Ethernet switch vendors like Cisco are riding high on AI network economics
(2025/08/15)
- Reference: 1755286209
- News link: https://www.theregister.co.uk/2025/08/15/ethernet_ai_gpus/
- Source link:
Nvidia is expected to ship somewhere north of 5 million Blackwell GPUs in 2025. But before those GPUs can train the next GPT, Gemini, or Llama, they need to be networked — and that's quickly becoming big business for Ethernet switch vendors like Cisco, Arista, HPE ... and Nvidia itself.
Speaking with analysts on the company's earnings call for the fourth quarter of its 2025 fiscal year on Wednesday, Cisco CEO Chuck Robbins [1]boasted that AI infrastructure orders received from web-scale customers exceeded $800 million during the fourth quarter and $2 billion for the 2025 fiscal year. That's "more than double our original $1 billion target stated in Q4 of fiscal year 2024," he added.
The economics behind this growth is actually pretty simple. For every B200 or GB200 GPU sold, Ethernet switch vendors have the opportunity to sell around three to five switch ports.
[2]
Actual numbers vary depending on port speed, resiliency ,and the size of the overall cluster. Up to 8,192 GPUs, only a leaf and spin layer are required when using modern 51.2Tb/s switches and 400GbE links. At the high end, that works out to 192 switches.
[3]
[4]
But for clusters exceeding that, the switch count explodes as a three-tier fat tree topology is now required to provide full-non-blocking connectivity for the AI back-end networks. For a 128,000 GPU cluster you're talking on the order of 5,000 switches — 10,000 if you're forced to settle for older 25.5Tb/s network kit.
And in most cases the switch count is going to be higher, as on top of the back-end network connecting all the nodes together, you'll also have a smaller front-end network for moving data around.
[5]
Then there are the optics and direct-attach copper cables required to stitch these networks together. For a 128,000 GPU cluster, we're potentially talking in excess of a million pluggable optics between the front-end and back-end networks.
According to Robbins, optics accounted for roughly a third of its $800 million AI networking revenues in Q4.
It's really no wonder why Nvidia is so keen to bring photonic switches to market, as they'll dramatically cut down on the number of optical transceivers required to stitch these transceivers together.
Plenty to go around
When it comes to AI networks, hyperscalers aren't particularly picky about who or where the kit comes from so long as it works. Remember, Ethernet is a standard which means that, in theory, you should be able to mix and match to your heart's content.
Having said that, the market for AI networks is growing about as fast as Nvidia and AMD can pump out GPUs.
[6]
Most of Cisco's $2 billion in AI infrastructure revenue is being driven by webscale customers for the moment, but Robbins waxed optimistic about its opportunities in more traditional enterprises. Here, Cisco is better positioned to upsell customers on value ads, service contracts, and software subscriptions than it can do with datacenter providers.
"In the enterprise specifically, while still early, AI orders are ramping and we have a growing pipeline in the hundreds of millions as these customers look to Cisco to provide simple scalable and secure solutions," Robins opined during this week's call.
Robbins is also keen to tap into the emerging neocloud segment, where he says Cisco has several large AI networking deals in the works.
For long-time rival Arista, the growth is even clearer. Earlier this month, the company [7]forecast full year revenues of $8.75 billion, of which AI related sales were expected to account for 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 is presumably riding high on AI infra demand as well, though to what degree is harder to assess amid its merger with HPE.
[8]Sysadmin cured a medical mystery by shifting a single cable
[9]The plan to make all networks optical is about to take two big steps forward
[10]Arista pushes Ethernet for AI, downplays effect of tariffs
[11]With Tomahawk Ultra, Broadcom asks who needs UALink when there's Ethernet?
Then there's the 800-pound gorilla of AI infrastructure, Nvidia. The company's GPUs aren't just driving Ethernet switching sales. It's now one of the biggest names in AI networking itself, thanks in large part to its acquisition of Mellanox for $6.9 billion back in 2019.
In Q1 alone, Nvidia CFO Colette Kress [12]reported network sales up 64 percent quarter over quarter to roughly $5 billion. To be clear, not all of that is Ethernet switching. In addition to its Spectrum-X Ethernet kit, Nvidia also sells a lot of InfiniBand switches and ConnectX NICs.
But short of an [13]AI bubble burst , there should be plenty of cash to go around for networking vendors. Dell'Oro group [14]predicts , AI networks will drive nearly $80 billion in network sales by 2030. ®
Get our [15]Tech Resources
[1] https://seekingalpha.com/article/4813420-cisco-systems-inc-csco-q4-2025-earnings-call-transcript
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[7] https://www.theregister.com/2025/08/06/arista_pushes_ethernet_for_ai/
[8] https://www.theregister.com/2025/08/15/on_call/
[9] https://www.theregister.com/2025/08/07/iown/
[10] https://www.theregister.com/2025/08/06/arista_pushes_ethernet_for_ai/
[11] https://www.theregister.com/2025/07/15/broadcom_ethernet_scale_up/
[12] https://seekingalpha.com/article/4790673-nvidia-corporation-nvda-q1-2026-earnings-call-transcript
[13] https://www.theregister.com/2025/08/15/boy_riding_bubble_realizes_what/
[14] https://www.delloro.com/news/ethernet-is-winning-the-war-against-infiniband-in-ai-back-end-networks/
[15] https://whitepapers.theregister.com/
Speaking with analysts on the company's earnings call for the fourth quarter of its 2025 fiscal year on Wednesday, Cisco CEO Chuck Robbins [1]boasted that AI infrastructure orders received from web-scale customers exceeded $800 million during the fourth quarter and $2 billion for the 2025 fiscal year. That's "more than double our original $1 billion target stated in Q4 of fiscal year 2024," he added.
The economics behind this growth is actually pretty simple. For every B200 or GB200 GPU sold, Ethernet switch vendors have the opportunity to sell around three to five switch ports.
[2]
Actual numbers vary depending on port speed, resiliency ,and the size of the overall cluster. Up to 8,192 GPUs, only a leaf and spin layer are required when using modern 51.2Tb/s switches and 400GbE links. At the high end, that works out to 192 switches.
[3]
[4]
But for clusters exceeding that, the switch count explodes as a three-tier fat tree topology is now required to provide full-non-blocking connectivity for the AI back-end networks. For a 128,000 GPU cluster you're talking on the order of 5,000 switches — 10,000 if you're forced to settle for older 25.5Tb/s network kit.
And in most cases the switch count is going to be higher, as on top of the back-end network connecting all the nodes together, you'll also have a smaller front-end network for moving data around.
[5]
Then there are the optics and direct-attach copper cables required to stitch these networks together. For a 128,000 GPU cluster, we're potentially talking in excess of a million pluggable optics between the front-end and back-end networks.
According to Robbins, optics accounted for roughly a third of its $800 million AI networking revenues in Q4.
It's really no wonder why Nvidia is so keen to bring photonic switches to market, as they'll dramatically cut down on the number of optical transceivers required to stitch these transceivers together.
Plenty to go around
When it comes to AI networks, hyperscalers aren't particularly picky about who or where the kit comes from so long as it works. Remember, Ethernet is a standard which means that, in theory, you should be able to mix and match to your heart's content.
Having said that, the market for AI networks is growing about as fast as Nvidia and AMD can pump out GPUs.
[6]
Most of Cisco's $2 billion in AI infrastructure revenue is being driven by webscale customers for the moment, but Robbins waxed optimistic about its opportunities in more traditional enterprises. Here, Cisco is better positioned to upsell customers on value ads, service contracts, and software subscriptions than it can do with datacenter providers.
"In the enterprise specifically, while still early, AI orders are ramping and we have a growing pipeline in the hundreds of millions as these customers look to Cisco to provide simple scalable and secure solutions," Robins opined during this week's call.
Robbins is also keen to tap into the emerging neocloud segment, where he says Cisco has several large AI networking deals in the works.
For long-time rival Arista, the growth is even clearer. Earlier this month, the company [7]forecast full year revenues of $8.75 billion, of which AI related sales were expected to account for 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 is presumably riding high on AI infra demand as well, though to what degree is harder to assess amid its merger with HPE.
[8]Sysadmin cured a medical mystery by shifting a single cable
[9]The plan to make all networks optical is about to take two big steps forward
[10]Arista pushes Ethernet for AI, downplays effect of tariffs
[11]With Tomahawk Ultra, Broadcom asks who needs UALink when there's Ethernet?
Then there's the 800-pound gorilla of AI infrastructure, Nvidia. The company's GPUs aren't just driving Ethernet switching sales. It's now one of the biggest names in AI networking itself, thanks in large part to its acquisition of Mellanox for $6.9 billion back in 2019.
In Q1 alone, Nvidia CFO Colette Kress [12]reported network sales up 64 percent quarter over quarter to roughly $5 billion. To be clear, not all of that is Ethernet switching. In addition to its Spectrum-X Ethernet kit, Nvidia also sells a lot of InfiniBand switches and ConnectX NICs.
But short of an [13]AI bubble burst , there should be plenty of cash to go around for networking vendors. Dell'Oro group [14]predicts , AI networks will drive nearly $80 billion in network sales by 2030. ®
Get our [15]Tech Resources
[1] https://seekingalpha.com/article/4813420-cisco-systems-inc-csco-q4-2025-earnings-call-transcript
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/networks&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aJ-t8xQsUo37S8glt1tZeAAAANQ&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[7] https://www.theregister.com/2025/08/06/arista_pushes_ethernet_for_ai/
[8] https://www.theregister.com/2025/08/15/on_call/
[9] https://www.theregister.com/2025/08/07/iown/
[10] https://www.theregister.com/2025/08/06/arista_pushes_ethernet_for_ai/
[11] https://www.theregister.com/2025/07/15/broadcom_ethernet_scale_up/
[12] https://seekingalpha.com/article/4790673-nvidia-corporation-nvda-q1-2026-earnings-call-transcript
[13] https://www.theregister.com/2025/08/15/boy_riding_bubble_realizes_what/
[14] https://www.delloro.com/news/ethernet-is-winning-the-war-against-infiniband-in-ai-back-end-networks/
[15] https://whitepapers.theregister.com/