Telcos are using gen AI to reduce costs, become more efficient, and offer new services. Some are building new gen AI data centers to sell training and inference to others. What role does connectivity play in these data centers?

There is a gen AI gold rush expected over the next five years. Spending estimates range from hundreds of billions to over a trillion dollars on the physical layer required for gen AI: chips, data centers, and electricity.¹⁶ Close to another hundred billion US dollars will likely be spent on the software and services layer.¹⁷ Telcos should focus on the opportunity to participate by connecting all of those different pieces of hardware and software. And shouldn’t telcos, whose business is all about connectivity, be able to profit in some way?

There are gen AI markets for connectivity: Inside the data centers there are miles of mainly copper (and some fiber) cables for transmitting data from board to board and rack to rack. Serving this market is worth billions in 2025,¹⁸ but much of this connectivity is provided by data centers and chipmakers and have never been provided by telcos.

There are also massive, long-haul fiber networks ranging from tens to thousands of miles long. These connect (for example) a hyperscaler’s data centers across a region or continent, or even stretch along the seabed, connecting data centers across continents. Sometimes these new fiber networks are being built to support sovereign AI—that is, the need to keep all the AI data inside a given country or region.

Historically, those fiber networks were massive expenditures, built by only the largest telcos or (in the undersea case) built by consortia of telcos, to spread the cost across many players. In 2025, it looks like some of the major gen AI players are building at least some of this connection capacity, but largely on their own or with companies that are specialists in long-haul fiber.

Telcos may want to think about how they can continue to be a relevant player in the part of the connectivity space, rather than just ceding it to the gen AI behemoths. For context, it is estimated that big tech players will spend over US$100 billion on network capex between 2024 and 2030, representing 5% to 10% of their total capex in that period, up from only about 4% to 5% of capex for a network historically.

Where the opportunities could be greater are for connecting billions of consumers and enterprises. Telcos already serve these large markets, and as consumers and businesses start sending larger amounts of data over wireline and wireless networks, that growth might translate to higher revenues. A recent research report suggests that direct gen AI data traffic could be in exabyte by 2033.²⁴

The immediate challenge is that many gen AI use cases for both consumer and enterprise markets are not exactly bandwidth hogs: In 2025, they tend to be text-based (so small file sizes) and users may expect answers in seconds rather than milliseconds,²⁵ which can limit how telcos can monetize the traffic. Users will likely pay a premium for ultra-low latency, but if latency isn’t an issue, they are unlikely to pay a premium.

A longer-term challenge is on-device edge computing. Even if users start doing a lot more with creating, consuming, and sharing gen AI video in real time (requiring much larger file transmission and lower latency), the majority of devices (smartphones, PCs, wearables, or Internet of Things (IoT) devices in factories and ports) are expected to soon have onboard gen AI processing chips.²⁶ These gen accelerators, combined with emerging smaller language AI models, may mean that network connectivity is less of an issue. Instead of a consumer recording a video, sending the raw image to the cloud for AI processing, then the cloud sending it back, the image could be enhanced or altered locally, with less need for high-speed or low-latency connectivity.

Of course, small models might not work well. The chips on consumer and enterprise edge devices might not be powerful enough or might be too power inefficient with unacceptably short battery life. In which case, telcos may be lifted by a wave of gen AI usage. But that’s unlikely to be in 2025, or even 2026.

Another potential source of gen AI monetization is what’s being called AI Radio Access Network (RAN). At the top of every cell tower are a bunch of radios and antennas. There is also a powerful processor or processors for controlling those radios and antennas. In 2024, a consortium (the AI-RAN Alliance) was formed to look at the idea of adding the same kind of generative AI chips found in data centers or enterprise edge servers (a mix of GPUs and CPUs) to every tower.The idea would be that they could run the RAN, help make it more open, flexible, and responsive, dynamically configure the network in real time, and be able to perform gen AI inference or training as service with any extra capacity left over, generating incremental revenues. At this time, a number of original equipment manufacturers (OEMs, including ones who currently account for over 95% of RAN sales), telcos, and chip companies are part of the alliance. Some expect AI RAN to be a logical successor to Open RAN and be built on top of it, and may even be what 6G turns out to be.

…………………………………………………………………………………………………………………………………………………………………………….