Cisco Announces G302 Silicon and USFP Optics With NVIDIA at OCP 2026

Cisco unveiled new networking silicon, systems and optics for open AI infrastructure at the OCP Global Summit 2026 in San Jose on October 12, 2026, including a 51.2T top-of-rack switching chip, a new out-of-band management system, and a planned 400Gb/s-per-lane pluggable optics module designed with NVIDIA.

Kevin Wollenweber, Cisco’s senior vice president and general manager of data center and internet infrastructure, said AI is changing what it means to scale infrastructure, shifting the focus from individual data centers or clusters to how compute is connected at massive scale while performance and resilience are preserved.

Scale-Across Architecture and P200 Production

Cisco’s scale-across work targets AI jobs that run across multiple facilities rather than inside a single site. The approach builds on conventional data center interconnect technology, but the workload demands differ: when accelerators across sites share one coordinated job, congestion, latency and recovery can directly affect compute utilization and job progress. To address this, Cisco pairs its Silicon One P200 with deep-buffered systems and coherent optics, connecting AI fabrics in different locations so they operate as one unified cluster.

Cisco said it is the first in the industry to announce 51.2T systems powered by the Silicon One P200 running in a production environment at leading hyperscalers. “Microsoft has moved into production on systems powered by the Cisco Silicon One P200,” said Anand Ramakrishna, corporate vice president of engineering at Microsoft. He added that the P200’s architecture supplies the speed and enhanced buffering needed as Microsoft’s cloud and high-performance services continue to grow.

Alibaba also described a P200 deployment. Dennis Cai, vice president and head of network infrastructure at Alibaba Cloud Intelligence Group, said Alibaba has deployed its in-house Quantum Cat router at massive scale on the Silicon One P200 for its eCORE backbone network. The routers deliver 51.2 Tbps of routing capacity with a fully shared deep-buffer architecture, Cai said, and Alibaba pairs Cisco’s silicon with its own custom systems and the SONiC operating system.

G300 Sampling and New G302 Silicon

The 102.4T Silicon One G300, announced earlier in 2026, is available for customer sampling, and the first G300-powered systems are shipping to customers, Cisco said.

Cisco described the new Silicon One G302 as the first 51.2T switching silicon with 200G SerDes designed for top-of-rack deployments. The chip delivers 51.2T of capacity with Cisco’s 1.6T optics and integrates direct-to-chip liquid cooling. Top-of-rack systems operate at the rack level, collecting server traffic and passing it into the wider data center network. The G302 is planned to begin shipping to customers before the end of 2026.

Cisco said it is a member of OCP’s ESUN and is collaborating on the 1.0 and future ESUN specifications for scale-up workloads. The company introduced Silicon One in 2019 and characterizes it as the industry’s most scalable and programmable unified networking architecture, able to scale from 100G to 100T on a single unified silicon design.

USFP Pluggable Optics With NVIDIA

In a separate joint release, NVIDIA and Cisco detailed plans for the Ultra-Dense Small Form Factor Pluggable (USFP), a high-density 400Gb/s-per-lane optical module. The companies said USFP is designed to be the first module to support 400Gb/s-per-lane electrical signals, with capacity of up to 12.8Tb/s.

The first USFP-based transceivers on the roadmap will fit 16 full-duplex lanes inside a footprint the size of an octal small form factor (OSFP) module, serving both switches and SuperNICs. Lane count can later expand to 32 or more, a scaling path enabled by the connector design. The module’s new two-dimensional array connector uses a two-step insertion mechanism in which high-speed contacts engage only after the module is seated, which the companies say reduces contact wear, improves signal integrity and makes insertion force more predictable. The design supports power dissipation of up to 100 watts without a dedicated cooling circuit embedded in the transceiver, which the companies say simplifies switch design and improves reliability.

According to the companies, USFP delivers 4x the bandwidth density and roughly 1.3x the power efficiency of 200Gb/s-per-lane pluggable modules in a comparable OSFP-sized module, allowing AI builders to grow compute scale within the same network footprint. “AI builders need an interconnect architecture that can scale performance without locking them into a more complex operating model,” said Bill Gartner, Cisco’s senior vice president and general manager of optical systems and optics. NVIDIA senior vice president Gilad Shainer said the two companies are building an ecosystem to reinvent pluggable optics.

TE Connectivity and Amphenol are collaborating with the two companies on USFP cages, connectors and modules, with more partners expected. NVIDIA and Cisco plan to create an open industry specification for USFP through an upcoming community-led multi-source project, extending existing work intended to keep NVIDIA Spectrum-X Ethernet networking and Cisco Silicon One-based Ethernet networks interoperable for enterprises, neoclouds and sovereign clouds.

SONiC, Live Protect and the Cisco 8001

Cisco described SONiC (Software for Open Networking in the Cloud) as an open-source operating system used as the foundation of deployments by hyperscalers and, increasingly, neoclouds, service providers and enterprises. Customers can adopt Cisco-validated SONiC with Nexus One, the company’s next-generation data center architecture, and that support will be extended with Live Protect starting mid-year 2027. Cisco said Live Protect uses its runtime technology to shrink the exposure window between a vulnerability disclosure and a patch on supported products, allowing SONiC customers to validate and roll out the permanent fix while operations continue.

Operators rely on out-of-band management networks for remote access when the production network goes down. At AI scale, Cisco said, those environments can turn into a patchwork of legacy systems, tools and proprietary interfaces that leaves thousands of GPU nodes harder to manage and repair. The Cisco 8001, a new Silicon One A100-powered management system offered in 10G or 1G versions, brings SONiC to that management network, so operators can use the APIs, automation and operating practices they already know across the fabric and the rack without maintaining a separate management environment.