Particle.news

Cisco Unveils 102.4 Tbps Silicon One G300 to Take On Broadcom and Nvidia in AI Data Centers

Built on TSMC’s 3nm process, the switch introduces congestion controls plus P4 programmability that Cisco says can cut AI job times by up to 28%.

Overview

  • Cisco’s G300 switch offers 102.4 Tbps capacity with 512 200 Gbps SerDes and is slated to ship in the second half of 2026 in new N9000 and 8000 systems with 64 1.6 Tbps OSFP ports.
  • Cisco highlights a fully shared packet buffer and a path‑based load balancer, claiming roughly 33% better link utilization and up to 28% faster AI training versus packet‑spraying designs.
  • The launch adds new optics, including 1.6 Tbps pluggables and 800G linear pluggable optics, with Cisco saying the combined systems can reduce switch power by about 30%.
  • Nexus One management gains job‑aware observability and Splunk integration expected in March, with AgenticOps updates over the next quarter for autonomous troubleshooting and optimization.
  • The move positions Cisco against Broadcom and Nvidia in AI networking, as analysts note demand for vendor diversity and point to competing 102.4 Tbps gear, including HPE’s Juniper‑branded switches using Broadcom silicon, arriving on shorter timelines.