-
Data centers will need a lot more fiber-rich interconnects to accommodate the large language models that enable artificial intelligence.
-
Corning is looking to use thinner fibers to allow it to cram 24 fibers in half the amount of space currently required.
-
According to Dell’Oro Group data center capex is expected to grow roughly 6% year on year to $266 billion in 2023.
It’s no secret that artificial intelligence (AI) is poised to take over the data center. In fact, we’ve been banging the drum about how liquid cooling is emerging as a way to solve the power and heat density issues that high-powered GPU-based AI servers will bring with them. But it turns out data centers will also need more fiber – lots more – to cope with the demands of AI.
“We believe artificial intelligence changes the game for the data center market,” Mike O’Day, CTO at Corning Optical (NYSE: GLW), told Silverlinings.
“At the end of the day, the data centers are going to have to add a lot more fiber-rich interconnects to accommodate the large language models that enable artificial intelligence to come true,” he continued. “Our initial estimates are at least five times more optical connections are going to be required in a data center as a result of the applications are going to access and want to use artificial intelligence.”
According to O’Day, the reason for all this new fiber is the need to create an entirely new network to connect the AI servers to one another. Today, data center operators have fiber for what they call a front-end network or traditional cloud computing network. For AI, though, they’re looking at bringing new back-end networks or dedicated AI networks online.
But there’s a problem. While data centers are huge buildings, O’Day noted there’s not a ton of unused space within them. So, hyperscalers and their vendors are left not only trying to solve for heat and power density issues associated with high performance computing, but also how to cram more fiber into less space.
Racing the AI clock
“In a new data center, you can plan better and logically plan for the space you need. But most of our customers have existing data centers and the AI market is now, not five years from now. So, they’re having to find space and find power,” O’Day said.
Corning is racing the clock to meet demand from hyperscale customers, with O’Day noting that it expects a fair amount of capex spending on the fiber front to come next year despite an overall capex slowdown across the industry.
O’Day said solving the density issue will be at the core of Corning’s second generation of data center solutions. Specifically, it is looking to use thinner fibers to allow it to cram 24 fibers in half the amount of space currently required.
It’s also exploring products that will help data center operators slash installation time, by removing the need for field splicing, he added.
He added these solutions will be commercialized by the end of this year to support the expected builds coming next year.
Forecasting for the data center
According to Dell’Oro Group data center capex is expected to grow roughly 6% year on year to $266 billion in 2023. That growth is slower than the 17% the segment saw between 2021 and 2022, but Dell’Oro noted the upward trajectory will continue despite the near-term deceleration, hitting more than $500 billion by 2027. This forecast figure includes servers, storage systems, networking equipment such as switches, routers, and transport systems, physical infrastructure, other data center facilities and supply chain costs.
Dell’Oro Senior Research Director Baron Fung told Silverlinings he expects servers “will have the biggest positive influence [on capex] compared to other technology areas.” That’s in part because the GPU-based servers used for AI “cost a lot more than general-purpose servers, which are servers without accelerators.”
“We are projecting capex on these accelerated [servers] to grow at a five-year CAGR in the range of 30-40%, compared to less than 15% for general-purpose infrastructure,” he said.
The proliferation of accelerated server infrastructure has implications for the rest of the data center.
“There needs to be continuous development not just with accelerated servers, but also with supporting infrastructure such as networking and DCPI,” Fung concluded. “These accelerated servers are not a drop in solution and requires the supporting infrastructure to be upgraded as well.”
Learn more about data centers by watching our Cloud Data Center Strategies Summit on demand.