Cisco has introduced the Cisco 8223, a fixed-routing system with 51.2 Tbps designed specifically for the challenge of distributed AI: interconnecting multiple data centers separated by hundreds of kilometers with more bandwidth, better energy efficiency, and end-to-end security. At the core of the device is the Silicon One P200, Cisco’s new generation ASIC with deep buffers, programmability, and a design aimed at “scale-across” (scaling between data centers) when upward (more capacity per system) or outward (within a single campus) scaling is no longer feasible.
“AI computing is surpassing the capacity of a single data center, requiring reliable and secure connection of distant centers,” summarized Martin Lund, EVP of Cisco’s Common Hardware Group. “With the 8223 and the new P200, we provide the bandwidth, scale, and security needed for this distributed architecture.”
Why it matters: from “scale-up/scale-out” to scale-across
Large AI clusters have reached power and space limits. Even optimizing within the same site (vertical and horizontal scale), the growth of models and datasets demands distributing loads across multiple data centers. This pattern increases pressure on the Data Center Interconnect (DCI) and metro/long-haul networks: any bottleneck hampers performance and TCO.
The 8223 addresses this point:
- Capacity and density: 64 ports of 800 GE in a 3RU chassis, processing >20 billion packets/sec with coherent 800 GE optics for DCI/metro up to 1,000 km.
- Efficiency: switch-class power consumption on a fixed routing platform with deep buffers— critical for absorbing bursts that generate training and AI checkpoints.
- Programmability and future-proofing: the P200 supports new protocols/standards without hardware changes, accelerating adoption of features as AI traffic evolves.
- Security: line-rate encryption with post-quantum resilient algorithms, built-in safeguards, and observability for faster detection and response.
8223: specifications and approach
- Throughput: 51.2 Tbps in a 3RU (fixed system).
- Ports: 64× 800 GE, supporting coherent 800 GE optics for DCI/metro (~1,000 km).
- Performance: >20 Gpps (packets per second).
- Silicon: Cisco Silicon One P200 (deep-buffer routing silicon, programmable).
- Aggregated scale: architectures exceeding 3 Exabits/sec (network level).
- Software: SONiC initial release; IOS XR “on the horizon.” The Nexus family will add systems with NX-OS based on P200 “soon”.
- Security: line-rate post-quantum encryption, HW/SW safeguards, and integration with Cisco’s observability platforms.
Use cases
- AI/ML backbone between data centers: distributed training, inference, and data repositories across multiple campuses.
- High-capacity DCI / Metro with coherent 800 GE optics.
- Hype-scale WANs for operators requiring deep buffers to absorb bursts without performance degradation.
Silicon One P200: beyond the 800 GE port
Since its launch in 2019, Silicon One has established itself as a unified architecture (DC, WAN, AI/hyperscale, enterprise, SP). The P200 strengthens this lineup with:
- Deep buffers: essential for AI traffic, with frequent peaks and micro-congestions.
- Programmability (e.g., P4) to introduce new protocols and features without cannibalizing hardware.
- Energy efficiency: designed for power-constrained environments; the 8223 aims to maximize Tbps per watt and Tbps per RU.
Cisco will also deploy the P200 in modular and disaggregated chassis platforms, maintaining architectural consistency across small networks to macro networks. The company indicates its Portfolio Nexus will support NX-OS-based systems with P200 soon.
End-to-end security
Connecting distributed data centers with sensitive AI information requires encryption and control without performance sacrifice. The 8223 includes:
- Line-rate encryption with post-quantum resistant algorithms (PQC).
- Defense-in-depth: HW/SW safeguards, continuous monitoring, and telemetry integrated into Cisco’s observability platforms for quicker detection and remediation.
Operational flexibility: SONiC today, IOS XR and NX-OS tomorrow
The company will launch the 8223 with SONiC (open-source) for clients seeking disaggregation and “a la carte” operation. In subsequent phases, IOS XR will become available, and the Nexus portfolio will add P200 systems with NX-OS. The idea: support both disaggregated and integrated deployment options, unified by a common silicon architecture.
Industry voices
- Microsoft (Azure Networking) highlighted that the growth of cloud + AI demands more buffering to handle bursts, and praised the P200’s flexibility in maintaining a common ASIC architecture across DCI, WAN, and AI/ML.
- Alibaba Cloud describes the P200 as the first 51.2 T routing ASIC with high bandwidth, lower power consumption, and full programmability, aligned with its eCore architecture to replace traditional chassis routers with clusters of P200.
- Lumen explores how the 8223 might fit into its high-scale multi-cloud alongside the 8000 series, Silicon One, and Cisco’s pluggable optics.
- Moor Insights & Strategy regards this as a significant step in AI-distributed networking: the first fixed router 51.2 T aimed at secure, “scale-across” deployment.
Critical reading: implications and questions for 2026
- Consumption and density: the promise of “switch efficiency in fixed router” will be crucial in W/RU-constrained rooms. Tbps/W and Tbps/RU will be key metrics comparing with other 51.2 T systems.
- Buffers vs latency: deep buffers help absorb AI bursts, but architecture must balance with low queues to prevent latency inflation on sensitive traffic.
- Coherent 800 GE optics: the availability of coherent modules and their eco-system (cost, power, interoperability) will be decisive for 1,000 km DCI/metro deployments.
- Multi-OS operation: coexistence (SONiC / IOS XR / NX-OS) broadens options for clients with varied preferences and tools, but requires consistent automation/telemetry standards.
- PQC security: accelerating post-quantum algorithms at line rate is strategic; monitoring adoption curves and computational costs will be important.
Availability
Cisco states that the 8223 is already shipping to initial hyperscalers, and the P200 will also be deployed in modular platforms and disaggregated chassis. The software roadmap begins with SONiC; IOS XR and NX-OS will follow within their respective families.
In a nutshell
With the 8223 and the Silicon One P200, Cisco aims to make inter-data center networks cease being the bottleneck for AI: 51.2 Tbps per fixed system, 64× 800 GE, coherent optics up to 1,000 km, deep buffers, PQC encryption, and programmability all serve a clear goal: enabling distributed AI clusters to perform like a single super data center.
via: newsroom.cisco