Zayo, Nvidia Target Fiber Capacity as AI Infrastructure Spreads

AI data center construction is creating new demand for dedicated connectivity, while optical equipment constraints could limit expansion of existing routes.

Zayo is building more than 8,000 miles of long-haul fiber across emerging AI corridors with anchor customer Nvidia as demand for dedicated connectivity grows.

The Denver-based digital infrastructure provider says it will build six new long-haul routes and expand capacity across 10 high-demand markets. The broader program now spans more than 15,000 route miles across North America, according to Zayo.

The buildout adds networking to the infrastructure required to support AI, alongside power, data center capacity and computing hardware.

Analysts differ on how quickly network constraints are developing.

Jimmy Yu, vice president at Dell’Oro Group, said he has not heard of wide-area network bottlenecks caused by AI traffic. The industry remains in the early stages of the AI infrastructure cycle, he said.

Ron Westfall, vice president and practice lead for networking and infrastructure at HyperFrame Research, said existing routes are already running into capacity limits as AI factories move into new markets to access available power.

Existing fiber routes were designed around legacy data center topologies, Westfall said. As AI factories spread geographically, wide-area optical capacity is becoming a rate-limiting dependency for distributed computing clusters.

The longer-term requirement is massive data center interconnect capacity between AI facilities.

Yu said purpose-built routes with low latency and thousands of fiber pairs could become critical to connecting AI data centers, other facilities and exchange points.

“Providers are building ahead of anticipated demand,” Yu said.

Network Becomes Part of AI Infrastructure #

Zayo said its new routes will connect emerging AI corridors where existing long-haul capacity is limited or does not yet exist. It also plans to expand capacity on existing routes where it expects demand to increase.

“AI is fundamentally reshaping where and how network infrastructure needs to be built across the U.S.,” Zayo CEO Steve Smith said in a statement.

The company said it has spent the past 18 months expanding its AI-focused network footprint, with construction and overbuild projects now spanning more than 15,000 route miles.

Nvidia will have significant access to capacity on new routes, while Zayo will build, own and operate the network and make remaining capacity available to other customers.

That gives Zayo an anchor customer while opening the infrastructure to AI developers, cloud providers, neoclouds and enterprises.

AI training can span multiple facilities, while inference places computing capacity closer to users and enterprise applications. For large AI systems, this creates a data center interconnect requirement beyond individual campuses.

Yu said large AI factories could require dedicated, low-latency routes and thousands of fiber pairs to connect facilities and exchange points.

Westfall said the biggest gaps are emerging along secondary corridors where high-capacity long-haul fiber does not exist.

At the wide-area level, distributed AI clusters require high-capacity connections to move data between facilities. At the metro level, regional data centers may lack the dense fiber and low-latency optical connections needed to connect AI infrastructure to corporate networks and edge environments, Westfall said.

A new AI campus can bring hundreds of megawatts or more of computing capacity to a market that lacks the network infrastructure to connect it to other facilities.

Fiber operators have an incentive to identify AI corridors and build before demand arrives.

Optics Face Near-Term Constraint #

Optical equipment is putting pressure on existing routes.

Yu said pump lasers used in amplifiers and coherent transponders are constrained in the near term. Ciena and Nokia have product lead times extending beyond 12 months, with growing backlogs, he said.

Existing fiber routes can carry more traffic when operators add optical equipment. Equipment availability can limit how quickly that capacity becomes available.

New AI campuses create a separate requirement for physical routes that equipment upgrades cannot address. “I think both are constrained,” Yu said of fiber and optical equipment. “It’s just a question of near term and long term.”

Westfall said AI workloads also produce traffic patterns that differ from traditional cloud networks.

Distributed GPU clusters can require significant bandwidth to coordinate workloads across regions, he said. Training runs can generate bursts of traffic as systems synchronize data between geographically separated clusters.

Those requirements can make dedicated, high-capacity routes important as developers distribute AI computing across multiple locations.

Building new routes takes time.

Yu said new fiber routes can take more than a year to build and approach two years – roughly the same order of magnitude as new data center construction.

That creates a potential timing problem for AI developers.

A data center can be complete, powered and equipped with GPUs while lacking the connectivity needed to operate as part of a larger distributed AI system.

“Providers need to begin building these new fiber plants now otherwise data centers will sit idle,” Yu said.

Zayo Builds Ahead of Demand #

Zayo’s strategy is to build ahead of projected demand.

Smith said the company has modeled where AI-driven demand will emerge and is expanding infrastructure ahead of that demand.

For fiber providers, a new route can serve multiple customers and an entire emerging market. Zayo’s expansion follows its acquisition of Crown Castle’s Fiber Solutions business, which added 90,000 metro route miles and 40,000 on-net enterprise locations, according to the company. Zayo has said the additional metro density will support AI inference workloads.

The company has also introduced an AI Infrastructure Blueprint focused on connecting training, inference and interconnection environments.

Zayo said the expanded network is intended to serve more than hyperscalers, including neocloud providers, frontier model developers and enterprises across health care, finance, manufacturing and other industries.

GPU clouds, model developers and enterprises are deploying computing capacity across a growing number of locations, including markets selected for access to power.

That geography creates a network planning requirement alongside the AI buildout.

For now, AI traffic has not overwhelmed wide-area networks, according to Yu. Optical equipment is putting pressure on efforts to add capacity to existing routes, while new AI facilities are creating demand for physical fiber in markets that lack it.