AI RAN (Radio Access Network) is attracting major investment from telecom providers as they seek to improve efficiencies to reduce their expenditure, with Juniper Research recently predicting that operators would plough $21 billion into AI technologies this year.

While the tech provides networks with a greater degree of autonomy, the application differs to true autonomous networks. To learn more about how AI RAN is defined, we spoke to Juniper Research and Ericsson, a major player in AI RAN.

What is AI RAN?

Sam Barker of Juniper Research concedes that there isn’t really a market definition, but the term AI RAN essentially denotes any time that AI is used at the edge of the radio access network for automation and efficiency to enable cost savings. He notes that while many operators and vendors have gone down their own paths with varying results, there is not much in the way of definitive standards.

Zoran Lazarevic, Chief Technology Officer at Ericsson Middle East and Africa, explains that the Swedish vendor defines the term as the use of machine learning directly in the radio layer to optimise things like scheduling, interference management, coverage and energy consumption in real time; it can be thought of as focused, domain-specific intelligence working inside the RAN itself.

Autonomous networks, on the other hand, describe the broader vision that operators are working toward: networks that configure themselves, optimise their own performance and heal their own problems across RAN, core, transport and operations.

“In practice, AI RAN is one of the key building blocks that makes that autonomy possible”, says Lazarevic. “It brings intent-based automation and closed-loop control to radio functions, which helps operators deliver predictable performance and differentiated connectivity. The autonomous network vision then extends those same principles, policy management, service assurance, lifecycle orchestration across the entire network stack.”

Barker acknowledges that there is overlap between the concepts of AI RAN and autonomous networks, but agrees that the latter is a wider definition in that it covers the entirety of the network, including the core.

“When I think of autonomous networks, I think of things like more resource allocation, so taking things away from geographical areas. If there's loads of strain on the network that wasn't really anticipated, an autonomous network can do all of the network, whereas AI RAN is just for the edge of the network.”

Automated efficiency

In terms of enabling efficiencies, Barker notes that the first point to make is that RAN probably consumes the most energy for a network. The larger a network’s geographical coverage, the more base stations it will require – and so the more RAN. Using AI to automate the management of power consumption across the base stations can reduce the energy usage during downtime, or when the network is idle.

Lazarevic agrees that energy is a key efficiency that AI can enable; AI-driven resource allocation and intelligent sleep modes can cut RAN energy consumption by up to 40% when paired with sustainable network technologies. Energy optimisation directly addresses one of the largest lines in operating budgets - RAN’s power consumption is about 70% of its operating expense. He notes that AI can put radios into sleep mode when traffic is light, fine-tune coverage and eliminate wasted energy, meaning power consumption is actually managed better rather than made worse.

On the operations side, automated parameter tuning, anomaly detection and predictive maintenance reduce manual work and the number of site visits required, as well as lowering failure rates and avoiding unplanned outages. Performance is also improved with smarter cell coordination and load balancing that boost throughput and latency, directly improving the customer experience.

Spectrum and site utilisation can also be made significantly more efficient via AI.  More intelligent scheduling allows capacity to be used more effectively – and essentially increased - without a proportional increase in capital expenditure.

“Cloud-native AI applications scale compute resources up or down based on actual demand, so you are not stuck paying for capacity you do not need”, notes Lazarevic. “Business agility is also a major beneficiary, with intent-based automation enabling the delivery of predictable performance tiers that align with SLAs and specific use cases. This speeds up time to value for both consumer and enterprise services.”

Barker agrees that AI can enable further cost savings by detecting equipment failures and network anomalies sooner, and adds that it can turn infrastructure into a revenue generating asset via inferencing and edge computing. On the other side, by enabling more virtualisation at the edge of the network, AI reduces reliance on hardware and thereby reduces capital expenditure. However, Barker underlines that the main benefit is via automation, whether for shutting down the network, improving spectrum efficiency, or automating resource management so it can be done quickly with minimal human oversight.

“The impact of AI at the edge of the network wholly depends on how virtualised the network is itself. In developing markets, it's going to be less virtualised, it's going to be older technologies, and AI is not going to be able to automate as many of the processes, the functions on a network as something from 5G - or in the future 6G - because it's not software defined”, says Barker.

“AI can’t look at it and say, ‘I can shut that part of the network down for now to save a bit of money’, because it's not virtualised. So in developing regions where they do lag behind on 5G - 4G is still being rolled out in some places - there is an issue with ROI. There's going to be less incentive for the operators in developing regions to invest in AI RAN, because they'll get fewer benefits, less savings, so that ROI is further away than [for operators] who have more virtualised networks.”

Lazarevic argues that on the revenue side, there are factors that, taken together, will shorten ROI cycles while improving service quality, claiming that exposing network capabilities through APIs opens the door to performance-based products and network slices that improve margins without linear cost increases. In this way, AI RAN can help to transform networks into a programmable, data-driven asset that costs less to operate and is easier to monetise.

AI RAN needs two main components: energy efficient compute at the radio sites to run AI features, and a central control layer that sets policies and learns from the data the network generates. Most of the decision-making happens at the edge, close to the radios themselves. The heavier work, analytics and model updates, sits in the core or cloud, which means operators do not need to build massive new data centres right from the start.

AI RAN in emerging markets

For emerging markets, adoption can be modular. Operators can begin by adding energy-saving software to existing sites, then gradually scale toward cloud-based RAN functions. Options like vendor-managed services, shared infrastructure and solar or hybrid power setups can lower the upfront investment.

Barker argues that it will be much easier for operators in developed markets to secure a return on investment into AI RAN, noting that the main issue is geographical reach since the AI sites are at the edge of the network.

For operators in smaller European countries, less investment is needed at the base station to enable AI RAN - even fitting a new bit of hardware is easier when there's fewer base stations. In larger countries, deploying to every base station at once is not feasible – it will take time.

Barker also suspects that AI RAN will not be a priority for many operators in developing countries – not for technological or even monetary reasons, but because operators in developed markets are already focusing more on the enterprise side. In the developing world, there’s still room for growth on the consumer side, and Barker expects operators to focus on increasing revenue from this segment before eventually moving towards enterprise.

“AI RAN can benefit the enterprises more than the consumers, from a user perspective. It's not really a big sell to consumers saying that you might get slightly quicker mobile internet speeds if they introduce AI RAN, but for enterprises that could be very important, especially in places like smart cities where latency might be really key for new services - and AI can handle that. It should eventually start lowering costs for these users, consumers and enterprise users as well. What we're seeing when we go to events and we speak to operators, is that AI RAN is part of a wider efficiency strategy, making their networks more efficient to save costs. In the long term, that's not necessarily a strategy for operators in developing countries.”

Lazarevic counters that AI RAN is central to 5G Advanced, and in many ways the foundation for 6G, so operators who embed the technology now gain a real competitive advantage in markets where 5G is scaling quickly. By waiting, there is the risk of higher operating costs, slower feature development and weaker positioning in mission-critical and industrial segments.

“The operators who come out ahead will be the ones who combine programmable networks, AI automation and open APIs to turn connectivity into a platform for innovation and new revenue. The direction is quite clear, and the gap between early adopters and everyone else is only going to widen.”

Barker is perhaps more sanguine about the opportunities presented by the technology. “Operators in different countries face different market opportunities and different challenges. It's not necessarily about catching up, but they're all in their own unique situations. They're all going to have their own different strategies for profit maximisation or revenue maximisation. For operators in developed countries, AI RAN is a much more valuable proposition for them than operators in developing countries.”

AI RAN will of course be part of a wider strategy for operators looking to make networks more efficient, and Barker notes that he expects AI to be a huge part of 6G alongside satellite.

AI RAN is going to be a technology that becomes a necessity, rather than something that can save money. As networks get more complicated to manage, and more challenging to maximise for efficiency, human intervention is not going to be enough. It's a case of how much access AI models are given to automate certain network functions - they can implement the technology, but in terms of what AI can and can’t change, the growth in trust will be slow.