- Hyperscalers are looking to connect GPUs across multiple data center campuses
- Ciena's business is benefitting from these networking needs
- Inferencing could propel future cloud growth for Ciena
Forget one ring to rule them all and bring on the rings of power. As artificial intelligence training clusters grow, hyperscalers are having to connect GPUs across multiple data center campuses. And that’s good news for the likes of networking vendors like Ciena.
In addition to general cloud growth, CEO Gary Smith told Fierce Ciena is benefitting from “dedicated provisioning for machine-to-machine [deployments] where the GPUs can’t be in the same campus.”
“We’re seeing it across all the major four players,” he said, adding they’re looking for “very high speed connectivity” and “clearly dedicated capacity” to support those needs. For Ciena, that means orders for its WaveLogic 6 platform and reconfigurable line system.
Smith’s comments track with observations made in a note from SemiAnalysis this week. The firm pointed out that as training clusters have ballooned to include 100,000+ GPUs, leading AI companies including Google, OpenAI and Anthropic have had to expand their training setups to include multiple data center campuses.
“Gemini 1 Ultra was trained across multiple datacenters,” they wrote. “In 2025, Google will have the ability to conduct Gigawatt-scale training runs across multiple campuses.”
And just how far apart are the data centers that make up these new mega-clusters?
Smith said Ciena has seen hyperscalers look at options both within the same metro region as well as across the country.
Indeed, SemiAnalysis noted of Google Cloud’s infrastructure: “There are three sites ~15 miles from each other, (Council Bluffs, Omaha, and Papillon Iowa), and another site ~50 miles away in Lincoln Nebraska… combining all four campuses will form a GW-scale AI training cluster by 2026.”
The analysts added Google could potentially connect its data center campus hubs in Nebraska/Iowa and Ohio to deliver “multiple gigawatts of power for training a single model.”
Meanwhile, Microsoft and OpenAI are collaborating to interconnect several ultra-large data center campuses together to run “giant distributed training runs,” the analysts wrote.
Sun peeks through the cloud
Both revenue and profit fell year on year in Ciena’s fiscal Q3, a slide Smith attributed to service providers going through an absorption cycle. This is expected to abate and normalize sometime next year, he said. But notably, service providers are accounting for less and less of Ciena’s revenue these days, a trend that’s expected to continue moving forward.
According to Smith, about 40-50% of Ciena’s revenue now comes from cloud providers. That’s up from 1-2% a decade ago and 10-15% just two or three years ago. Eventually, Smith said he expects cloud providers will overtake service providers as the company’s largest revenue generating engine.
“It’s a very positive set of dynamics around traffic growth, both in cloud, some of the consumption on the AI piece but specifically machine-to-machine,” Smith told us. “And that’s before seeing any traffic growth in the inference. That’s going to be when you see the real monetization of the large language models. “That’s all in front of us and that will obviously probably drive large cloud growth.”