As AI clusters scale, network silicon becomes the bottleneck. Here’s a look at Cisco’s Silicon One G300 and P200, alongside Broadcom’s Tomahawk 6 and Nvidia’s Spectrum-6.
Adding thousands of GPUs to a data center can be a colossal waste of money if the network fabric can’t keep up. It’s like pairing a modern laptop with a 28k or 56k dial-up modem: throughput and latency constraints will dominate. Vendors such as Cisco, Broadcom, and Nvidia are converging on 51.2-102.4 Tbps-class Ethernet switching silicon with deeper buffering, richer telemetry, and load-balancing strategies to keep GPU utilization high and job completion times low.
“In AI data centers, the network is becoming a primary determinant of cluster efficiency,” said Sameh Boujelbene, Vice President of Research at Dell’Oro Group. “A company can spend billions on GPUs, but if the fabric cannot deliver predictable bandwidth and low latency, they did not buy an AI supercomputer; they bought an expensive collection of stranded chips.”
How AI Infrastructure Scales: Up, Out, and Across #
AI infrastructure typically grows along three complementary dimensions. Scale up packs more compute per server or rack, often via high-bandwidth intra-node links. Scale out adds more racks and combines their resources to run larger models or concurrent jobs. Scale across connects multiple data centers over optical networks, enabling clusters to work together across regions. Each dimension puts different requirements on the network, from intra-rack latency to wide-area buffering and telemetry.
Cisco Silicon One Roadmap and Timing #
To meet the demands of AI networking, Cisco has developed new Silicon One chips for its high-end switches and routers. Initially announced in February 2026, Cisco used its Cisco Live event in June to outline availability: the new devices – part of the Silicon One family – are expected to ship broadly before the end of the year.
G300 for Scale Out
Cisco positions the Silicon One G300 for scale-out within the data center. “The Cisco Silicon One G300 chip is designed for scale-out compute to manage the backend networking between racks of GPUs and other devices inside the data center,” said Nick Kucharewski, Senior Vice President and General Manager, Cisco Silicon One.
The G300 class exposes a large number of high-speed channels and emphasizes congestion avoidance and burst tolerance. “As congestion can happen randomly, packets are instantaneously redirected to avoid delays and enable GPUs to do more work,” Kucharewski said.
In this class, aggregate bandwidth reaches 102.4 Tbps via 512 lanes at 200 Gbps per lane, and the silicon integrates real-time telemetry, identity-aware forwarding, and traffic visibility. In system form factors, Cisco targets large-cluster environments across training, inference, and real-time agentic workloads.
“The networking enables GPUs to work as a supercomputer,” Kucharewski said. “No matter the AI traffic bursts, we can deal with it.”
Cisco's G300 chip and Nexus system. (Image: Cisco)
P200 for Scale Across
For scale-across, Cisco’s Silicon One P200 focuses on inter-data center connectivity over optical networks. The goal is sustaining high-throughput transfer over serialized wide-area links with deep buffering to minimize perceived latency. The P200 provides 51.2 Tbps via 512 x 100 Gbps links and is paired with external high-bandwidth memory (HBM) for deep buffers and traffic management features aligned to service provider requirements. “The Cisco 9000 with the P200 inside allows optical fiber to operate close to capacity for long periods,” Kucharewski noted. “Its deep buffer is effective in dealing with congestion.”
Select customers have access to early P200 systems. In Q3 2026, 28.8 Tbps switches featuring the P200 are slated to arrive, with 51.2 Tbps models due before the end of the year.
Liquid Cooling for High-Density Switches
AI-class processing generates more heat than air cooling alone can handle, so Cisco’s top-end switches incorporate liquid cooling. Much like GPUs, a cold plate is mounted to the chip and connected to the rack or data center’s liquid cooling system. In addition to the G300 and P200, cold plates are used on other heat-intensive components in the Nexus 9000 and other high-end Cisco Silicon One switches.
“There are differing needs from different customers, so we offer a range of different switches – some that are 100% liquid cooled, some partially liquid cooled, and some only air cooled,” Kucharewski said.
Who Buys First: Hyperscalers Lead, Enterprises Follow #
Most data centers don’t need to purchase these high-end switches immediately. Many lack the AI workloads to justify them today. Boujelbene expects initial adoption from hyperscalers, as well as neocloud and sovereign AI cloud providers. Over time, these technologies will trickle down into more mainstream environments.
“The latest chips and servers are making the network a first-order design decision,” Boujelbene said. “Enterprises do not need to refresh everything overnight, but they do need to stop treating AI networking as an incremental upgrade. For serious GPU deployments, the network has to be designed as part of the compute system — not bolted on afterward.”
Competitors to Watch: Broadcom Tomahawk 6 and Nvidia Spectrum-6 #
Cisco is far from the only game in town. Broadcom and Nvidia are pursuing similar scale-out and scale-across targets.
Broadcom's Tomahawk 6 targets scale-out and scale-up AI networks used for training and inference. Offered with 512 or 1024 lanes and 64 ports, it enables 1.6 T Ethernet switching and routing. Like Cisco’s G300, it delivers up to 102.4 Tbps on a single chip. Its load balancing and congestion management aim to maximize network utilization and minimize job completion time.
Broadcom's Tomahawk 6 / BCM78910 Series. (Image: Broadcom) “By enhancing link stability and energy efficiency, we’re enabling smoother, more cost-effective AI model training,” said Near Margalit, Vice President and General Manager, Optical Systems Division, at Broadcom, in a press release. “We designed this platform to scale large AI clusters by delivering on the three imperatives for optical interconnect: higher model FLOPs utilization, reduced job interruptions, and improved cluster reliability.”
Nvidia’s Spectrum-6 is another 102.4 Tbps networking ASIC for Ethernet switches, built to accelerate AI and cloud workloads. As part of the Nvidia Spectrum-X Ethernet Platform, it ties together real-time predictive performance features to handle bursty traffic and minimize jitter.
Nvidia's Spectrum-6 chip. (Image: Nvidia) “Ethernet was designed primarily for enterprise applications and north-south traffic moving between users, servers and storage. It wasn’t created for the synchronized, collective-heavy communication patterns of gigascale AI; Spectrum-X Ethernet changes that,” said Scot Schulz, Senior Director of HPC and Technical Computing at Nvidia, in a blog post. “Purpose-built for AI, it transforms Ethernet into a high-performance scale-out fabric engineered to keep every GPU fed with data.”
Network Silicon on the Critical Path #
Data center designs have scaled up rapidly; rack densities exceeding 100 kW are increasingly common. The next frontier is scale-out and scale-across networking that sustains high utilization across racks and sites. Cisco’s Silicon One G300 addresses scale-out fabrics within the data center, while the P200 targets inter-data center links. Broadcom and Nvidia offer parallel 102.4 Tbps-class ASICs aimed at similar outcomes. The trajectory is clear: network silicon sits on the critical path to turning large GPU fleets into productive AI infrastructure.