Cisco Unveils Deep Buffer Router to Interconnect AI Data Centres

• 13 Oct 2025
• TM Team
• 0 Comments

The AI factory buildout is exposing all kinds of bottlenecks, including in the WANs that connect far-flung data centers. Cisco is addressing this market with a new router, dubbed the Cisco 8223, which it says is the industry’s first fixed 51.2Tb router supporting deep buffer routing that allow hyperscalers to lash their AI factory networks together, or a data center interconnect.

We’re in the midst of a massive effort by hyperscalers to build as many huge data centers to house racks of GPUs for AI training and inference, as fast as they can. OpenAI is party to the Project Stargate AI alliance, along with Oracle and SoftBank, which has committed to building up to 10 gigawatts worth of data centers (equal to the electricity needed to power all the GPUs and systems at full build-out), at a cost of $500 billion. The group is currently building its first data center in Abilene, Texas, dubbed Stargate I, which will consume 5 gigawatts of energy.

Not to be outdone, Facebook parent Meta is investing in giant data centers of its own. The first one, dubbed Prometheus, is a 1-gigawatt campus in Columbus, Ohio, which is currently being built and is due to come online in 2026. It will follow that up with a much larger data center, dubbed Hyperion, that will consume 5 gigawatts of energy when it’s completed in Richland Parish, Louisiana.

Nvidia will be called on to provide much of the computing horsepower in these data centers with its current Blackwell line of GPUs. But there will be a mix of other AI accelerators, including AMD’s forthcoming MI450 GPU, which OpenAI committed to buying in great numbers this week in a deal that will see OpenAI buying 6 gigawatts of AMD GPUs in exchange for a 10% equity stake in the company. Meta, also a big Nvidia customer, also recently nabbed Rivos, which was developing a promising new AI accelerator based on RISC-V technology and was reportedly worth $2 billion.

There will be lots of big and fast storage purchased to go along with these massive installations of GPU and AI accelerators, as we’ll be writing about in the weeks to come (stay tuned). And of course the network gear providers will also have something to crow about, as Meta, OpenAI, and others invest massive sums in copper wiring, and probably a bit of cooler and more energy efficient co-packaged optics (CPO), to string all these GPUs, CPUs, and storage devices into a coherent whole. And let’s not forget the massive amounts of passive and active cooling that will be needed to keep all this silicon, copper, and glass from melting down–let along the gigawatts worth of power to make it all hum.

With all this exciting AI infrastructure going into place, it’s easy to lose sight of the need for external networking. These data centres are being built in some fairly remote locations, where land and electricity is cheaper, so the hyperscalers will need some pretty big network pipes, or data centre interconnects (DCIs), to connect them to their other massive data centres for AI training and inference workloads.