NVIDIA and Cisco have announced plans for a new high-density optical hardware design intended to help AI data centers move more data while managing networking power consumption and infrastructure complexity. Unveiled on October 12, 2026, the proposed Ultra-Dense Small Form Factor Pluggable (USFP) technology is designed for next-generation Ethernet networks connecting large AI computing systems.
The announcement targets a growing challenge in AI infrastructure: as data centers add more accelerators and run increasingly demanding AI workloads, their networks must handle much higher volumes of data without proportionally increasing power use, physical space or operational complexity.
According to the companies’ official announcement, USFP is designed to support 400-gigabit-per-second electrical signaling per lane and up to 12.8 terabits per second of aggregate capacity.
What Is the New USFP Optical Hardware?
USFP stands for Ultra-Dense Small Form Factor Pluggable. It is a proposed form factor for optical transceiver modules used to transmit data between networking equipment and computing systems.
Optical transceivers convert electrical signals into optical signals and back again, allowing data to travel through fiber-optic links. These connections are essential in large AI data centers, where thousands of processors may need to exchange information rapidly during AI model training and inference.
NVIDIA and Cisco say the initial planned design will use 16 full-duplex lanes within a footprint comparable to an Octal Small Form Factor Pluggable (OSFP) module. The companies also describe a potential path to 32 or more lanes in future designs.
The aim is to increase the amount of network capacity available in a given physical space while retaining a familiar pluggable-module approach.
Key Specifications and Expected Benefits
The companies have outlined several headline specifications for the planned USFP design.
- 400 Gb/s per lane: The design is intended to support electrical signaling at 400 gigabits per second per lane.
- Up to 12.8 Tb/s aggregate capacity: The proposed 16-lane configuration is designed to deliver a combined capacity of up to 12.8 terabits per second.
- Four times the bandwidth density: NVIDIA and Cisco say USFP is expected to provide four times the bandwidth density of 200 Gb/s-per-lane pluggable optics.
- Around 1.3 times the power efficiency: The companies say the design is expected to improve power efficiency compared with the same 200 Gb/s-per-lane reference.
- A familiar deployment model: The module is intended to preserve the serviceability and operational flexibility associated with pluggable optics.
These are company-provided design expectations, not independently verified measurements from broad commercial deployments. Actual results will depend on the final hardware, network configuration and operating conditions.
Why Optical Networking Matters for AI Data Centers
AI infrastructure depends on more than powerful processors. Accelerators must exchange model parameters, training data and intermediate results across high-speed networks. As clusters become larger, networking can become a major factor in overall system performance and energy consumption.
When existing switches reach their capacity limits, operators may need to add networking tiers or additional links. Those changes can increase costs, complexity, power requirements and latency.
Higher-capacity optical connections could help data-center operators connect more computing resources within a similar physical footprint. If the promised density and efficiency gains are achieved, USFP could give infrastructure builders another way to expand AI clusters without increasing networking overhead at the same rate as compute capacity.
However, optical modules are only one part of the system. Switch silicon, cabling, cooling, software, power delivery and network design all influence the performance of a complete AI data center.
NVIDIA and Cisco Plan an Open Hardware Ecosystem
NVIDIA and Cisco say they are working with component suppliers TE Connectivity and Amphenol to support a multi-source hardware ecosystem around USFP.
The companies also plan to develop an open industry specification through a community-led project. A shared specification could allow multiple suppliers to build compatible components and give data-center operators more choice when sourcing hardware.
A multi-vendor ecosystem may also reduce dependence on a single supplier and help customers integrate future networking generations into existing infrastructure. Those benefits will depend on industry participation, specification maturity, interoperability testing and the availability of compatible products.
The announcement was made around the Open Compute Project Global Summit 2026, where companies are discussing infrastructure designs intended to support the next generation of AI computing.
When Will USFP Hardware Be Available?
The October 12 announcement describes plans for the USFP form factor and its supporting ecosystem. It does not, by itself, establish a confirmed general-availability date, final product pricing or broad customer deployment.
Before organizations can assess the hardware for production use, they will need further details on final specifications, supported networking equipment, interoperability, power consumption and supply availability.
Data-center operators considering next-generation optical hardware should compare real-world performance, total cost of ownership and compatibility with their existing network architecture rather than relying only on announced specifications.
What the Announcement Means for the AI Hardware Market
The NVIDIA-Cisco initiative reflects a wider shift in AI infrastructure design. As demand for AI computing expands, companies are increasingly working across processors, networking, optics and system architecture instead of treating each component as a separate upgrade.
The potential significance of USFP is not simply higher link capacity. It is the possibility of fitting more network bandwidth into existing space while improving the amount of data moved per unit of power.
Whether the design delivers those benefits at scale will depend on implementation and adoption. For now, the announcement establishes a proposed direction for next-generation AI networking rather than proof of a completed industry-wide transition.
Read More:- UK Gains Access to Advanced AI Models for Safety Testing
Frequently Asked Questions
What did NVIDIA and Cisco announce?
They announced plans for USFP, a high-density pluggable optical hardware form factor intended to support faster and more efficient networking for AI data centers.
What is the maximum capacity of USFP?
The companies describe a planned 16-lane design with aggregate capacity of up to 12.8 terabits per second, using 400 Gb/s-per-lane signaling.
Will USFP make AI data centers use less electricity?
The companies expect the design to improve power efficiency by around 1.3 times compared with 200 Gb/s-per-lane pluggable optics. Actual savings will depend on the final implementation and the wider data-center network.
Is USFP available to buy now?
The announcement outlines plans for the technology and its ecosystem. It does not confirm general commercial availability or final pricing.
Why are optical transceivers important for AI?
They enable high-speed data transmission over fiber-optic links between networking equipment and computing systems. Fast, efficient connections help large AI clusters exchange information at scale.
