AI: Fiber Shortage, Corning Invests $585 Million

The demand related to AI puts fiber optics under pressure: doubled lead times, megacontracts, and increased capacity are piling up. Corning invests $585 million in North Carolina and multiplies agreements, while Relativity Networks deploys a hollow-core fiber announced to operate at 99.7% of the speed of light.

Immediate Pressure: Doubled Lead Times, Megacontracts, and Rising Sales

According to Lauren Clark, fiber orders now take twice as long to fulfill compared to about three years ago, a sign of supply strain that she says will impact the entire market. In this context, Corning Optical Communications announced last month a 44% year-over-year increase in sales, with sales of GenAI products more than doubling. To secure availability, Jason Eichenholz observes massive pre-commitments and multi-billion dollar contracts spread over several years. On the demand side, Relativity Networks recently completed an initial deployment with a major hyperscaler in 2024 and received a $40 million order to connect data centers with this same client. Meanwhile, Gordon Bell notes that hyperscalers are assessing the reliability of fiber delivery in a timely manner and in sufficient quantities.

Multi-Year Contracts and Shared Capacity Increases Among Manufacturers

According to Mike O’Day, fiber optic producers are increasing capacity in a measured way: they are making commitments with their customers before expanding their factories and hiring, which allows them to share costs and risks with hyperscalers. In 2024, Corning secured long-term contracts to supply fiber and cables to Amazon, Meta, and Nvidia, and is investing $585 million to expand its manufacturing site in North Carolina. At the same time, hyperscalers like Meta are securing their supply through multi-year contracts, while manufacturers expand their capacities and startups develop solutions to accelerate transmissions. Beyond agreements, industry players are also innovating on the product front.

Denser Cables and Hollow-Core: Two Paths to Increase Throughput

Corning claims to have developed smaller diameter fibers and denser cables to increase capacity at a constant volume. The company has invented a fiber with a surface area 40% smaller that can bend in spaces ten times smaller, leading to the Contour Flow Cable announced in March. Meanwhile, Relativity Networks has developed ChronoCore, a hollow-core fiber presented by Jason Eichenholz as enabling signals to travel at 99.7% of the speed of light, much faster than through glass fiber where light travels at about two-thirds of its speed. The company has partnered with Prysmian to amplify production. This speed difference is considered crucial for connecting data centers and could allow them to be spaced further apart without compromising the performance of AI models. On the demand side, O’Day observes that clients are connecting enormous numbers of GPUs and seeking to integrate much more fiber and connections in the same space.

Data Center Architecture Under Constraints of Energy, Water, and Acceptability

While mega-centers exist, such as xAI's Colossus in Memphis, Jason Eichenholz highlights the physical limits related to electricity and water needs, as well as opposition from some communities to these projects. Unable to grow indefinitely, companies are therefore multiplying data centers connected by fiber. Gordon Bell reminds us that fiber connects GPUs and chips to operate as a gigantic computer. However, he identifies the main bottleneck on the production capacity side, which encompasses manufacturing, connectors, testing, and installation work. The pressure is increasing as offices, industrial sites, and housing also consume fiber for connectivity.

Why Fiber is Becoming Critical for Daily AI Operations

AI relies on hundreds of thousands of servers and chips, and the greater the fiber density, the faster the exchanges between machines and the responses of language models, according to Lauren Clark. Mike O’Day links the increasing complexity of models to the need for denser optical networks. Data centers, already relying on fiber to transport signals, need to consume more to keep pace with the growing demands. Otherwise, chatbots freeze or return errors, a phenomenon already observed, and Gordon Bell anticipates, in the longer term, more expensive tokens and extended adoption cycles if supply remains constrained. The rise of AI has also placed fiber back at the center of the game since 2022, the year ChatGPT was launched, and Clark describes AI as the greatest transformative factor for data centers and digital infrastructure since the industry's inception. This dynamic is redefining how centers are built, against a backdrop of a shared bet that fiber will be essential, in a context of demand described as "insatiable" by Jason Eichenholz.

What Could Follow: Stabilization of Construction, Sustained Fiber Demand

Gordon Bell anticipates a stabilization in the pace of data center construction. However, he does not see the appetite for fiber weakening in the short term. While highlighting the many uncertainties still weighing on AI infrastructure, he believes that fiber is among the few certainties in volume: more will be needed.