Cisco targets high-speed AI networking with 102.4Tbps chip, liquid-cooled switches – News

Cisco Systems Inc. today introduced a new generation of networking silicon, systems and optical modules aimed at supporting hyperscale-class artificial intelligence infrastructure as AI workloads shift from model training toward broader inference and agentic applications.

At Cisco Live EMEA in Amsterdam, the company unveiled its Silicon One G300, a 102.4-terabit-per-second Ethernet switching chip designed to support massive “scale-out” AI clusters. Cisco said the chip will power new Nexus 9000 and Cisco 8000 switching platforms that can be deployed in high-density, air- and liquid-cooled configurations to reduce power consumption and simplify operations for organizations building large AI networks.

Cisco framed the announcements as a response to the expanding AI ecosystem, in which the demands of AI infrastructure are no longer confined to hyperscale cloud providers. Instead, enterprises, neocloud providers and sovereign cloud operators are increasingly investing in their own AI clusters and require more efficient networking to support graphics processing unit-heavy workloads.

“The last two or three years, we’ve mainly been focused on building out massive training clusters with hyperscalers,” said Kevin Wolterweber, senior vice president and general manager of Cisco’s data center and internet infrastructure business. “What we’re starting to see now is a shift toward agentic AI workloads, and more adoption within enterprise service providers and a broader customer base.”

Gigawatt-scale clusters

Cisco said the G300 chip is designed to improve the efficiency of AI clusters by maximizing GPU utilization and reducing delays caused by network congestion that stalls distributed training or inference jobs.

The chip supports a feature Cisco calls Intelligent Collective Networking, which combines a shared packet buffer, path-based load balancing and telemetry capabilities intended to better handle the bursty traffic patterns typical of AI workloads. Cisco claimed the approach can deliver a 33% increase in network utilization and a 28% reduction in job completion time compared with simulated environments using non-optimized path selection.

“The G300 represents more than a typical incremental upgrade,” said Sameh Boujelbene, vice president at Dell’Oro Group Inc. “Cisco is pushing intelligence directly into switching silicon, which addresses real pain points for large AI clusters like unpredictable east-west traffic, congestion hotspots and scaling limits in legacy fabrics.”

Wolterweber said the differentiator is less about raw bandwidth, since other chipmakers are also moving toward 100-terabit-class devices. Cisco sees its primary value being the intelligence built into the silicon.

“It’s about the efficiency of being able to load balance workloads across the network efficiently, doing very granular telemetry to feeding back to the job schedulers,” he said.

The G300 is also programmable, meaning customers can add new networking functionality after deployment as standards evolve.

Liquid cooling options

Alongside the silicon announcement, Cisco introduced new fixed and modular switching systems in its Nexus 9000 and Cisco 8000 product lines powered by the G300. Cisco said the systems deliver 102.4Tbps switching speeds and are aimed at customers building large AI fabrics.

The new systems will be available in both air-cooled and fully liquid-cooled designs. The company claimed the liquid-cooled configuration enables significantly higher bandwidth density and can improve energy efficiency by nearly 70% compared with prior generations, delivering the same bandwidth that previously required six systems.

Wolterweber said the move toward liquid-cooled switching aligns with the direction of GPU server platforms. “The future generations of GPUs are all going to be liquid cooled,” he said,  “so we’re starting to build liquid cooling into our switching devices as well.”

The Nexus switches run the company’s NX-OS software, a modular, Linux-based network operating system designed for Nexus-series data center switches and Multilayer Director Switch storage networking devices. The 8000 models can also support alternative network operating systems, including the open-source Sonic platform.

Cisco also introduced a new set of optical modules designed to support higher-density interconnects inside AI clusters.

The 1.6-terabit octal small form factor pluggable optical modules are aimed at switch-to-network interface card links and switch-to-server connections. OSFP is a type of transceiver that connects a network device to fiber or copper cable.

Cisco claimed new 800-gigabit linear pluggable optics can reduce optical module power consumption by 50% compared with retimed modules, reducing overall switch power by 30%.

Unified management

Cisco also updated its Nexus One management platform, positioning it as a unified operating model for AI networks that span on-premises and cloud deployments. The platform is designed to simplify the deployment and management of AI fabrics while improving observability and security.

New features include a unified fabric capability, application programming interface-driven automation and AI job observability that correlates network telemetry with AI workload behavior. Cisco also added native integration with the Splunk data analytics platform it acquired in 2024, enabling network administrators to analyze network telemetry where the data resides rather than moving it into external systems

Wolterweber said the integration is designed to address the cost and complexity of ingesting large volumes of network data into Splunk. “Organizations can’t ingest everything they have,” he said.

The move brings together Nexus Dashboard and Nexus Hyperfabric into a single platform that can manage both on-premises and cloud-managed devices, while also supporting multiple operating systems including Sonic.

Dell’Oro’s Boujelbene said Cisco is making inroads into an AI networking market that has been dominated by Nvidia Corp., Celestica Inc., Juniper Networks Inc. and Arista Networks Inc.

“Few competitors can match Cisco’s integrated stack of silicon, systems, optics, software and operations tooling, which matters for customers seeking end-to-end visibility and unified support,” she said. “Cisco’s deep penetration in enterprise data centers also gives it leverage, as IT teams often prefer incremental evolution of vendors they already trust rather than introducing entirely new suppliers.”

Wolterweber said the growth of agentic AI workflows will drive major changes in how networks are secured and monitored, since enterprises will need to manage identities and permissions for large numbers of autonomous agents operating continuously.

“We expect to see a lot more utilization of the network itself, because now you’re going to have a multiplicative effect of agents doing things for you,” he said. That means security models will need to evolve beyond centralized firewalls, with enforcement distributed across the network.

“We’re embedding a lot of the technologies that we used to put into things like centralized firewalls around the network to allow us to do policy enforcement,” Wolterweber said.

Cisco said many of the products and features announced today are in development and will be made available as they are finalized.

Image: Cisco