Telecommunications infrastructure is entering a different investment cycle. Carrier capital spending and data center networking demand are now moving on different paths. Large computing environments generate enormous volumes of machine-to-machine communication inside facilities. Growing clusters also need fast connections between campuses, cloud regions and storage locations. Traditional telecom spending does not capture all of this infrastructure. Cloud providers and data center operators now buy significant amounts of networking equipment directly. Dell’Oro Group reported flat nominal telecom investment among covered service providers during 2025. It expects worldwide telecom capital expenditure to decline in 2026. Yet optical transport equipment revenue grew 10% in 2025. Direct WDM purchases for data center interconnection also increased by nearly 40%.

These figures reveal an important split in infrastructure spending. Weak aggregate telecom investment can coexist with strong demand in selected networking markets. Optical systems, switches and interconnection equipment are benefiting from cloud and accelerated computing expansion. This does not mean every networking segment will experience the same growth. Instead, spending is becoming more closely connected with the location and architecture of computing resources. Accelerator clusters need networks capable of moving data quickly and consistently. Those requirements exist inside facilities and between separate computing sites. As a result, connectivity decisions increasingly influence how effectively expensive computing capacity can operate. The network is becoming part of the compute architecture itself.

The Network Is Moving Closer to the Compute

Large accelerator clusters tie networking infrastructure more closely to data center locations. They concentrate switching, optical connectivity and transport requirements around computing sites. Conventional telecommunications architectures often start with users and move traffic through several network layers. Access networks feed aggregation systems, while carrier backbones transport data toward applications. Distributed AI creates another major traffic environment inside the computing infrastructure. Training systems require accelerators to exchange information through high-performance fabrics. Networking therefore becomes an active component of the computing system. It is no longer simply the route that connects an application with its users. Poor communication performance can also limit the value delivered by expensive processors.

The scale of projected investment shows how important this layer has become. Dell’Oro Group updated its outlook in February 2026. It expects spending on AI back-end data center switches to surpass $100 billion by 2030. That forecast covers scale-up, scale-out and scale-across architectures. These switches sit directly around the servers generating demanding workloads. Their performance affects how effectively accelerators communicate during distributed computing operations. Network designs must handle bandwidth, congestion and increasingly large accelerator domains. They also need predictable behavior when thousands of processors exchange information at once. NVIDIA’s Spectrum-X architecture illustrates this changing relationship between networking and computing. The platform addresses communication within AI environments and across distributed infrastructure.

Inside the Facility, Traffic Behaves Differently

Much of the traffic associated with accelerated computing can remain inside a facility. That helps explain an important feature of current network demand. More AI infrastructure does not produce an equal increase across every public telecommunications network. Distributed training creates large amounts of east-west traffic between computing systems. Accelerators exchange data with other accelerators, storage platforms and supporting infrastructure. These flows differ from north-south traffic associated with users accessing applications. The distinction becomes increasingly important as accelerator clusters grow. Internal networks must move huge volumes of data without relying on public carrier infrastructure. That creates networking demand even when the traffic never leaves the data center.

Meta has demonstrated the scale that these internal networks can reach. The company documented two clusters containing 24,576 GPUs each for generative AI workloads. One implementation used RoCE, while another used InfiniBand. Both supported 400 Gbps endpoints. Those deployments show why network design becomes a major architectural choice at large cluster scale. Facilities need switches, adapters, optical components and structured cabling to support those systems. However, internal communication is only part of the network requirement. Datasets, checkpoints and workloads may still move outside one computing domain. Architects must therefore distinguish traffic within clusters from traffic crossing buildings or regions. That distinction influences where external transport capacity becomes necessary.

Data Center Interconnection Becomes Strategic Infrastructure

Networking becomes a geographic infrastructure problem when computing extends across separate facilities. A single site may not contain every resource required by a workload. Organizations can connect compute campuses with storage, cloud regions and recovery environments. They may also need access to additional accelerator capacity at other locations. Power availability can influence where new computing infrastructure gets constructed. This can create demand for routes connecting those sites with established digital infrastructure. Data sources and user markets may remain far from the available power. The network must bridge those locations without compromising workload requirements. Long-haul connectivity therefore becomes relevant to the design of distributed computing systems.

TeleGeography expects used international bandwidth to grow at a 24% compound annual rate from 2025 through 2035. The company also provides an important qualification to that forecast. AI-specific demand cannot be separated cleanly from all other traffic in those totals. Cloud applications, video and social platforms continue to generate substantial network demand. Connected devices and broadband adoption contribute as well. Network planners must support this established growth while preparing for new compute locations. Fiber availability can influence development decisions when sites require large interconnection capacity. Route diversity also matters when workloads depend on links between separate facilities. Optical capacity determines how much information those paths can carry. These factors become particularly relevant for distributed computing projects.

Optical Spending Shows Where Demand Is Appearing

Optical transport provides measurable evidence of changing network investment. Dell’Oro reported that the market reached $16 billion in 2025. Revenue increased 10% during the year. The research firm identified data center interconnection as the primary driver of that recovery. Direct purchases by cloud providers increased by about 50%. Those customers bought more WDM transponders, ZR optics and optical line systems. Direct WDM purchases for data center interconnection increased by nearly 40%. These numbers provide stronger evidence than assumptions based only on projected AI traffic. They show actual spending around the infrastructure connecting computing environments. The growth is also occurring while broader telecom capital expenditure remains restrained.

Dell’Oro expects optical transport revenue to continue growing during 2026. It also expects data center interconnection revenue to outperform the overall optical market. Meanwhile, Nokia has reported strong demand from AI and cloud customers. Its second-quarter 2026 sales to that customer group increased 105% year over year. Optical Networks grew 20% on a constant-currency basis. IP Networks grew 16% using the same measure. Nokia also reported €2.8 billion in AI and cloud order intake during the quarter. The company said demand remained strong while supply represented an important industry constraint. These figures do not suggest equal growth across every telecommunications category. They show that specific networking markets are already benefiting from accelerated computing investment.

Private Backbones Complicate the Telecom Opportunity

Higher data volumes do not automatically produce equivalent carrier revenue. Major cloud and content providers operate extensive private network infrastructure. They can also obtain capacity through several different commercial models. TeleGeography reported that content providers represented 75% of used international bandwidth in 2025. Its definition includes hyperscalers, cloud providers, AI platforms and neocloud operators. The research firm expects international content-provider bandwidth to increase ninefold from 2025 through 2035. That forecast covers several forms of digital traffic rather than AI alone. Large operators can connect facilities through private backbones and dedicated capacity. Direct interconnection can keep additional traffic away from the public internet. Traffic growth alone therefore cannot identify which provider will capture the associated network spending.

This creates a more complicated commercial environment for telecommunications companies. Internet transit providers can face different demand from suppliers of dedicated infrastructure. Dark fiber, wavelengths and optical systems serve different parts of the connectivity stack. Colocation interconnection also addresses requirements that generic internet transit cannot always satisfy. Private networks still depend on extensive physical infrastructure. They require fiber, cable systems and rights-of-way across their routes. Subsea networks also depend on landing infrastructure and specialized operating capabilities. Those requirements leave commercial opportunities across several layers of the network. The opportunity depends partly on who owns or supplies infrastructure around high-capacity compute locations. Network demand is growing, but its economic value can land in different places.

Fiber Procurement Is Moving Up the Infrastructure Agenda

Fiber procurement offers another indication of connectivity’s growing role in infrastructure development. Meta announced a multiyear agreement with Corning in January 2026. The agreement has a potential value of up to $6 billion. It supports fiber-optic infrastructure for Meta’s U.S. data centers. The arrangement includes optical fiber, cable and connectivity products. Meta also said Corning would expand manufacturing capacity in North Carolina. The commitment connects optical supply directly with a major data center expansion program. It does not mean every operator needs the same procurement model. Still, it shows how large connectivity requirements can justify early planning for fiber supply.

A powered campus can still encounter connectivity limitations. Insufficient fiber diversity may restrict links to other computing locations. Delayed optical equipment can create another constraint. Therefore, developers can evaluate network infrastructure alongside electrical and mechanical requirements. Available routes provide one part of that assessment. Carrier presence and conduit capacity provide additional information about connectivity options. Physical path diversity deserves separate attention because nominal carrier choice may not guarantee independent routes. Delivery schedules also matter when optical equipment has to match facility commissioning. The important question extends beyond whether a facility can reach the internet. Distributed workloads may require substantial dedicated capacity between strategically important computing sites. Early planning can expose those constraints before they affect operating capacity.

Scale-Across Architecture Changes the Meaning of a Cluster

Distributed computing is beginning to challenge the assumption that one cluster must occupy one facility. NVIDIA describes Spectrum-XGS Ethernet as an architecture for connecting separate data centers. Those sites can occupy different buildings or sit much farther apart. The design addresses communication between geographically distributed computing resources. NVIDIA says the architecture combines congestion control, latency management and telemetry. Its reported performance improvements remain vendor claims based on its platform and testing conditions. The broader engineering problem exists regardless of the supplier. Geographic separation introduces latency, routing and recovery requirements. It also creates demand for predictable inter-site bandwidth. Those conditions differ from communication inside a single server hall.

Cloud providers already operate networks that connect multiple regions and facilities. Distributed AI adds another reason to examine those links carefully. Power constraints can make it difficult to place every accelerator at one site. Physical capacity can create another limit on expansion. Network designers must then determine which workloads tolerate geographic separation. Some communications remain highly sensitive to latency. Other operations can run across longer distances without the same performance impact. Route engineering becomes more important when a workload depends on multiple sites. Failure recovery also needs to account for inter-site network conditions. Scale-across architectures already allow computing environments to span several facilities. Adoption will depend on workloads, latency requirements and infrastructure constraints.

Inference Creates a Different Network Problem

Training attracts attention because it concentrates large numbers of accelerators. User-facing inference creates a different set of network requirements. Application design has a major influence on those requirements. A text interaction can create less traffic than a bandwidth-intensive media workload. Accelerator counts alone therefore cannot predict public-network demand. Voice applications can produce different traffic patterns from text systems. Generated video introduces another set of bandwidth requirements. Machine vision and persistent agents may behave differently again. Payload size, interaction frequency and latency sensitivity all influence network demand. The location of inference changes the transport pattern as well.

Processing close to users creates different traffic paths from centralized inference. Hyperscaler applications may also keep substantial traffic on private infrastructure. Public-network demand therefore depends on more than overall AI adoption. TeleGeography reported international internet bandwidth of 1,835 Tbps in 2025. That represented a 23% increase during the year. The firm attributed continued growth to several factors rather than AI alone. Connected devices, broadband penetration and bandwidth-intensive applications remain important contributors. AI-specific bandwidth demand cannot yet be separated cleanly from this broader expansion. Usage patterns for many emerging applications are also continuing to develop. Capacity planning works better when it reflects these individual workload characteristics.

Network Planning Has to Start With Workload Architecture

Infrastructure buyers should model telecommunications capacity around actual workload behavior. A generic bandwidth target cannot describe every requirement of a distributed computing environment. Planners first need to know where training will occur. They also need to identify where inference, datasets and storage will reside. Checkpoint movement can create another network requirement. Recovery infrastructure may sit at a separate location. Each decision affects the amount and type of connectivity required. Some projects need dense internal switching. Others require metro interconnection, long-haul wavelengths or private fiber. Large environments may need several of these layers at once.

Physical route diversity deserves particular attention. Services purchased from different carriers can still share portions of the same infrastructure. Separate commercial contracts therefore do not always represent completely independent physical paths. Latency targets also need precision because workloads tolerate distance differently. A high-capacity connection alone does not guarantee predictable application performance. Congestion and failure conditions can alter the behavior of a network path. Optical capacity must also align with fiber characteristics and transceiver choices. Equipment availability can affect when those connections become operational. These dependencies make networking part of compute-utilization planning. Evaluating power, compute and telecommunications together can expose dependencies that separate assessments may miss.

The Next Network Cycle Will Be Distributed Unevenly

The current investment pattern does not point toward a universal telecommunications boom. Different parts of the network stack face different demand conditions. Dell’Oro expects worldwide telecom capital expenditure to decline in 2026. At the same time, optical transport and data center networking have shown strong growth. Direct purchases from cloud providers add another dimension to this market. The divergence matters to equipment suppliers and carriers. It also matters to data center developers and infrastructure investors. Fiber operators serving major compute markets can encounter different demand from mature access-network businesses. The location of network assets is becoming important to the commercial opportunity. Connectivity demand follows the infrastructure that needs to communicate.

Cloud providers already purchase networking equipment directly. Dell’Oro has also identified neoclouds as an emerging group making more optical-network purchasing decisions. Finally, evidence already shows substantial investment in networking systems surrounding large compute deployments. The harder issue is determining where the resulting data will move. Some traffic will remain inside private infrastructure. Other workloads will require third-party transport between computing sites. Different providers may therefore capture value from different network layers. Optical transport, private fiber and interconnection can benefit without matching growth across traditional carrier spending. Current market data links significant networking growth with compute locations and data center interconnection. Connectivity architecture has consequently become a more important part of digital infrastructure planning.