Cisco Ships Silicon One G300, Debuts G302 at OCP 2026
3 min readCisco used the Open Compute Project Global Summit in San Jose this week to show how far its AI networking roadmap has moved in eight months. The Cisco Silicon One G300, the 102.4 terabit switching chip unveiled in February, is now sampling with customers and the first G300 systems are shipping. Alongside it, Cisco introduced a new 51.2T chip, the G302, and said its P200 routing silicon is already running in production at Microsoft and Alibaba Cloud.
From one chip family to a full AI fabric
Silicon One launched in 2019 as a single architecture meant to cover everything from enterprise campuses to hyperscale backbones. The AI build-out has turned it into Cisco’s main answer to Nvidia and Broadcom in the data center, where clusters now stretch across buildings and even cities. The question for every network vendor at OCP this year was how to keep tens of thousands of GPUs talking as fabrics jump from 800G to 1.6T links.
What Cisco announced
The headline is the Cisco Silicon One G302, a 51.2T switching chip built with 200G SerDes for top-of-rack switches. It integrates Cisco’s 1.6T optics and direct-to-chip liquid cooling, and Cisco plans to ship it before the end of 2026. The G300 moves from announcement to availability, with systems now reaching customers.
Cisco also said it is the first vendor with 51.2T systems on the Silicon One P200 running in hyperscaler production. Microsoft has put P200 systems into service, and Alibaba Cloud deployed its in-house Quantum Cat router on the P200 for its eCORE backbone, pairing the chip with SONiC. Cisco calls the architecture “scale-across”: deep-buffered P200 systems plus coherent optics link separate AI fabrics into one cluster that spans multiple sites.
In optics, Cisco and Nvidia plan to co-develop USFP, a pluggable module running 400 gigabits per second per lane. Cisco says it should deliver four times the bandwidth density and 1.3 times the power efficiency of today’s 200G-per-lane optics, with an open specification to follow. Cisco also extended SONiC support to its Nexus One architecture, promised Live Protect for SONiC customers from mid-2027, and introduced the Cisco 8001, an out-of-band management box on Silicon One A100 that keeps operators connected during outages or cyber incidents. Details are in Cisco’s announcement.
Why it matters
Named hyperscaler deployments are the detail to watch. Chips get announced all the time, but Microsoft and Alibaba Cloud running P200 in production shows Cisco’s silicon winning real backbone slots, not just lab tests. The G302’s liquid cooling and built-in 1.6T optics signal where the next generation of AI racks is headed. And the USFP partnership puts Cisco and Nvidia on the same side of the optics standard fight, which matters as Broadcom pushes its own full-stack story.
Watch for G302 shipments late this year and for whether other hyperscalers follow Microsoft onto P200. Kevin Wollenweber, Cisco’s data center chief, put it simply: AI is redefining what it means to scale infrastructure.