For the past couple of years, the telco industry has arguably watched from the sidelines while hyperscalers and chipmakers absorbed the AI infrastructure spend. Now the operators are moving. Verizon has struck a $1 billion data center interconnect deal with Google, with cloud and edge sales expected to ramp in 2027. SK Telecom went a step further and created a sub-brand, SK Hyper, to run its AI data center business end to end.

The reason telcos suddenly have a seat at the table comes down to the shift from training to inference. Training happens in a handful of giant, centralized facilities; inference happens everywhere, all the time, close to the user. That’s distributed infrastructure — data centers, metro sites, edge premises, and the connectivity between them — which happens to be exactly what telecoms already own. AT&T made the point explicitly on its Q2 earnings call, arguing its network is already built for the agentic AI wave. Notably, its optimization prioritizes upstream traffic over the download speeds previous-generation networks were built around, and the strategy treats inference as a primary network workload with orchestration, rather than bigger models, as the real breakthrough.

That upstream theme shows up again on the radio side. Huawei and China Unicom Beijing have commercially deployed what they describe as the world’s largest 5G-A 100 MHz GigaUplink network, and Huawei projects mobile AI will drive a whole new wave of uplink investment globally. Take the projection with the usual vendor-forecast grain of salt, but the direction of travel matches what AT&T is saying independently — networks designed for consumption are being reoriented for devices that generate and send data as much as they receive it. Verizon, meanwhile, is pointing its AI ambitions inward as well as outward, with CTO Yago Tenorio targeting Level 4 automation in critical parts of the core network, built on the contextual data only an operator has.

Of course, not every shiny piece of infrastructure is ready for prime time. HKT’s 3.2Tbps hollow-core fiber “superhighway” is a genuinely impressive showcase of ultra-low-latency DCI — but analysts warn that at-scale adoption could be a decade away. It’s a useful reminder that in this market, the gap between a working demo and a deployable business case can be very wide indeed.

Read more below.