Nvidia expects to ship around 5 million Blackwell GPUs by 2025. But before these GPUs can train GPT, Gemini or Llama they need to be connected — and that’s becoming a big business for Ethernet switches vendors like Cisco Arista HPE and Nvidia.
On Wednesday, Cisco CEO Chuck Robbins told analysts during the company’s earnings conference for its fourth quarter of fiscal year 2025 that the AI infrastructure orders received by web-scale clients exceeded $800 million in the fourth quarter. He added “more than double our original $1 billion target stated in Q4 of fiscal year 2024,” .
This growth is based on a simple economics. Ethernet switch vendors can sell three to five switch port for every B200 or GB200 graphics processor sold. Actual numbers can vary depending on port speeds, resilience, and the size of the cluster. With modern 51.2Tb/s switch and 400GbE link, a leaf layer and spin layer is sufficient for up to 8,192 GPUs. On the high-end, that’s 192 switches.
For clusters above that size, the switch count will explode as a three-tiered fat tree topology becomes necessary to provide non-blocking connectivity in the AI back-end network. For a cluster of 128,000 GPUs, you’re looking at 5,000 switches – or 10,000 if you have to settle for an older 25.5Tb/s kit.
In most cases, the switch count will be higher because on top of the larger back-end network that connects all the nodes, you’ll have a smaller network for moving data.
There are also the copper cables and optics required to connect these networks. We’re talking about more than a million pluggable optical cables between the front-end network and the back-end network for a 128,000 GPU Cluster.
Robbins claims that optics accounted roughly for a third (800 million dollars) of the AI networking revenues during Q4.
Nvidia’s desire to bring photonic switches into the market is understandable, as they will reduce the number of optical transmitters required to connect these transceivers.
There’s plenty to go around
Hyperscalers don’t care who or where the kit is from as long as it works. Ethernet is a standard, so you can mix and match as much as you want.
The market for AI networks grows about as fast as Nvidia or AMD can produce GPUs.
Cisco’s $2 Billion in AI Infrastructure revenue is currently driven by webscale clients, but Robbins was optimistic about the opportunities for traditional enterprises. Cisco is in a better position to upsell its customers on value ads and service contracts as well as software subscriptions. This is more than datacenter providers can do. Robins made his opinion known during this week’s conference call.
Robbins wants to tap into the emerging Neocloud segment where, according to him, Cisco has several large AI network deals in the works.
The growth for Arista, a long-time competitor, is even more evident. Earlier this year, the company projected a full-year revenue of $8.75 billion. Of that, AI-related sales were expected to make up about 17 percent.
“As large language models continue to expand into distributed training and inference use cases, we expect to see the back end and the front end converge,” Ceo jayshree ulla said. “This will make it increasingly difficult to parse the back end and the front end precisely in the future, but we do expect an aggregate AI networking revenue to be ahead of $1.5 billion in 2025 and growing in many years to come.”
Juniper Networks’ AI infra demand is also likely to be high, but to what extent is difficult to assess given its merger with HPE. Arista promotes Ethernet for AI and downplays the effect of tariffs. Broadcom asks, with Tomahawk Ultra, who needs UALink when there’s Ethernet? The GPUs of the company aren’t only driving Ethernet switch sales. It is now one of the most prominent names in AI networking, thanks to its acquisition of Mellanox in 2019 for $6.9 Billion. Colette Kress, Nvidia’s CFO, reported that in Q1, network sales were up 64 percent from quarter to quarter. That’s about $5 billion. It’s important to note that not all of this is Ethernet switching. Nvidia sells InfiniBand switches, ConnectX NICs and Spectrum-X Ethernet kits in addition to its Spectrum-X Ethernet Kit.
But if the AI bubble doesn’t burst there should be enough cash for networking vendors. Dell’Oro predicts that AI networks will drive $80 billion worth of network sales by 2030. (r)
