Cisco launches Silicon One G300 and liquid-cooled N9000/8000 systems to accelerate AI data centers

Cisco unveiled a new generation of switching silicon and systems designed to support the scale and operational needs of modern AI deployments, coupling raw on‑chip bandwidth with thermal, optical and operational innovations. The Silicon One G300 delivers 102.4 Tbps of on‑chip bandwidth and introduces larger buffering and control/data‑plane techniques intended to smooth bursty AI traffic and reduce congestion‑related GPU stalls. Cisco pairs that silicon with N9000 and 8000 systems offered in air‑cooled and fully liquid‑cooled configurations to push port density, thermal headroom and power efficiency for GPU‑dense clusters. The hardware roadmap includes 1.6T‑capable optics and lower‑power linear pluggable (LPO) modules engineered to cut optical module power and overall switch energy consumption. On the management side, Cisco expanded its Nexus One fabric to provide a centralized control plane with job‑aware telemetry from network to GPU and tighter integrations with observability platforms. Cisco is aligning those telemetry and automation capabilities with broader AgenticOps and agent‑governance initiatives to help operators manage AI agents, runtime model behavior and policy enforcement across multi‑site fabrics. The vendor also highlighted Outshift and SASE/security enhancements that bring continuous runtime monitoring, adaptive testing and proactive policy recommendations into network and edge workflows, aiming to reduce model and agent risk in production. Cisco said some software features will roll out in stages — with datacenter capabilities entering controlled availability in the months ahead and additional observability integrations following — while the systems and optics are due to ship during the year. Partners across compute and storage ecosystems are working to validate end‑to‑end stacks that reduce network‑induced GPU starvation and sustain throughput in production AI workloads. For operators, the combined approach of deeper buffering, path‑aware load balancing, richer telemetry and agent‑aware automation targets measurable improvements in utilization, job completion time and operational predictability without changing GPU hardware. Cisco frames the offering for hyperscalers, sovereign and neoclouds, service providers and large enterprises moving from pilots to continuous, cost‑sensitive AI services.