During the Age of AI, Telco Late Adopters May Be the Winners of Tomorrow

Frontier Artificial Intelligence (AI) models are now operating in large data center clusters, where processing concentration rewards the economics of running an AI business. Graphics Processing Unit (GPU) utilization and cost-per-token improve with scale and that is why there is a shortage in AI processing that is expected to last many years into the future. It is also the reason xAI decided to lease its spare GPU capacity to Anthropic and Google, making US$2 billion per month while powering its massive data centers with gas turbines that have not received a permit to operate. Training frontier models are what set AI providers apart and this business is now riding a massive investment and innovation wave. Inference is now a minute part of this business and centralized, because the mainstream applications that these AI companies provide are neither latency sensitive nor require data sovereignty. However, there are market indications that interest in inference is increasing; for example, NVIDIA releasing inference-specific servers (Groq 3 LPX).

There are a few applications that require low-latency inference, as low as 1 Millisecond (ms). This means that inference must take place very close to the application and typically must not be more than one network hop away. In telco terms, this translates to near-edge (central office or aggregation point), or even far-edge (cell site or street cabinet). This typically applies to small-model, low-latency applications like robotics, machine vision, and autonomous driving, all of which are prospects and being driven by the telco narrative to reframe idle real estate as strategically important for AI inference.

Telefónica has repurposed its former copper exchanges to become edge locations and now offers this capability in 17 nodes across Spain. On top of this, Deutsche Telekom, Orange, TIM, and Vodafone have announced the European Edge Continuum (EEC), a cluster of edge nodes across Europe, all of which can be accessed by the edge portal of any of these operators. The European arm of the Linux Foundation has also created NeoNephos, an umbrella project that includes several open-source projects that aim to progress the objectives of the European Union’s (EU) Important Project of Common European Interest on Next Generation Cloud Infrastructure and Services (IPCEI-CIS). The EU is trying to create its own cloud and AI sovereign stack, deployed across telco core and edge locations. Assuming that all four operators have similar assets in their home countries, and that each of these edge locations can host 2 Megawatts (MW) of inference capacity, then the collective edge inferencing capability of Telefónica and the EEC is 34 MW and 136 MW, respectively. These locations can run expert Small Language Model (SLM) and Visual Language Model (VLM) inference at scale and become the proof point for telco inference. As shown in ABI Research’s Telco Tokenomics report (AN-6610), cell-site inference can result in a 4-year Return on Investment (ROI), after which token generation in the network becomes profitable. The business case for central office inference will be even easier to prove.

Moreover, other operators, including NTT and SK Telecom, are progressing with their AI strategies and testing new deployment architectures. For example, NTT DOCOMO Business is trialing a distributed AI training concept using its own All-Photonics Network (APN) across its GPU locations across eight cities, meaning it can potentially deploy a geographically distributed large cluster. Tier-One operators are not sitting idle, but they are actively exploring opportunities in the AI market. The question is, what are their key assets and what are the biggest challenges they are facing?

Core and central office locations are prime opportunities for edge AI inference, both of which can be used as strategic foundations to breathe new life into their business model. ABI Research has identified four potential AI strategies for telcos going forward:

1. Connectivity Provider: This option maintains the status-quo and most telco operators remain in the business they are in today: to provide fixed and mobile connectivity at the risk of increased competition, lower profit margins, and a capital-intensive business.
2. AI Landlords: Telcos can lease their core and central office real estate to third parties, especially enterprise vertical solution providers with applications that require locality, low latency, and data sovereignty.
3. AI Infrastructure Providers: Telcos deploy AI infrastructure themselves—in the same manner Telefónica has deployed its edge locations—and lease capacity to third parties, including AI hyperscalers.
4. AI Application Enablers: Telcos host both infrastructure and software platforms for AI applications. 6G discussions in The 3rd Generation Partnership Project (3GPP) are starting to include concepts like Tokenization-as-a-Service (TaaS) and Application Programming Interface (API) access to AI infrastructure.

Capital requirement, effort, and complexity increase exponentially when going from option 1 to option 4, but so do profits. However, the biggest challenge that telcos will have in this thesis is to prove the business case and wait until these Physical AI applications enter the market. Another challenge is power and cooling in the central office locations, some of which may still be running their fixed or mobile network exchanges. The early adopters that have sold off their data center assets and central office locations may be at a considerable disadvantage compared to other telcos that own these assets and can modernize them for AI. Nevertheless, many telcos will likely invest in multiple strategies in this list, depending on their assets, willingness to risk, industry partnerships, and market location.