AI-RAN gets real

From Intelligent RAN Forum: AI-RAN is becoming less futuristic and more practical, with Optus reporting live-network gains in link adaptation, coverage prediction and compensation. But scaling those AI-for-RAN capabilities across multi-vendor networks remains a challenge, notes SK Telecom. In sum – what to know: Useful RAN work – Optus is deploying AI for dynamic link adaptation, coverage prediction, and automated coverage compensation with Ericsson, reporting major field gains in spectral efficiency. AI-native ambition – More speculative networks-for-AI projects are a while off; the near-term value is improving existing network performance and operations, without requiring new RAN hardware. Scaling problem – Different hardware and architectures make isolated POCs difficult to scale, increasing the importance of interoperability, common validation methods, and industry collaboration. One could make the case that the interesting thing about AI-RAN right now is that it is less interesting. Not because the tech is unimportant, or any less important than it was, but that the conversation is developing along two paths: the practical and the fanciful, and the former – arguably, probably sensibly – holds more sway. Yes, the industry is still rubbing its hands, in desperately excitable fashion, at the prospect of inward-looking gains from ‘level-five’ AI autonomy and ‘AI-native’ 6G, and outward-looking gains from all the physical and agentic AI that will be stood-up on ‘AI grid’ infrastructure someday. But it is also asking itself, very seriously: what can AI do in a live network now? Which was a major highlight of a discussion between Ericsson, Optus, and SK Telecom at Intelligent RAN Forum this week (September 22, available on-demand). And it was interesting – to hear the two operators, prompted by Ericsson, present their (Ericsson-backed) AI-RAN gains in their live networks, in Australia and South Korea. There is a helpful distinction, of course, to keep in mind: AI-RAN splits roughly as AI-for-networks, where AI is embedded in the network for its own ends, and networks-for-AI, where the network is adapted and rebuilt to support new AI traffic and apps – to deliver telcos to a promised land of AI milk and money. Like a version of the 5G dream, on steroids. Speaking on another session at the RCR event, Dr Alex Jinsung Choi, research fellow at Softbank, called it a little differently: AI for RAN to optimise radio performance, AI on RAN for running AI workloads in the network, and, nominally, AI and RAN as a shared compute fabric between the two. The point is only that the first category, AI for networks / AI for RAN, is where most practical value is emerging. The rest of it, including more ambitious pursuits in the network domain, will be covered in a separate post. (The message, there, is that the telecoms industry’s AI journey might yet go anywhere, but that it should also learn from its 5G car-crash – and not mess up again.) As for progress with AI-for-networks, or AI-for-RAN, there are already some fairly concrete examples. Sriharan Amirthalingam, chief technology officer for networks at Optus, provided three good ones: link adaptation, coverage prediction, and coverage compensation. See: the telco tech-verbiage makes it seem uninteresting, maybe; except each is ‘dynamic’ and ‘automated’, to some degree, and all sound like giant leaps in their own right. Amirthalingam said: “If you look at customers milling about at Circular Quay in Sydney, say, you look at their link and apply AI to optimize their experience. Which is different to what you do currently with ML algorithms – because, with AI, you adjust the antenna gains as well as modulation techniques. This is where the collaboration comes in.” The collaboration reference is to its work with Ericsson on the solution; the panel was convened and steered by Ericsson, and all the triumphs here invariably involve the Swedish firm. As an aside, Amirthalingam quite helpfully categorised AI RAN roles and responsibilities into “three buckets”: “The stuff the vendors do, which only they can do – and we shouldn’t – [related to] the software, the scheduler, stuff around spectral efficiency, energy savings, uplink optimization; some operational domains like optimization, [and] plan-design-deploy [which operators are directly involved in]; and [ecosystem work on] operational features in the RAN… to marry the intelligence of the network with the AI.” Optus is rolling out Ericsson’s ‘dynamic link adaptation’ solution across its network in Australia, he said. The solution adjusts transmission rates, modulation schemes, and coding to match changing channel conditions; it integrates neural networks directly into baseband or cloud hardware; it is supposed to deliver a throughput boost of up to 20 percent, and a spectral-efficiency gain of 10-15 percent. AT&T, Bell Canada, and T-Mobile are also testing it, variously. Melike Erol-Kantarci, strategic product manager for AI RAN at Ericsson, moderating the session, suggested the spectral gains might be even higher. “Being able to see those gains in the field – up to 10% spectral efficiency across the sites and even up to 25% in certain places, in certain conditions – these are huge numbers.” Another Optus field project, for coverage prediction, has yielded “eye-opening” gains, she said. Back to Amirthalingam: “We are working to see how to use AI to optimize handovers so you can drag the signal further than you could, and provide a better experience on call continuity. So AI takes the decision away from the handset – about which cell to go to next. The AI suggests the best signal, and the next cell. So you’re enhancing existing features with AI.” His third AI-RAN result, for ‘coverage compensation’, puts the manual process to “change the tilts and pans and antenna gains” when a base station goes down into the hands of an AI agent. “When the base station comes back, [it] automates the reversal,” he said. “The network becomes a living organism – where it knows it is hurt [somewhere] and it can compensate and self-heal, and then go back to where it was before.” He added: “Those are three examples – where we are trying to collaborate and bring in the third category of AI into the RAN… To me, those are bird-in-hand things we can do now while we work for a longer-term orchestrated AI engine.” Bird-in-hand, but not fluffy stuff either – if call-centre sales and support, away from the network, might be (dubiously) dismissed as such. He said: We are not talking about front-of-house processes – activation, assurance; park that. This is [about the] network, in two categories: everything in the operations, whether in network operations or field operations, [where] you’ll have all kinds of different AI to help diagnose, recover, manage the network; and then you’ve got domain AI in the RAN, in the IP, in the core, [which all has to be] cross-domain.” The latter requires “AI orchestration and meta structure”, and will be dealt with in the write-up, tomorrow or next week, about the funner future-gazing (as perceived) of networks-for-AI. Ericsson’s new AI-in-RAN software package, launched in June, embeds telco-specific AI models into existing 5G basebands and radios – with all the promised throughput and spectral gains discussed above. Kantarci said the product, available on subscription, was designed around field proof points that did not require additional hardware and could run across the operator’s installed base. So no new servers, accelerators, RAN equipment – just yet; to eke out significant operational improvements. Just new software, applied to existing infrastructure; and then lots of kicking the tyres. Erol-Kantarci said: “You can’t evolve AI with a pen and paper. You can’t evolve AI on simulations. You have to be in the field, you have to get your hands dirty, and you have to work with your customers.” But there was a warning, too, from SK Telecom about that classic AI-era schism for telcos, about proprietary kit and operational silos, and the challenge to achieve interoperability and scale. “Every partner has a different hardware architecture, and a different approach for AI RAN implementation. Even if one POC is successful with a specific vendor platform, it does not automatically scale across the whole network. Technologies that work well in one environment may not work the same way in another,” said Dr Dongwook Kim, director of the 6G tech team at the Korean operator. Going from virtualised RAN towards AI-native RAN means bringing together RAN software, operator data, AI platforms, and AI models, he said. Industry collaboration, including work through the AI-RAN Alliance, is crucial for agreeing common use cases, data points, and validation methods. “So the industry can build a more open and reusable ecosystem, instead of creating isolated POCs that are difficult to scale.” There is another issue, too: an operational telco ecosystem where multiple AI agents work together will require an orchestration layer to manage all the agents, and resolve conflicts and turn intent into outcomes. Which is a bigger proposition, and another story – to be covered next time.