Cisco and NVIDIA Unveil USFP Module to Scale AI Networks to 12.8Tb/s Capacity

Cisco and NVIDIA Introduce USFP for Next-Generation AI Networks

Cisco and NVIDIA announced plans for the Ultra-Dense Small Form-Factor Pluggable (USFP), a high-density optical module designed to address the escalating connectivity demands of AI infrastructure1

1

2

2

. The collaboration was unveiled at the Open Compute Project Global Summit, where Cisco delivered one of the event's most anticipated keynotes3

3

. USFP represents the industry's first pluggable form factor capable of supporting 400G-per-lane electrical signals, enabling AI data centers to scale from today's 1.6T fabrics to achieve 12.8Tb/s capacity1

1

5

5

.

Addressing Critical Bandwidth and Power Challenges in AI Factories

As AI models expand and inference demand grows, AI factories must connect more computing systems while managing network complexity, power consumption, and costs2

2

. Modern AI workloads depend on continuous communication across accelerators, applications, storage, and services, driving unprecedented requirements for bandwidth, responsiveness, and scale1

1

. When switches reach capacity, further network expansion typically requires additional switch layers and optical links, increasing latency, power consumption, and costs2

2

. USFP delivers 4x the bandwidth density and approximately 1.3x the power efficiency compared to 200Gb/s-per-lane pluggable modules, allowing AI factories to increase compute scale within the same network footprint while decreasing networking's share of the overall power budget2

2

4

4

.

Technical Specifications and Design Innovation

The first planned USFP-based transceivers will pack 16 full-duplex lanes in an OSFP-sized footprint supporting both switches and SuperNICs, with the ability to expand to 32 lanes and more in the future2

2

5

5

. By doubling both the lane count and bandwidth per lane, USFP enables AI networks to connect more 1.6Tb/s endpoints without adding network tiers2

2

. The module uses a two-dimensional array connector with a two-step insertion design that engages high-speed contacts after insertion, reducing contact wear, improving signal integrity, and making insertion force more predictable2

2

5

5

. USFP supports power dissipation of up to 100 watts without needing a separate, dedicated embedded cooling circuit within the transceiver, reducing switch design complexity and improving reliability2

2

5

5

.

Preserving Operational Simplicity While Advancing Performance

Cisco and NVIDIA designed USFP to deliver next-generation optical technology without requiring customers to overhaul their AI data centers or abandon established operational models1

1

2

2

. Organizations continue to value independently field-replaceable modules, mature supply chains, interoperability across suppliers, and familiar serviceability models that become even more critical as AI workloads grow in scale and complexity1

1

. "AI builders need an interconnect architecture that can scale performance without locking them into a more complex operating model," said Bill Gartner, senior vice president and general manager of optical systems and optics at Cisco2

2

. The connector technology is designed for 400Gb/s performance while offering improved performance at 200Gb/s per lane, providing a migration path for existing deployments2

2

.

Building an Open Ecosystem for AI Infrastructure

Cisco and NVIDIA are collaborating with TE Connectivity and Amphenol to build a multi-source hardware ecosystem around USFP and develop cages, connectors, and modules2

2

5

5

. The companies plan to create an open industry specification for USFP through an upcoming community-led multi-source project5

5

. This approach aligns with Cisco's commitment to open infrastructure that can be deployed by hyperscalers, neoclouds, and enterprises alike3

3

. "Building the world's largest AI factories starts with rethinking the smallest of connections," said Gilad Shainer, senior vice president at NVIDIA2

2

5

5

. This collaboration builds upon the existing commitment to NVIDIA Spectrum-X Ethernet networking and Cisco's Silicon One-based Ethernet networks working together seamlessly4

4

.

Strategic Implications for Distributed AI Computing

The introduction of pluggable optics and pluggable transceivers through USFP arrives as networking has become a major bottleneck in AI data centers3

3

. Frontier models require enormous clusters made up of thousands of graphics processing units and AI accelerators that must act as one giant processor3

3

. Kevin Wollenweber, Cisco's senior vice president and general manager of data center and internet infrastructure, emphasized that AI requires a new approach to scaling that incorporates the reality that many AI workloads are now distributed across multiple physical sites3

3

. The 12.8Tb/s capacity enabled by USFP positions AI infrastructure operators to handle growing data movement requirements without proportional increases in hardware costs or power consumption, addressing critical concerns as AI adoption accelerates across industries.