5
SK Telecom warns telcos not to repeat their 5G mistakes with AI-RAN – while Optus considers how they should re-architect their infrastructure as “social networks for agents”.
In sum – what to know:
Lessons and warnings – SK Telecom warns that AI-RAN could repeat 5G’s mistakes if operators prioritize technical capability over applications, demand, and sustainable revenue.
Traffic and architecture – Optus says AI will create new traffic patterns and network demands, from agent-to-agent comms and multimodal data to distributed edge inference.
Platforms and alliances – AI needs programmable platforms for traffic management; common architectures should stop vendor-specific AI-RAN becoming hard-to-scale silos.
Where were we? “You can’t evolve AI on pen and paper. You can’t evolve AI on simulations. You have to be in the field, [and] you have to get your hands dirty.” This was the line from Ericsson during a session with Optus and SK Telecom at Intelligent RAN Forum last week. As covered already, there were some practical examples (notably from Optus) about how AI is being put to work to raise efficiency in RAN operations – via link adaptation, coverage prediction, and coverage compensation. There was some smart advice too, and stark warnings (notably from SK Telecom), about the sector’s rangier ambitions to devise and monetize a new role in the AI infrastructure game.
And the one that will hurt most is the one that is most familiar: that the industry can’t mess up this time, like it did with 5G. “Remember 5G,” said Dr Dongwook Kim, director of the 6G tech team at SK Telecom. “The biggest risk [with] AI RAN is not technical failure, but repeating the mistakes [of] the early 5G era.” The industry lost its head, he warned – overly-enanoured by its tech, not mindful enough of its application. It is a classic cautionary tale of well-engineered technology, which applies just as well to IoT, in recent memory – or to Sony Betamax or Google Glass, or whatever else. With 5G, Kim pointed to (4G-era?) edge computing (MEC) and network slicing as examples of the same.
“The goal was to provide differentiated connectivity, including low latency and customized performance. But despite technical progress, [there has not been] significant revenue growth for MNOs,” said Kim. And then he put the boot in, effectively, with a hard truth for the sector: “There aren’t many services that truly require those advanced network capabilities.” Of course, one could make the argument, too, that once-hyped technologies like IoT and private 5G, closely related to network slicing, have reasserted some kind of buoyancy of late, partly buffeted by the prevailing hype around edge-based physical AI and robotics. Plenty in the industry that would make the case, certainly.
5G realities
Nonetheless, Kim’s message was a reality-check. “For AI-RAN… we should investigate services that are possible when [connectivity] works with AI inference in real time. Once they are defined, the network can evolve to support their requirements… People are discussing autonomous robots in factories and logistics centers, delivery robots in urban environments, and spatial awareness services based on sensing technology. These cases are attracting attention because they require both advanced communication and AI inference at the same time. Even so, it is still too early to say which will scale and become sustainable businesses. Practical verification is very important.”
The session last week split loosely into two parts, according to standard AI-RAN segmentation: ‘AI for networks’ (or AI for RAN), as discussed before, and ‘networks for AI’ (AI on RAN). The nature of discussion veered between the two respectively, as quite practical on one hand and rather speculative on the other. But the second half also dealt with the bigger questions, and the most pressing. “It’s good to imagine a world we’re going to be working in,” said Sriharan Amirthalingam, chief technology officer for networks at Optus. He explained, later: “We have gone from 1G voice to 2G SMS, to 3G browsing, to 4G video, to 5G.” (Note: 5G, unattached to a ‘killer’ app, is the odd one out.)
AI introduces a new traffic-type, the argument goes, whether on 5G or 6G (or fiber). “A token-based traffic: machines talking to machines, all the time,” said Amirthalingam. “Agents calling agents; a social network for agents…. With 4G we created a social network for people; we will [do the same] for agents.” He added: “There’s a token-economics coming.” But the architecture will have to change – as the industry is discussing, in private and at every telco forum programmable networks to respond to demanding apps; edge compute to run inference closer to the action, or where latency or privacy matters; platform orchestration to know where data is stored, accessed, exposed.
As above, the case for 4G/5G MEC-style compute gets remade with AI. “LLMs will do the reasoning, but the context and inference might be at the edge. Or you could have smaller LLMs at the edge to serve different applications.” Uplink traffic is a worry, as well – particularly as AI systems ingest video, images, and other sensor data; which could also change where computing happens. “It will certainly change the traffic flow, and traffic engineering.” Kim at SK Telecom said the same: “[There will be] many more physical and agentic AI devices – humanoid robots and AI glasses and AI assistants. Networks will need to handle [more] multimodal data and support real-time AI services.”
AI arbitrage
Amirthalingam described the shift as “AI arbitrage”. Not every problem needs a GPU cluster running a fully-featured frontier model. A small model close to the network will do for some workloads, not others. The task for telcos – matching the model, compute, location, capacity and latency to the workload – is no mean feat. They should consider the architecture now, even if the pieces are in flux. “We’re not going down a blind alley. The thing is to get started [and] work towards [it]. These are the elements that will shape AI on the network: use cases, sensing, agents, the social network for them.” Which poses a bigger architectural question: how to manage this rush of agents?
He sketched a hierarchy where agents interact with one another, in service of a “super agent” or “meta agent”, tasked with control of the wider environment. Which creates this concept of a knowledge plane to manage conflicts between them, and connect intent to outcome. It is speculative, still – as said at the top. But it sounds practical, as presented, and it points to a fundamental change for telcos, where they don’t just connect applications, but become a part of the AI system to enable and coordinate agents, between agents and machines. Which is the hype story they are telling right now. The discipline, as Kim said, is to make sure the apps stick this time.
Meanwhile, telcos have to make sure their magical tech does not only work in silos. Kim said: “The risk is every company develops AI RAN in its own way without common alignment. Today, each vendor has different hardware architecture and different approaches to AI RAN. Individual POCs may succeed in each company, but they are difficult to scale across the network.” The other risk, he suggested, is that all their good engineering work brings proper “operational value” to their own businesses – which is like the flipside of his previous warning about the outward-facing failure of 5G, and is covered already in the previous write-up about practical AI-RAN gains.
Kim said: “For operators, the most important things are not just spectrum efficiency or technical performance, but whether AI reduces operational costs, improves energy efficiency, and delivers a stable service.” But hence all the collaboration, in proofs and alliances – to face-off all of these risks. “We hope to create a practical result that can eventually lead to successful commercialization of the AI native network and especially AI RAN,” he said.
