Telecom operators could power the AI economy’s backbone, from fiber and data centers to GPU cloud services, in a market worth up to $70 billion by 2030.
For over a decade, telecom operators have watched a strange pattern repeat: every leap in tech, from smartphones to 5G, sent data traffic soaring, yet operator revenue barely budged. Between 2010 and 2022, global mobile data traffic grew roughly 60% a year while telecom revenue crept up at just 1% a year. Hyperscalers and other new entrants captured most of the upside.
A new industry analysis argues that AI could finally break that pattern, but only for operators who move with real speed and precision. The opportunity is described as building the physical and digital backbone of the AI economy, and it comes with four distinct paths, each carrying its own risk, investment size, and payoff.
Why This Time Might Be Different
The scale of what’s coming is hard to overstate. Global data center demand is projected to more than triple by 2030, reaching at least 170 gigawatts under a low-range scenario and potentially as high as 298 gigawatts under a high-range one. Even the midrange estimate implies a jump from 55 gigawatts in 2023 to 219 gigawatts by 2030, a 22% annual growth rate driven heavily by AI workloads.
Building and running that many data centers requires fiber to connect them, physical space and power to house them, and GPUs to actually run AI training and inference. Every one of those requirements is something telecom operators are, at least in principle, well positioned to provide.
Competition, however, is intensifying from cloud service providers, hyperscalers, cloud exchange operators, and other new entrants, all chasing the same opportunity. The report’s authors are candid that success is not guaranteed, and that delays or missteps could leave telcos even further behind than they already are.
Path One: Getting Paid to Connect the New Data Centers
Hyperscalers and colocation providers have announced plans to break ground on more than 2,600 new data centers, and roughly a quarter of them will land in cities that have no data center infrastructure today. By the early 2030s, the total number of data centers operated by colocation firms and hyperscalers combined is expected to approach 11,000 worldwide.
None of those facilities are useful in isolation. They need to be connected to each other and to end users by fiber, and that connectivity work is projected to create a $30 billion to $50 billion global market opportunity.
Operators have real leverage here. In much of Europe and Asia-Pacific, hyperscalers need a telco’s license just to lay fiber, which hands existing operators first pick of which markets to invest in directly and which to lease out. In the United States, operators with existing network footprints may be able to sidestep a lot of the permitting and construction headaches that come with building from scratch.
Hyperscalers, for their part, often prefer to lease or buy existing “dark fiber,” meaning unused fiber optic cable already in the ground, rather than build their own. Verizon’s infrastructure agreements with Google Cloud and Meta, and Lumen’s fiber deal with Microsoft that includes dark fiber access, are cited as examples of this dynamic already playing out.
Path Two: Selling Smarter Networks, Not Just Bandwidth
As enterprises push more AI workloads onto the cloud, their networking needs get more complicated. Businesses increasingly want intelligent network services that let them dynamically route AI workloads to meet regulatory requirements, monitor and reduce the data transfer fees known as egress costs (estimated to run $70 billion to $80 billion a year industry-wide), and automatically request more network capacity for latency-sensitive jobs like AI inferencing.
That last point matters more than it might first appear: AI inferencing is projected to account for the majority of AI workloads by 2030, up sharply from just 15% to 30% of workloads in 2023.
For telcos, this represents a chance to reverse a decade of shrinking B2B wireline revenue by shifting from flat monthly billing to value-based pricing for these smarter network capabilities. Lumen’s ExaSwitch platform, which lets customers configure and route their own network traffic through a self-service portal, and British Telecom’s managed security and networking service for corporate clients, are both offered as working examples.
The catch is that this market is still being defined. There’s no consensus yet on what features customers actually need or what business model works best, which means operators pursuing this path will likely need real investment in market research before committing to a product strategy.
Path Three: Turning Empty Space Into Rent
Many operators already sit on underused data center space and central office real estate. Instead of selling it off for residential or commercial redevelopment, some are finding it more lucrative to lease that space and power directly to hyperscalers, colocation providers, GPU cloud firms, and large enterprises that need capacity immediately.
The timing favors telcos: new data center construction can take upwards of five years, and power grids in many markets are already running at capacity, unable to support new builds. That combination gives operators sitting on existing space and power a real, if temporary, advantage.
Verizon’s mobile-edge computing partnership with AWS illustrates one version of this model. Verizon supplies the space and power, AWS brings the compute and customers, and Verizon stays actively involved in running AI workloads over its own network rather than acting as a purely passive landlord.
Not every property is equally attractive for this purpose, though. Data centers that support AI inferencing at the network edge typically need at least 500 kilowatts of on-site power capacity to make economic sense, and spaces built for older, air-cooled server racks often require retrofitting with liquid cooling equipment to handle the density of modern AI hardware.
Path Four: Building an Actual GPU Cloud Business
The most ambitious path is standing up a GPU-as-a-service (GPUaaS) business: renting out remote access to high-performance GPUs, either by the hour or on longer reserved contracts, so customers can train or run AI models without buying their own hardware.
The addressable market here, specifically the slice available to telcos rather than hyperscalers, is projected to range from $35 billion to $70 billion globally by 2030, roughly 2.5 times its expected 2025 size. Most of that demand is expected to come from North America and Asia.
Some operators are partnering their way in, like Verizon’s tie-up with GPUaaS players in North America. Others are building and marketing their own offerings directly: Indosat Ooredoo Hutchison, Singtel, and Softbank have each launched GPUaaS products, often in partnership with GPU makers.
Norway’s Telenor is positioning itself around sovereign AI for the Nordic region through a collaboration with NVIDIA, while Swisscom’s AI platform lets companies keep their data processing inside Switzerland, both plays aimed at governments and enterprises wary of storing sensitive data with non-national providers.
The financial case has to be built carefully. Whether an operator leverages its own existing data centers or rents space from a colocation provider, projected returns on invested capital for GPUaaS ventures range from 6% to 14%, depending heavily on the operator’s existing asset base and local electricity costs.
A related, more experimental idea covered in the report is “AI-RAN,” which would repurpose the GPU-based hardware inside mobile network equipment to also run AI workloads when it isn’t busy handling network traffic. The concept is still being tested, and its economic viability will vary significantly by operator, but early industry vendor testing has pointed to improvements in network efficiency alongside the added revenue potential.
It’s Not Just About the Technology
Pursuing any of these four paths, the report argues, requires real operating model changes, not just new products bolted onto existing sales teams. Operators selling space and connectivity to hyperscalers will likely need dedicated sales teams built for fast, specialized deal cycles, while those selling GPUaaS to enterprises need a shift from traditional solution-selling to a more advisory, consultative sales approach.
New partnerships also become essential, spanning GPU makers, data center developers, power suppliers, systems integrators, and specialized construction contractors, since few operators have all the in-house expertise to build a full-stack AI infrastructure offering alone.
On the financial side, evaluating these opportunities often means abandoning traditional return calculations in favor of longer-view models, like predicting demand for connecting multiple data centers along a shared fiber route rather than judging a deal purely on its first customer.
Finally, the report is blunt that this is a communications challenge as much as an operational one. It points to AT&T’s CEO publicly discussing the company’s AI strategy, Lumen highlighting AI-relevant network performance metrics on its own website, and Verizon foregrounding its AI plans on recent earnings calls.
These are all examples of operators trying to actively shape how investors and enterprise customers see their role in the AI economy, rather than waiting to be asked.
The Real Risk Is Standing Still
Not every operator will find every path attractive, and the report is careful to note that some options may simply be too risky for certain companies right now given their asset base, market structure, and appetite for risk.
But the clearest message is that inaction carries its own cost. Telecom operators missed out on the bulk of the revenue created by the smartphone, video-streaming, and 5G waves. This report frames AI infrastructure as a chance to not repeat that pattern, provided operators move with real urgency.