Cisco drops Silicon One G300 to speed up AI clusters

When people talk about the AI boom, GPUs always steal the spotlight. But inside massive data centers, the real headache is often the network. If thousands of chips stall waiting for data transfers, power and money burn for nothing.

At Cisco Live EMEA in Amsterdam on Feb. 10, Cisco took direct aim at that cluster bottleneck by unveiling the Silicon One G300, a flagship 102.4 Tbps switching chip built for gigawatt-scale AI data centers.

The new silicon powers updated Cisco N9000 and Cisco 8000 systems tailored for AI training, inference, and real-time agentic workloads. Cisco claims its Intelligent Collective Networking features—like a shared packet buffer and path-based load balancing—boost network utilization by 33% and cut AI job completion times by 28% compared to non-optimized setups.

Squeezing six boxes into one

AI clusters are running out of power and rack space. To address this, Cisco designed these systems to slash energy consumption drastically.

The new N9000 and 8000 systems offer a 100% liquid-cooled option. Cisco says combining liquid cooling with new optics delivers a nearly 70% energy efficiency boost, packing the bandwidth that previously required six systems into a single chassis.

The optics upgrades are equally aggressive. Cisco introduced 1.6T OSFP optical modules alongside 800G Linear Pluggable Optics (LPO). By bypassing traditional retimers, LPO cuts optical module power by 50% and drops overall switch power consumption by 30%.

Cisco also expanded its 51.2T P200 lineup with new 28.8T modular line cards and 800G coherent optics. The G300, its systems, and the new optics are scheduled to ship later this year.

AI agents to troubleshoot AI networks

Managing an AI cluster network is notoriously painful when links drop during critical training runs. Cisco is attempting to ease operations through upgrades to its Nexus One platform.

One standout addition is AgenticOps via an AI Canvas interface, enabling engineers to troubleshoot complex network incidents through guided conversational interactions.

Cisco is also integrating its Splunk acquisition directly into Nexus One starting this March. This allows operators to analyze network-to-GPU telemetry right where the data lives without exporting it, a crucial capability for sovereign clouds and tightly regulated enterprises.

Strategic partners backing the launch include AMD, Intel, Nvidia, DDN, NetApp, and VAST, showcasing broad ecosystem support across compute and storage.

Why it matters

For enterprise builders, sovereign clouds, and neoclouds, Cisco's release offers a viable off-the-shelf path to hyperscaler-grade networking. Building massive clusters usually requires painful trade-offs in power and rack footprint; a liquid-cooled 102.4T box that saves significant power lets teams scale without redesigning entire facilities from scratch.

For the wider industry, this is an important battle line in the Ethernet versus Nvidia InfiniBand fight. Ethernet backers want an open, multi-vendor standard to scale AI clusters, and Cisco brings serious silicon firepower to that coalition alongside partners like AMD and Intel.

We think Cisco's full-stack execution is smart: rather than just selling raw silicon, it bundles hardware, LPO optics, liquid cooling, and Splunk observability into a cohesive story. The primary risk lies in adoption among tier-one hyperscalers, who often lean toward white-box hardware and in-house chips. Cisco will need to prove that its promised 28% job-speed gains hold up outside the test lab to justify the enterprise price tag.