The next limit on AI scale may not sit inside a processor. Data centres are running out of floor space, electrical capacity, and cooling, forcing companies to spread enormous workloads across separate facilities. Once that happens, the network between buildings becomes part of the computer—and a slow link can leave costly accelerators waiting.
Cisco has entered the emerging “scale-across” market with the 8223 routing system and its Silicon One P200 chip. The fixed router delivers 51.2 terabits per second in a compact three-rack-unit chassis and is designed specifically for the immense, uneven traffic created by distributed AI training and inference.
The company is joining Broadcom and Nvidia in a contest over the connective tissue of future AI infrastructure. Broadcom’s Jericho 4 and Nvidia’s Spectrum-XGS address the same problem, making inter-data-centre networking a new strategic layer rather than a background utility.
Distance changes the engineering equation. Inside one building, specialised links can be extremely fast and tightly controlled. Across cities, traffic encounters optics, routing, encryption, and more opportunities for congestion or failure. The network must preserve enough predictability that distributed compute still behaves like a coordinated system.
Every extra millisecond and dropped packet can be multiplied across thousands of expensive processors waiting for the next synchronised step in a long-running job.
AI Is Becoming Too Large for One Building
Traditional scaling adds more capability to one system or connects more systems inside the same facility. Both approaches eventually collide with physical limits. Large clusters consume enormous power and generate heat that local grids and cooling plants cannot always support. Cisco executive Martin Lund argues that reliable links may need to join data centres hundreds of miles apart.
AI traffic makes that difficult. Training generates sudden waves of data followed by quieter periods. A router that cannot absorb a burst creates congestion, stretching job completion time while GPUs sit underused. Conventional equipment often compromises among raw throughput, traffic management, and energy use; scale-across networks need all three.
Deep Buffers Catch the Traffic Storm
The 8223 provides 64 ports of 800-gigabit connectivity and can process more than 20 billion packets per second. The P200’s deep buffers act like reservoirs, temporarily holding traffic during a surge instead of allowing packets to overwhelm the link. Cisco also claims switch-like power efficiency, an important consideration when every additional watt competes with compute.
Support for 800G coherent optics extends connections to 1,000 kilometres, making regional distribution possible. Line-rate encryption uses post-quantum-resilient algorithms, while integration with Cisco observability tools helps operators investigate performance. Security cannot be an afterthought when model data and internal traffic leave one controlled campus and cross a wider network.
Programmability and Ecosystem Will Decide the Contest
The P200 is programmable, allowing support for emerging protocols without replacing the silicon. That flexibility matters because AI networking standards are still changing while routing hardware represents a long-lived capital investment. The system initially supports open-source SONiC, with Cisco’s IOS XR planned later, and the chip will appear in fixed, modular, and Nexus platforms.
Microsoft already uses the broader Silicon One architecture across data-centre, wide-area, and AI environments. Alibaba Cloud plans to use P200 devices to extend its eCore design, while Lumen is exploring the technology. These relationships give Cisco practical routes into large networks, but Broadcom and Nvidia bring their own established ecosystems.
The winner will not be determined by 51.2 terabits alone. Operators need software, optics, telemetry, security, support, and evidence that expensive training jobs remain efficient across distance. As AI escapes the walls of one facility, the best router will be the one that makes several data centres behave least like several data centres.
