Data centers have long relied on fiber to transmit signals. But the compute necessary to meet AI’s demand requires more fiber than ever before.
AI pulls from hundreds of thousands of servers and chips. Lauren Clark, a managing director in Brown Gibbons Lang & Company’s infrastructure vertical, told Business Insider that the higher density of fiber, the faster computers can talk to each other— and the faster a large language model can provide a response.
And as AI models become more complex and sophisticated, they require denser optical networks, said Mike O’Day, senior vice president and general manager of Corning Optical Communications.
Without sufficient fiber, AI chatbots stall and show users a busy signal or error, which is already happening today. Longer-term, a lack of fiber could mean pricier tokens and longer adoption cycles because supplies are constrained, said Gordon Bell, a partner and principal leading digital infrastructure engagements at EY-Parthenon.
“You can’t have AI for an end user without a powered data center. You can’t have it without connectivity,” Bell said. “They all need to come together at the same place at the same time in order for everything to work.”
As AI models become more complex and enterprises and consumers use the technology daily, AI tools will require more infrastructure to respond to chatbot queries. To meet demand, hyperscalers — the companies that provide cloud computing and data storage such as Amazon and Meta — are locking in agreements with fiber manufacturers such as Corning. In turn, fiber suppliers are expanding manufacturing capacity, while some startups are developing technologies to send data at faster speeds. Both strategies reflect the industry’s bet that fiber will be critical to keeping pace with AI’s growth.
“There’s just an insatiable demand for optical fiber,” said Jason Eichenholz, the cofounder of Relativity Networks, an optical networking and fiber-optics startup.
AI changes the fiber game
For decades, optical fiber was a quiet corner of telecommunications, Clark said. When ChatGPT launched in 2022, it brought fiber into the spotlight.
“AI is the biggest change factor in data centers and digital infrastructure that we’ve seen since the origins of the industry,” Clark said.
That demand is also reshaping how data centers are built. While single mega-facilities exist — like xAI’s Colossus data center in Memphis, Tennessee — there are limits to how large one facility can be due to the amount of electricity and water needed to support each data center, Eichenholz said. Communities are also pushing back against this type of data center construction in their neighborhoods, he added.
The lack of mega-facilities means companies build multiple data centers with fiber connecting them, Bell said. Fiber gets data from point A to point B and connects GPUs and chips, allowing them to “function essentially as one gigantic computer,” Bell said.
The challenge to scaling up fiber supply lies in production capacity, Bell said. That includes making the fiber, sourcing connectors, testing the fiber, and the physical labor of laying fiber in the ground.
Additionally, non-data-center industries, such as office buildings, industrial sites, and residential homes, use fiber for connectivity — such as fiber-optic internet that many people use at home.
“Because of data centers increasing their demand for fiber, it is putting a strain on the overall system,” Clark said. “There is definitely a dearth of supply that’s going to trickle through the entire market.” She said fiber orders today take twice as long to fill compared to roughly three years ago.
Solutions take shape
Hyperscalers are assessing if they can trust that fiber will be delivered on time and in the quantities they need, Bell said. To solve the crunch, tech giants and fiber manufacturers are signing enormous pre-commitments, Eichenholz said. “You’re seeing massive multi-year, multibillion-dollar take-or-pay contracts on a product that was a commodity three years ago,” he added.
Fiber manufacturers are scaling up capacity, but in a disciplined manner. They’re locking in commitments from customers before they increase manufacturing and labor capacity, so that hyperscalers and suppliers share the costs and risks of expansion, Corning’s O’Day said.
Corning is investing $585 million in expanding manufacturing in North Carolina. This year alone, the optical fiber manufacturer inked long-term agreements to supply fiber and cable to Amazon, Meta, and Nvidia. Last month, the company’s optical communications division reported a 44% year-over-year jump in sales, while GenAI product sales more than doubled.
Travis Dove for The Washington Post via Getty Images
“Our customers are connecting enormous numbers of GPUs, and they need to fit much more fiber and many more connections into the same physical space,” O’Day said.
Beyond agreements, fiber suppliers are innovating the product itself. O’Day said Corning has developed smaller-diameter fiber and denser cables that put more capacity within the same footprint. The company invented a fiber with a 40% smaller surface area that can bend into 10 times smaller spaces. That led to the development of its Contour Flow Cable, which Corning announced in March.
While incumbents like Corning are packing more fiber into limited spaces, new entrants are rethinking optical physics altogether.
Relativity Networks, an AI network startup, developed ChronoCore, a hollow-core fiber designed to bypass the physical speed limits of conventional glass. The fiber allows signals to travel at 99.7% the speed of light, Eichenholz said. He calls it “an HOV lane for fiber.” By comparison, light moves through glass fiber at two-thirds the speed of light. Relativity is partnering with Prysmian, one of the world’s largest producers of fiber-optic cables, to scale production.
The difference in speed matters when connecting data centers. Light traveling through traditional fiber has a slight delay, but the hollow core moves at nearly the speed of light, which means operators can build data centers farther apart without harming the performance of an AI model.
Relatively completed its first deployment with a major hyperscaler in 2024 and recently landed a $40 million order to connect data centers with the same hyperscaler.
Bell predicts the pace of data center buildouts will eventually level off. But they don’t see the appetite for fiber shrinking in the near term.
“There’s a lot of unknowns with respect to AI infrastructure,” Bell said. “But the use of fiber is really kind of one of the highest conviction areas where I think we’re going to clearly need more of.”
