Cisco Unveils 102.4 Tbps Silicon One G300 Chip to Challenge Nvidia and Broadcom in AI Data Center Networking
Cisco's 3nm G300 switching silicon promises 33% better network utilization and 70% energy savings with liquid cooling, targeting the $600 billion AI infrastructure buildout.
Overview
Cisco on February 10 unveiled the Silicon One G300, a 102.4-terabits-per-second switching chip fabricated on TSMC’s 3-nanometer process, marking the company’s most aggressive move yet into the AI data center networking market dominated by Nvidia’s InfiniBand and Broadcom’s Tomahawk silicon, according to Data Center Knowledge. The chip anchors a new family of liquid-cooled Ethernet switches and routers designed to interconnect hundreds of thousands of GPUs inside AI training and inference clusters.
“With Silicon One, the differentiation isn’t just bandwidth. It’s how efficiently we drive AI workloads,” said Rakesh Chopra, Cisco’s SVP and Silicon One fellow, as quoted by Data Center Knowledge.
What the G300 Does
The G300’s headline feature is Intelligent Collective Networking, a combination of fully shared packet buffering, path-based load balancing, and proactive network telemetry that Cisco says delivers a 33 percent increase in network utilization and a 28 percent reduction in AI job completion time compared to non-optimized routing configurations, according to Cisco. By rerouting data at millisecond intervals to avoid congestion, the chip aims to minimize the time GPUs spend idle waiting for data — a problem that becomes more severe as AI clusters scale into the tens of thousands of accelerators.
“The network must deliver scalable bandwidth and reliable, congestion-free data movement,” said Martin Lund, Cisco’s executive vice president of Common Hardware, per Cisco.
Systems and Cooling
The G300 will power new fixed and modular platforms in Cisco’s Nexus 9000 and Cisco 8000 series, including fully liquid-cooled configurations that deliver approximately 70 percent better energy efficiency than prior-generation air-cooled systems, according to SiliconANGLE. A single G300-based system can match the aggregate bandwidth of six previous-generation units, per SiliconANGLE.
Cisco also announced 1.6-terabit OSFP optical modules and 800G Linear Pluggable Optics (LPO) that reduce optical power consumption by 50 percent compared to retimed modules, with switches cutting overall power by 30 percent, per Cisco.
Alongside the G300, Cisco expanded deployments of the Silicon One P200, a 51.2 Tbps chip for long-distance data center interconnection, with new 28.8-terabit modular line cards, according to SiliconANGLE.
Software and Observability
The hardware launches are paired with IOS XE 26, a new operating system release featuring post-quantum cryptography and hardware-fused security, per Data Center Knowledge. Cisco’s updated Nexus One management platform will integrate with Splunk starting in March 2026 to provide job-aware, network-to-GPU observability, and a new AgenticOps capability uses AI Canvas to assist with troubleshooting, according to Cisco.
Competitive Landscape
The G300 positions Cisco squarely against Nvidia’s InfiniBand fabric, which has dominated high-performance AI cluster networking, as well as Broadcom’s Tomahawk switching silicon. Analyst Ron Westfall said that “by reaching the 102.4 Tbps threshold, Cisco has closed the performance gap” between Ethernet and InfiniBand, potentially shifting the industry toward unified Ethernet-based standards rather than proprietary interconnects, as reported by Data Center Knowledge.
An analyst quoted by SiliconANGLE noted that “few competitors can match Cisco’s integrated stack of silicon, systems, optics, software and operations tooling.” Cisco’s approach bundles custom silicon with purpose-built switches, optics, and management software into a single vendor stack — a strategy designed to simplify deployment for hyperscalers building gigawatt-scale AI data centers.
What We Don’t Know
- Pricing for the G300-based systems and how it compares to Nvidia and Broadcom alternatives.
- Whether hyperscale customers — particularly the largest GPU buyers like Microsoft, Meta, and Google — will commit to Cisco’s Ethernet approach over Nvidia’s entrenched InfiniBand ecosystem.
- Whether Cisco’s 28 percent job completion improvement holds in real-world multi-tenant production environments, as opposed to simulated benchmarks.
- The G300 is expected to ship in the second half of 2026, leaving competitors time to respond with their own next-generation silicon.