Cisco introduces Silicon One G300 for AI networking

On February 10, 2026, Cisco announced the Silicon One G300 and said the device could power gigawatt-scale Artificial Intelligence (AI) clusters for training, inference, and real-time agentic workloads while improving Graphics Processing Unit (GPU) utilization and reducing job completion time by 28%.

Responses within the networking industry were largely positive, and analysts and partners cited the G300 as a factor in arguments that Cisco could reclaim dominance in the AI infrastructure market. Cisco said Intelligent Collective Networking could increase network utilization by 33% versus simulated non‑optimized path selection and would generate more tokens per GPU‑hour to improve profitability for large-scale AI data centers.

The G300 offered 102.4 Tbps switching capacity and implemented Intelligent Collective Networking, combining a fully shared packet buffer, path‑based load balancing, and proactive telemetry to absorb bursty traffic, respond faster to link failures, and prevent packet drops over long distances. The chip used TSMC 3nm manufacturing, was programmable, and was described as able to support deployments of up to 128,000 GPUs with security fused into the hardware.

Cisco said the G300 would power new Cisco N9000 and Cisco 8000 systems that included liquid cooling and support for high‑density optics to achieve new efficiency benchmarks, and the company said it had enhanced Nexus One to simplify operation of AI networks on‑premises or in the cloud.

“As AI training and inference continues to scale, data movement is the key to efficient AI compute; the network becomes part of the compute itself. It’s not just about faster GPUs – the network must deliver scalable bandwidth and reliable, congestion-free data movement,” said Martin Lund, Executive Vice President of Cisco’s Common Hardware Group.

Cisco described plans to use G300 programmability to enable equipment upgrades after deployment so Silicon One‑based products could support emerging use cases and multiple network roles, and it said security fused into the hardware would allow at‑speed protection to keep clusters up and running.