Cisco has officially stepped into the fast-growing competition to build better connections between AI data centres. On October 8, the company introduced its new —a powerful piece of hardware designed to link AI workloads across multiple locations.
At the heart of this system is Cisco’s new Silicon One P200 chip. It’s their answer to a growing problem in the AI world: data centres are running out of space, power, and cooling capacity. So instead of building bigger, companies are now trying to build smarter—by spreading AI tasks across several facilities and connecting them efficiently.
A Three-Way Tech Showdown
Cisco isn’t the only one chasing this goal. Broadcom made its move in August with its Jericho 4 chip, offering high bandwidth and smart buffering to handle heavy traffic. Then Nvidia followed with its Spectrum-XGS network, aimed at scaling AI across multiple centres. Now, Cisco’s entry sets up a three-way battle between these tech giants.
Why AI Needs More Than One Building
Modern AI systems—like those used to train large language models—require thousands of processors working together. That means lots of heat, lots of electricity, and lots of space. But data centres are hitting their limits. They can’t expand fast enough, and power grids can’t keep up.
“AI compute is outgrowing the capacity of even the largest data centre,” said Martin Lund, Cisco’s EVP of Common Hardware Group. That’s why companies are turning to “scale-across” strategies—spreading AI workloads across multiple centres, sometimes hundreds of miles apart.
Old Routers Can’t Keep Up
AI traffic is different. It comes in bursts—huge waves of data followed by quiet periods. Traditional routers weren’t built for this kind of load. They either focus on speed or traffic control, but not both. And they often use too much power.
Cisco’s 8223 system changes that. It’s compact, energy-efficient, and packed with 64 ports of 800-gigabit connectivity. It can handle over 20 billion packets per second and scale up to three exabytes of bandwidth. Its deep buffering—thanks to the P200 chip—helps absorb traffic spikes, keeping networks smooth and GPUs busy.
Real-World Use and Adoption
Big names are already on board. Microsoft has been using Cisco’s Silicon One chips and says they’ve helped expand across different roles in data centres and AI environments. Alibaba Cloud plans to use the P200 to upgrade its core network. And Lumen is exploring how the 8223 could improve its services.
Built for the Future
One standout feature of the P200 chip is its programmability. As AI evolves, so do its networking needs. Instead of replacing hardware, companies can update the chip to support new protocols—saving time and money.
Security Matters Too
Connecting data centres over long distances raises security concerns. Cisco’s 8223 includes encryption that’s designed to withstand future threats, including those from quantum computing. It also works with Cisco’s monitoring tools to quickly spot and fix issues.
Can Cisco Win the Race?
Broadcom and Nvidia have already made strong moves, but Cisco has some advantages: deep experience in networking, a mature chip portfolio, and strong relationships with major cloud providers. The 8223 supports open-source software now, with more options coming soon.
Whether Cisco becomes the go-to choice for AI data centre interconnects is still up in the air. But one thing’s clear—the need to connect AI systems across multiple locations is only going to grow. And the winner may be the company that offers not just great hardware, but a full ecosystem of tools, support, and flexibility.
Want to dive deeper? Check out:
