0

AI traffic is triggering a new fiber focus meant to support distributed training, inference and critical workloads

The first wave of AI infrastructure investment was overwhelmingly a compute story. Hyperscalers and neocloud providers raced to assemble GPUs into increasingly large AI factories, while networking investment concentrated on the scale-up and scale-out fabrics needed to keep those accelerators working efficiently. Now the effects are spreading beyond the data center.

As AI infrastructure becomes larger and more distributed — and as the industry shifts from training toward widespread inference — the networks connecting data centers, clouds, enterprises and eventually edge locations are becoming part of the AI infrastructure equation. That transition is beginning to produce a new investment cycle in fiber and optical networking.

The market data is increasingly difficult to dismiss. Dell’Oro Group reported that the global optical transport equipment market grew 15% year-over-year in the second quarter of 2026, while data center interconnect revenue from IP-over-DWDM and traditional WDM systems increased 45%. Dell’Oro expects the overall optical transport market to grow 16% this year and exceed $18 billion in manufacturer revenue, citing AI data center construction as a central demand driver.

Service providers see the same trajectory from the demand side. New research commissioned by Ciena found that 90% of more than 1,200 telecom, wholesale fiber and regional service provider respondents expect high-capacity AI network services to drive revenue growth during the next three to five years. More tellingly, 88% expressed a strong sense of urgency around optical network upgrades needed to support premium enterprise AI connectivity, while just 11% believe routine upgrades will be sufficient.

Some of the immediate pressure comes from a familiar networking problem at an unfamiliar scale. Power, real estate and cooling constraints increasingly make it difficult to expand the largest AI clusters indefinitely within a single facility or campus. The result is greater interest in scale-across architectures that connect separate compute domains and allow them to function more like a larger distributed system.

Side note: We recently published a report examining multi-dimensional scaling in the shift from training to inference; download “Building AI-era network fabrics”.  

The networking requirements can be extreme. Ciena’s research notes that distributed synchronous training of frontier models can require communications measured in tens of petabits per second, potentially involving hundreds of fiber pairs operating in parallel. Some 95% of surveyed providers expect hyperscaler scale-across requirements to contribute substantially to wholesale revenue growth.

But training is only part of the story. Nearly half of the surveyed service providers expect the growth of AI inference data centers to create new DCI revenue opportunities. As inference becomes more geographically distributed, network requirements expand beyond raw capacity to include latency, reliability, jitter and the ability to connect workloads dynamically across multiple infrastructure domains.

Operators are already responding. AT&T, for example, has expanded 400G wavelength services to more than 40 U.S. metros and enabled 400G handoff capability across 440,000 properties serving more than 2.3 million business tenants. The operator is also increasing capacities toward 1.6 Tbps across key metro and long-haul routes as it positions its network for AI and other data-intensive applications.

The broader implication is that AI is moving the networking bottleneck outward. Accelerators remain scarce and expensive, but compute alone has diminishing value if models, datasets and inference requests cannot move efficiently among locations. Fiber availability, optical capacity, route diversity and predictable network performance therefore become increasingly important components of AI infrastructure.

That creates an opportunity for network operators, but also an investment requirement. The next phase of the AI buildout will not happen entirely inside AI factories. It will increasingly happen between them, and across the networks connecting compute to clouds, enterprises, devices and users. For optical networking, the AI infrastructure cycle is only beginning.

Billion-dollar fiber agreements and purpose-built private networks show how hyperscaler AI infrastructure is changing the economics of carrier networks

The cloud era was supposed to diminish the strategic importance of telecom networks. Hyperscalers built enormous private backbones, pushed intelligence into centralized data centers and increasingly controlled the infrastructure connecting their own facilities. AI is complicating that trajectory.

The sheer scale and geographic distribution of AI infrastructure is creating demand for connectivity that cannot always be satisfied by adding capacity to existing networks. Increasingly, hyperscalers and AI companies are contracting with network operators for dedicated fiber infrastructure connecting data centers across metros and long-haul routes.

The result is something closer to a private AI backbone. Verizon, for instance, provided one of the clearest examples in July when CEO Dan Schulman disclosed a dark fiber agreement with Google worth more than $1 billion to connect data centers. Verizon expects additional AI infrastructure agreements anticipated by year-end to collectively represent multiple billions of dollars, with the resulting revenue beginning to contribute meaningfully to growth in 2027.

That follows an earlier agreement with Amazon Web Services under Verizon AI Connect. Rather than simply selling AWS more capacity on the existing network, Verizon is constructing new long-haul, high-capacity fiber pathways between AWS data center locations, with an emphasis on route diversity, low latency and resilience.

This is a relevant distinction because AI infrastructure is pushing some network relationships away from conventional connectivity services toward purpose-built physical infrastructure backed by long-duration customer commitments.

Lumen provides perhaps the clearest example of how far that model can go. The company has secured nearly $13 billion in Private Connectivity Fabric contracts with hyperscalers, neocloud providers, social platforms and AI companies. Its customers include Anthropic, which selected Lumen to expand a high-capacity, purpose-built network across North America. At the same time, Lumen plans to expand its intercity network from 17 million fiber miles at the end of 2025 to 47 million by 2028 and approximately 58 million by 2031.

That represents an important change in the investment equation. Building long-haul fiber is expensive and slow. Operators have to secure rights-of-way, install conduit and fiber, build regeneration and optical infrastructure, and create redundant routes. Historically, that meant investing capital based partly on forecasts of future traffic.

Large AI infrastructure agreements can shift some of that risk. A hyperscaler or AI provider effectively anchors the investment before construction is completed, giving the operator greater certainty that new fiber will produce revenue.

Service providers expect that model to become increasingly important. Ciena’s latest global survey found that 96% of respondents expect managed optical fiber network services to contribute revenue from connecting distributed AI compute clusters over the next three years. More than half identified managed optical fiber networks as their primary vehicle for delivering data center interconnect services.

The appeal is straightforward: a managed optical fiber network can give a hyperscaler infrastructure that behaves much like a private optical backbone while leaving much of the construction, operations and lifecycle management with the carrier. This creates an interesting reversal for telecom operators.

For years, hyperscaler private networks raised concerns that cloud companies would progressively disintermediate carriers. AI is demonstrating the continuing value of assets that cannot be reproduced in software, including rights-of-way, conduit, fiber routes, metro density and operational expertise.

The competitive opportunity, however, extends beyond owning fiber. Lumen is combining its physical network with programmable networking and cloud connectivity, while Verizon AI Connect packages dark fiber and wavelengths alongside data center, cloud, compute and security services.

That points toward a broader shift in the carrier-hyperscaler relationship. AI companies need enormous quantities of compute, but distributed compute has to be connected with infrastructure capable of moving models, datasets and inference traffic predictably between locations. For network operators, the opportunity is building the private infrastructure on which the AI economy increasingly depends.

Distributed inference, agentic AI and physical AI are changing the shape of network traffic

Training frontier AI models is fundamentally a problem of concentration. Thousands of accelerators are assembled into large clusters, and enormous amounts of data move among GPUs, storage systems and increasingly between adjacent data centers. Inference creates a different networking problem.

As the AI story expands from building models toward using them, requests originate wherever people, enterprises, devices and machines happen to be. Some inference workloads can still be served efficiently from centralized clouds. Others, particularly things like real-time voice and video, autonomous systems, industrial applications and agentic workflows, introduce requirements around latency, data locality, reliability and uplink capacity that can push compute closer to where data is generated.

For network operators, that makes the AI traffic opportunity considerably broader than connecting hyperscale data centers.

AT&T CEO John Stankey has framed the change explicitly. In the company’s second-quarter 2026 investor commentary. “The rise of agentic AI is fundamentally reshaping network traffic, not just in volume, but in shape, symmetry and criticality. The proliferation of agentic and autonomous AI workloads will require networks to sense, decide and act in near real time.” This  shift has an important physical consequence; the edge of the AI network needs more fiber.

The opportunity is already showing up in service provider expectations. Nearly half of respondents to Ciena’s latest global survey believe growth in AI inference data centers will create new data center interconnect revenue. But the opportunity extends farther outward: 88% expect edge-hosted services, including localized inference, edge compute and GPU-as-a-service, to account for at least 10% of enterprise revenue within three years. Meanwhile, 60% expect physical AI ecosystems such as industrial robotics and remote-control systems to represent more than 15% of enterprise AI revenue within five years.

Those applications connect the optical core to a very different set of endpoints. AT&T and AWS provide an early example. Their AWS Interconnect last mile offering is designed to extend AT&T fiber and 5G connectivity directly from enterprise locations into AWS environments, creating more predictable premises-to-cloud connectivity for AI applications. AT&T is simultaneously expanding capacities up to 1.6 Tbps across key metro, regional and long-haul routes.

The architecture illustrates the larger point. AI connectivity is more than scaling up, out and across. Enterprises increasingly need a high-performance path from where data is created to where inference takes place.That makes metro density and access infrastructure more strategically important. It also adds an AI dimension to the massive fiber investments already underway at the largest U.S. operators.

AT&T completed its $5.75 billion acquisition of Lumen’s mass-market fiber business in February, adding more than 4 million fiber locations and significant construction capabilities across 11 states. AT&T now expects its fiber footprint to exceed 60 million locations by the end of 2030. Verizon’s acquisition of Frontier similarly expanded its reach to approximately 30 million fiber passings, with Verizon targeting 40 million to 50 million over time.

Neither transaction should be recast as an AI acquisition. The immediate economics are driven primarily by broadband growth and fixed-mobile convergence. But distributed inference makes those assets more valuable.

Dense access and metro fiber can connect enterprises to clouds, backhaul growing uplink traffic from wireless networks and provide connectivity to edge compute locations. Verizon explicitly identified expansion of its intelligent edge network for applications including AI and IoT as one benefit when it announced the Frontier transaction.

The long-term implication is that AI could change where operators perceive network value. Training concentrated value inside massive compute facilities and the optical links connecting them. Inference distributes that value outward toward metro networks, enterprise locations, wireless aggregation points and ultimately the edge. For operators with dense fiber footprints, that puts the network edge in play.

Operators are pushing toward 400G, 800G and 1.6T optical networking, but AI workloads are also increasing the importance of deterministic performance, power efficiency, telemetry and automation

The first requirement AI places on optical networks is obviously more capacity. Connecting increasingly large data centers, distributing compute across multiple sites and moving enormous datasets among clouds and enterprises are driving traffic growth that conventional network expansion cycles were not designed to accommodate. Operators are responding with new fiber, higher-capacity wavelengths and more efficient coherent optics.

But simply adding bandwidth will not be enough. AI workloads are increasingly important to business operations, and their performance can depend on more than whether sufficient aggregate capacity exists. Latency, jitter, availability and consistency all become important when networks connect distributed compute resources or support real-time inference.

That shift is reflected in service provider priorities. In Ciena’s latest global survey, 35% of respondents identified network consistency, including guaranteed performance stability and low jitter, as the leading opportunity for monetizing premium connectivity. And 96% said advanced network automation, including agentic AI, will be important to capturing AI-driven network opportunities.

The emerging optical investment cycle is about simultaneously increasing capacity, efficiency and intelligence.The capacity transition is already visible in commercial networks. AT&T, for instance, has expanded 400G wavelength connectivity to more than 40 U.S. metros, with 400G handoff capability enabled across 440,000 properties serving more than 2.3 million business tenants. The operator explicitly positions that expansion around distributed AI, cloud and analytics workloads that require high throughput and predictable performance between locations where data is created, processed and stored.

The technology roadmap is moving quickly beyond 400G. Verizon has validated 1.6 Tbps transmission over a live metro network using Ciena’s WaveLogic 6 Extreme coherent optics. The trial demonstrated a single 1.6 Tbps wavelength operating through a dense ROADM environment, while the technology is also designed to extend 800G connectivity across substantially longer network distances.

Lumen is deploying the same generation of coherent technology as it expands infrastructure for AI workloads. Beyond increasing raw capacity, Ciena says WaveLogic 6 Extreme can double capacity within the same space and power envelope, reducing space and power consumption per bit by approximately 50%. Lumen is also using Ciena’s network-control software to automate management of its expanding fiber assets.

Those efficiency improvements are important because AI is putting pressure on more than bandwidth. Power and physical space are constrained throughout digital infrastructure, meaning the economics of optical upgrades increasingly depend on how much capacity operators can extract from existing fiber, racks and power envelopes.

The next challenge is operational. Traditional optical networks have largely been engineered around relatively predictable traffic requirements and long-lived circuits. AI introduces a more dynamic environment. Workloads can move between clouds and data centers, compute availability can change, and application requirements can vary considerably in latency, capacity and resilience. That puts greater value on telemetry and automation.

Service providers in Ciena’s survey increasingly favor dynamic connectivity models: 56% identified multi-cloud connectivity management as an important revenue driver and 50% pointed to consumption-based bandwidth services, compared with 38% for traditional fixed-capacity services. Nearly half expect movement of large AI datasets between clouds to increase demand for flexible, high-capacity connectivity.

Optical vendors are consequently optimizing not just transmission rates but the surrounding system. Nokia, for example, is developing application-specific coherent solutions spanning metro through subsea networks, alongside denser amplification systems intended to light large numbers of fiber pairs with lower power and footprint requirements.

The direction of travel is clear. AI is accelerating the industry’s move from 400G toward 800G and 1.6T connectivity. But the more consequential transition may be from optical networks designed primarily to transport capacity toward networks able to expose, optimize and automate that capacity according to changing application requirements.

For operators, the AI-era optical network will need to be faster, more predictable, more efficient and considerably more programmable.