0

As the center of AI gravity shifts from training to inference, AT&T is positioning its fiber network as connecting AI infrastructure to enterprises and physical systems

The first phase of the AI infrastructure boom has been concentrated around compute infrastructure, particularly rack-scale GPU systems for hyperscale data centers and massive scale-across fiber links connecting those facilities. AT&T sees an opportunity in that buildout, but recent comments from company executives suggest its larger AI thesis sits beyond the data center.

AT&T is positioning its fiber network for a world in which inference, agentic AI and physical AI distribute compute and data across enterprises, metropolitan areas and end devices. That changes the networking problem because AI traffic has to move between clouds and businesses, machines and models, centralized compute and distributed endpoints. The implication is that the infrastructure connecting compute to the broader economy becomes more important alongside the infrastructure connecting compute to compute.

For AT&T, that puts fiber at the center of its AI opportunity. The company is expanding its fiber footprint, investing in 400-gigabit-plus enterprise connectivity and adding higher-value intelligence on top of the physical network. The goal here is to translate deep fiber, network reach and customer relationships into differentiated, higher-margin services as AI workloads become more distributed.

Beyond data center interconnect

AT&T is participating selectively in the explosion of data center interconnect demand, but management is drawing a clear boundary around how central that business should become.

Speaking at the Goldman Sachs Communacopia + Technology Conference, AT&T CEO John Stankey said the company is again pursuing data center-to-data center and data center-to-metro interconnection where its network position makes sense. But he framed that as complementary to AT&T’s core franchise rather than its defining AI infrastructure opportunity.

“Our business should be built on the value of getting traffic to end-user customers and having a preferred position to do that,” Stankey said. “I think that’s the sustainable way to build a franchise that drives contribution and margin accretion into a company.” 

The operative distinction is that dedicated fiber between hyperscale facilities is an obvious beneficiary of AI infrastructure spending, but it is also a relatively narrow portion of the end-to-end network. AT&T’s addressable opportunity potentially becomes much larger as AI applications move outward into businesses, public-sector organizations, manufacturing facilities, cameras, vehicles and eventually increasingly autonomous physical systems.

At Citi’s 2026 Global TMT Conference, AT&T Business EVP and General Manager Melissa Arnoldi described enterprise demand already shifting in this direction. Customers are asking for 400-gigabit-plus connectivity, but capacity is only part of the requirement. They are also asking for resiliency and security, along with visibility into where AI agents operate, how they interact with employees and infrastructure, and how network resources are being consumed. AT&T is consequently investing in 400-gigabit capacity across multiple metro markets while developing software and intelligence capabilities on top of the network.

In other words, AI is creating demand for bigger pipes, as well as increasing the importance of understanding and controlling what moves through them.

Distributed AI changes the traffic model

There is also a more fundamental architectural argument underneath AT&T’s strategy. Telecom networks have traditionally reflected highly asymmetric consumer traffic patterns marked by users sending relatively small amounts of information upstream and receiving far larger volumes downstream. Fiber does not carry that same architectural constraint, and Stankey argued at Goldman Sachs that AI-related workloads will make uplink performance increasingly important.

Agentic solutions, robotics and video processing are useful examples. Cameras, machines and various classes of devices continuously generate data that has to travel toward compute resources for analysis. Stankey specifically highlighted those applications while discussing the need to engineer both fixed and wireless networks around greater uplink demand.

The broader point goes beyond any individual use case. AI inference is unlikely to exist in one location. Some workloads will remain in hyperscale clouds; some will run in regional or metro facilities; others will move onto enterprise premises or devices. Agents will interact with corporate applications and data spread across different environments. Physical AI will connect machines with centralized and distributed compute. That creates a much more heterogeneous traffic topology and potentially raises the strategic value of the network between endpoints and compute.

AT&T already owns several pieces of that path. Arnoldi pointed at business endpoints, metropolitan networks and AT&T’s nationwide backbone, while CFO Pascal Desroches argued at Citi that deep fiber gives AT&T an existing infrastructure base capable of supporting future workloads without requiring the same level of incremental network construction others may need.

The strategy appears to be based on AI becoming distributed enough that the network fabric connecting clouds, data centers, metros, enterprises and endpoints becomes a vital, and strategic, part of the broader AI stack.

AI gives AT&T’s fiber more jobs to do

The second-order opportunity is economic. AT&T was already investing aggressively in fiber before generative AI turbocharged infrastructure spending. Consumer broadband, business connectivity, wireless transport and convergence each independently support that investment. AI could increase the utilization and value of the same physical asset by giving fiber additional workloads to carry and additional services to support.

Desroches made the underlying cost argument at the Bank of America Media, Communications & Entertainment Conference. “The incremental cost of delivering a bit in fiber is lower than any other technology,” he said.

That means increased traffic can be economically attractive, but more bits alone are not enough. Arnoldi made that distinction particularly clear at Citi when asked about the large dark-fiber contracts being signed around AI infrastructure. AT&T will participate where it makes sense, but dark fiber itself does not necessarily offer the company’s most attractive margins.

“Where the margins get interesting is when you then sell on top of the dark fiber your lit services,” Arnoldi said. That suggests a fiber monetization stack. The physical infrastructure is the foundation. AT&T can then add lit capacity, high-speed enterprise connectivity, cloud access, resiliency and security before moving further up toward network visibility, control and intelligence.

The higher AT&T can move up that stack, the more consequential AI becomes to the return profile of its fiber investment.

More AI traffic does not automatically create carrier value

This is distinct from the more simplistic telecom AI thesis that AI creates more data, so carriers make more money. Bandwidth has historically commoditized, unit transport costs decline, and hyperscalers possess enormous purchasing power. It is entirely possible for network traffic to grow exponentially while much of the incremental economic value accrues elsewhere in the technology stack — which likely sounds familiar.

AT&T’s strategy depends on converting physical infrastructure into something more than transport. Management’s emphasis on network intelligence, security, visibility and performance suggests it understands that challenge. AT&T wants to participate in the data center buildout, but its more durable position comes from controlling a much longer portion of the path between compute and users.

That fits the company’s broader transformation. Stankey described AT&T’s end state at Goldman Sachs as essentially a metropolitan fiber provider combined with a nationwide wireless network, with legacy copper infrastructure and its associated complexity stripped away. AT&T’s September shareholder update similarly characterized years of fiber and 5G investment as creating a structural advantage for its next phase of growth.

AT&T does not need to become a hyperscale AI infrastructure provider to participate meaningfully in the AI economy. Its larger opportunity may emerge as AI leaves the data center. If inference, agents and physical AI distribute data and compute across enterprises and metropolitan networks, connecting those endpoints back to compute is a business unto itself. AT&T has spent years building precisely that layer. 

For a broader look at the role of fiber in the AI infrastructure supercycle, download this free RCRTech Trends piece, “The AI traffic effect: A new investment cycle for fiber networks.”