Saturday, 19 September 2026
The artificial-intelligence infrastructure boom has largely been discussed through the enormous investment flowing into chips, data centres, electricity and cooling. A less visible layer is now becoming equally important: the fibre and network infrastructure required to move huge volumes of data between those computing facilities. Speaking at Ai4 2026 in Las Vegas, Jim Fowler, Executive Vice President and Chief Technology and Product Officer at Lumen Technologies, argued that increasingly powerful AI models cannot deliver their full value unless companies can move information rapidly between users, applications, sensors, cloud platforms and computing infrastructure.
His analogy was straightforward. If an AI model is the brain, the communications network increasingly resembles the nervous system connecting that intelligence to the rest of the economy. That proposition has significant implications for telecommunications infrastructure investment. AI workloads are increasingly distributed rather than confined to one data centre. Corporate information may sit across several cloud platforms, private infrastructure, software-as-a-service applications and edge locations, while the computing capacity used to analyse that information could be located hundreds or thousands of kilometres away.
Connecting the two is becoming a physical infrastructure challenge. The availability of computing power alone does not solve the problem. A business can secure access to expensive GPUs and still be unable to use them efficiently if the data required for training or inference cannot reach those processors quickly enough. Fowler described a conversation with a large financial institution that had obtained GPU capacity for AI model training but found the processors were located separately from the required data. According to his account, network capacity became the limiting factor, leaving expensive computing resources underutilised while information was transferred.
The example illustrates a broader change taking place across AI infrastructure. Location decisions for data centres and computing capacity increasingly depend not only on available electricity and land but also on connectivity to other computing clusters and major concentrations of enterprise data. That means fibre routes, carrier-neutral interconnection facilities and cloud access points could become increasingly important components of the AI investment landscape.
Lumen itself is making a substantial bet on that direction. The company reported that it had deployed approximately 17 million intercity fibre miles by the end of 2025 and plans to increase this to around 47 million by the end of 2028 and approximately 58 million by 2031. The expansion forms part of a broader strategy to position its network for increasing AI and cloud traffic.
The company has also secured nearly $13 billion of Private Connectivity Fabric contracts, including agreements involving major AI and cloud companies. Anthropic has selected Lumen to expand high-capacity fibre infrastructure supporting its North American operations. For the property sector, this represents another infrastructure layer behind the data-centre development cycle.
A new hyperscale facility cannot operate in isolation. It needs connections to cloud regions, other data centres, corporate networks and ultimately the customers and machines generating the information being processed. As more computing facilities are developed in locations selected according to electricity supply, planning conditions, water availability or tax considerations, the distance between data and compute can increase. That creates additional demand for high-capacity fibre corridors capable of moving information between geographically dispersed facilities.
The result could strengthen the investment case around land and infrastructure positioned along major digital corridors, particularly where multiple data centres can connect to several independent fibre routes. This also helps explain why network architecture is beginning to change.
Traditional enterprise telecommunications frequently involved relatively fixed connections between offices, company data centres and external networks. The cloud fundamentally changed where applications and data were located, but much of the wider network infrastructure connecting those environments continued to operate through relatively static arrangements. AI is putting more pressure on that model.
Computing workloads can now move between cloud regions and different providers depending on available GPUs, electricity costs, data sovereignty, latency and capacity. Enterprises may need additional bandwidth for several hours or days rather than signing up to fixed capacity that remains unchanged for years. Fowler therefore expects telecommunications networks increasingly to operate more like cloud computing itself.
Instead of engineers manually configuring multiple pieces of networking equipment every time capacity needs to change, businesses would be able to activate additional connectivity through software, increase or decrease bandwidth and establish new connections between cloud platforms or data centres on demand. Lumen calls this transition Network-as-a-Service, and adoption of its programmable connectivity products has been growing. The company reported more than 3,000 NaaS customers by the second quarter of 2026, with increases in customers, active ports and services during the quarter.
The strategic significance is not simply that network services can be purchased online. It is that connectivity could eventually become another programmable component within a wider AI infrastructure stack. An AI workload might require additional compute capacity in one cloud region and automatically establish the necessary network bandwidth to move data there. Once the task is finished, that capacity could be reduced again.
Such a model is considerably different from traditional telecommunications procurement. Lumen accelerated this strategy in July by completing its acquisition of Alkira, a cloud-networking company that provides software for connecting cloud environments, data centres and corporate locations through a unified control layer. The transaction allows Lumen to combine physical fibre infrastructure with software capable of orchestrating connectivity across hybrid and multi-cloud environments.
The acquisition is significant because it demonstrates how the distinction between telecommunications and cloud infrastructure is becoming less clear. Physical fibre remains essential, but customers increasingly expect the network to behave like software. They want to be able to configure connections, security policies and capacity without manually coordinating several telecommunications carriers or cloud providers.
Fowler argued that the eventual network supporting AI will therefore require three elements: very high-capacity physical infrastructure, a programmable control layer and an ecosystem of security and network services that can be activated alongside connectivity. Security could increasingly become part of the same infrastructure model.
Businesses currently combine networks with separate firewalls, load balancers, domain services and other security products. Fowler’s vision is that these capabilities increasingly become software functions that can be deployed alongside connectivity through the same interface. The wider industry is already moving in that direction as enterprises attempt to manage networks spanning multiple public clouds, data centres and geographic locations.
Lumen’s acquisition of Alkira specifically targets this challenge by adding a carrier-neutral control layer capable of creating connections across cloud and hybrid environments. The company argues that bringing physical network infrastructure and cloud orchestration together will allow customers to activate and change connectivity much more rapidly.
AI could accelerate the need for these systems because machine-to-machine communication operates differently from conventional human internet traffic. A person opening a website or sending an email generates relatively modest and predictable network demand. AI applications can require enormous datasets to move between storage systems, training environments and inference infrastructure.
Agentic AI could add another layer. As software agents communicate with databases, applications, sensors and other agents, the volume and complexity of automated network traffic could rise considerably. This means the telecommunications infrastructure built around the internet’s previous growth phase may not necessarily be optimised for the next one.
Data-centre-to-data-centre and cloud-to-cloud connectivity could become particularly important. Fowler said Lumen increasingly sees growth coming from connections between computing facilities rather than the traditional model centred on connecting corporate premises to the cloud. This reflects a wider architectural change as enterprise systems become distributed across several cloud and computing environments.
The company’s recent expansion illustrates that strategy. In August, Lumen extended its Multi-Cloud Gateway service to more than 10 million US business locations, allowing enterprises to reach private cloud connectivity even where the underlying location does not sit directly on Lumen’s own fibre network.
The requirement for redundancy will also increase. A data centre hosting critical AI workloads cannot depend on a single physical fibre route. Large facilities increasingly require multiple carriers and geographically diverse connections so that a cable failure or equipment problem does not isolate expensive computing infrastructure.
For developers and investors, fibre availability may therefore become a more important part of site due diligence. Electricity remains the central constraint for many data-centre markets, but access to large quantities of power is less valuable if a site cannot also move enormous amounts of data reliably.
This could create another divide between development locations that merely have suitable land and electricity and locations that also sit within strong digital infrastructure networks. Secondary data-centre markets could benefit where new fibre investment reduces the connectivity disadvantage historically associated with being located further from established hubs.
The same dynamic could influence the emerging edge-computing market. Not every AI workload can operate from a small number of enormous hyperscale campuses. Applications involving industrial automation, autonomous vehicles, telecommunications networks and other real-time systems may require computing capacity closer to where information is generated.
That would create a more distributed network of smaller computing facilities connected to larger regional and hyperscale centres. Connectivity therefore becomes increasingly important at every level of the AI infrastructure hierarchy: between continents, between cities, between data centres and between edge facilities and the devices generating information.
Fowler’s broader argument is that telecommunications infrastructure itself is being transformed from a relatively static utility into a flexible computing resource. That transition resembles what happened to servers during the first wave of cloud computing. Companies once bought physical machines, configured them manually and installed software individually. Cloud computing converted much of that infrastructure into something that could be requested almost instantly through software.
Telecommunications could now experience a similar shift. Networks that once took weeks or months to provision could increasingly become programmable resources capable of responding to workload demand.
For property investors, developers and infrastructure funds, the consequence is that AI’s physical footprint may be much broader than the data-centre campuses attracting most attention today. The investment chain begins with electricity generation and grid connections, continues through land and data centres, and increasingly extends into long-haul fibre, metropolitan networks, interconnection facilities and cloud access infrastructure.
Each layer depends on the others. The fastest GPU cluster provides little economic advantage if insufficient electricity is available to operate it. A data centre with abundant power becomes less useful if information cannot reach it quickly. And an advanced AI model remains constrained if the systems generating its data cannot connect reliably to the computing infrastructure processing it.
The AI infrastructure race is therefore becoming a connectivity race as well as a computing race. As Fowler put it at Ai4, intelligence alone is not enough. AI needs a system capable of connecting people, machines, applications and data at very high capacity and low latency.
If the current investment cycle continues, the next generation of telecommunications infrastructure could become one of the least visible but most economically important components of the AI boom. The industry’s next question may therefore not simply be who owns the most advanced AI model or the largest data centre. It may increasingly be who owns, controls and can rapidly expand the infrastructure connecting all of them.
Source: CIJ.World Research & Analysis Team
