USFP module boosts Ethernet AI fabrics

Cisco and Nvidia introduced a new high-capacity pluggable optical system designed to accelerate the creation of Ethernet AI fabrics. This collaboration aims to simplify the infrastructure build-out for demanding AI workloads.

Unveiled at the Open Compute Project Global Summit 2026, the Ultra-Dense Small Form Factor Pluggable (USFP) module handles 400 Gb/s-per-lane electrical signals. This capacity allows for the construction of comprehensive infrastructures supporting networking, security, and high-performance AI fabrics with up to 12.8 terabits per second (T) of total capacity. The primary goal is to alleviate the pressure AI infrastructure places on connections between GPUs, switches, security appliances, and other critical systems.

Increasing bandwidth density allows customers to implement higher-capacity systems without a proportional increase in the physical footprint of connectivity hardware. Bill Gartner, Senior Vice President and General Manager of Cisco’s Optical Systems and Optics Business, and Will Eatherton, Senior Vice President of Networking Engineering, highlighted these benefits.

“USFP will provide a path supporting both 200G and 400G SerDes based systems and ASICs, helping customers scale from today’s 1.6T fabrics,” Gartner and Eatherton explained in a blog post. They further noted that with 16 lanes operating at 400 Gb/s per lane, USFP delivers 6.4T of aggregate capacity, which is four times the bandwidth density of current 1.6T pluggables.

The architecture can scale to 32 lanes, offering a pathway to 12.8T of aggregate capacity. This also presents an opportunity for approximately 1.3 times greater power efficiency through 400G SerDes technology, a crucial factor for operators managing increasing power and cooling demands in data centers. The collaboration between Cisco and Nvidia is guided by a shared objective: to help customers accelerate AI adoption while maintaining flexibility, openness, and choice.

Innovations in AI Fabric Connectivity

The USFP module preserves the serviceability of pluggable optics, meaning modules can be replaced in the field. This capability allows customers to continue using their existing deployment and maintenance practices, minimizing disruption and training requirements. The combination of higher-capacity fabrics, improved efficiency, and familiar operational procedures enables customers to scale AI infrastructure effectively. It also helps maintain the flexibility, interoperability, and consistency essential for large-scale deployments.

Nvidia emphasized that scaling AI factories requires increased bandwidth per connection and more connections per switch. When a switch reaches its capacity, further network expansion often necessitates additional switch layers and optical links, which can increase latency, power consumption, and overall costs. USFP addresses this challenge by allowing AI factory builders to scale from 200 Gb/s per lane to 400 Gb/s-per-lane Ethernet AI fabrics.

This technology extends a deployment model already proven across multiple generations of data center networks. With USFP, customers can adopt next-generation optical technology and meet the demanding performance requirements of AI workloads without completely overhauling their existing data center infrastructure. Gilad Shainer, Senior Vice President at Nvidia, stated that USFP represents a joint effort by Nvidia and Cisco to build an ecosystem that reinvents pluggable optics. This will connect more computing resources with greater efficiency to power the next generation of AI factories.

Cisco and Nvidia also plan to establish an open industry specification for USFP through a community-led multi-source project. This open approach ensures broader adoption and fosters innovation within the ecosystem. Interestingly, despite the joint announcement, Nvidia marketing director Brian Sparks clarified that the USFP technology is not part of the current partnership between Cisco and Nvidia. Instead, it reflects a broader belief among Cisco, Nvidia, and other ecosystem participants that USFP is the optimal solution for future advancements.

Expanding Silicon One Capabilities

In addition to the USFP announcement, Cisco unveiled an expansion of its Silicon One portfolio. This includes a new 51.2-terabit-per-second top-of-rack (ToR) Ethernet switching chip, specifically designed for high-density AI and cloud data-center networks.

The new Silicon One G302 chip delivers a remarkable 102.4T capacity. It integrates Cisco’s 1.6T optics and features direct-to-chip liquid cooling. This advanced cooling solution is essential for managing the significant heat generated by such high-performance components. The G302 is typically positioned at the rack level, where it aggregates traffic from servers and efficiently connects the switch to the broader data center network. This strategic placement ensures optimized data flow and network performance.

The G302 expands the existing 102.4 Tbps Silicon One G300 portfolio, which Cisco confirms is currently shipping. This portfolio incorporates Intelligent Collective Networking, a suite of features that supports shared packet buffers, path-based load balancing, and proactive network telemetry. These capabilities are crucial for handling the large-scale, bursty data center traffic associated with AI training, inference, and real-time agentic workloads.

The G302 chip forms the core of two new fixed-form 64-port switches. These include the 3RU air-cooled 8133 model and Cisco’s inaugural commercially available water-cooled switch, the 2RU 8132. The introduction of a water-cooled switch underscores the increasing need for advanced thermal management solutions in high-density data center environments. The Silicon One G302 is scheduled to begin shipping to customers before the end of 2026. This timeline provides data center operators with a clear roadmap for adopting this next-generation switching technology.