networking. Rather than requiring operators to abandon that deployment approach, NVIDIA and Cisco want to increase capacity while retaining established maintenance and replacement practices.

The practical benefits will depend on how the technology performs in real systems, how quickly suppliers adopt the design and how it compares with competing approaches to scaling AI networks. The companies have announced the proposed form factor and their ecosystem plans, but have not provided a confirmed commercial release date or pricing in the announcement.

For now, USFP represents a proposed step toward denser, more power-efficient networking for next-generation AI factories. As AI infrastructure expands, making every switch port and watt go further could become increasingly important alongside advances in the processors doing the actual computing.