Telecom networks are entering a new phase where intelligence is built into the architecture. The focus is shifting from pilots to artificial intelligence (AI)-native networks, where automation, data-driven optimisation and cost efficiency are integrated into one system. At the core of this shift lies the principle of total cost of ownership. Every decision, from silicon to software, is now assessed on the basis of its impact on power, cost and reliability. AI fits naturally into this framework, not as an experiment but as a driver of efficiency. Virtualisation enables this through a by moving workloads from proprietary hardware to commercial off-the-shelf compute, redu­cing the capex and opex while paving the way for AI-based functions such as planning, fault prediction and energy management.

Within the radio access network (RAN), the demand for intelligence is increasing. Platforms, silicon and software are being designed to be AI-ready, optimised to run compact models efficiently on general-purpose processors rather than relying on power-intensive graphics processing units (GPUs). This evolution is making compute systems more distributed, affordable and adaptable across the network. For instance, in high-density, interference-prone areas, predictive AI is already being explored to anticipate interference, forecast ducting conditions and shift users intelligently between ­frequency layers. Models like these demonstrate how networks can evolve from reacting to problems to preventing them altogether.

Evolving architecture

AI within RAN is evolving in three distinct ways. The first, AI for virtualised RAN (V-RAN), focuses on training and processing models off-site and then deploying them into RAN for non-real-time optimisation. The second, native AI within the baseband, embeds intelligence directly into the system, enabling real-time decision-making but demanding higher processing precision. The third, AI on RAN, positions GPUs or dedicated accelerators alongside the baseband to manage local inference and transfer learning.

Each approach represents a different balance of latency, power and cost, and most operators are combining them to meet their specific network goals. As 5G expands and open RAN (O-RAN) matures, standardisation has become the foundation of progress. Aligning the rapid evolution of AI with the slow development of telecom standards remains a challenge, but open interfaces such as A1, O1 and E2 are bridging that gap. They make it poss­ible for hardware and software from multiple vendors to operate together while preserving reliability and compliance. This modular architecture is accelerating innovation. AI-driven RAN controllers, self-optimising systems and closed-loop automation can now evolve independently and integrate through a standardised applications programming interface. Intelligent RAN extends from the application layer to Layer 1, embedding adaptability across the full stack.

The guiding principle is “sense, think, act”. Observability data feeds into AI and machine learning engines that detect changes in the network, classify traffic patterns and trigger automated actions. Re­inforcement learning continuously fine-tunes these responses, striking a balance between performance, latency and energy use. Early results from deployments indicate substantial gains in efficiency, faster issue resolution and notable improvements in throughput, demonstrating that intelligence within RAN is no longer theoretical but operational. With these architectural foundations in place, intelligence is moving deeper into operations, teaching networks to not just perform but also predict.

Networks are learning to anticipate repair works. Self-organising networks (SON) provide this foundation by collecting radio key performance indicators (KPIs) and user experience metrics that feed directly into AI models. What began as automation for load balancing and self-healing has now ­expanded into intelligent energy management. One example is the intelligent dynamic energy saving model, built to learn demand patterns and adjust power use dynamically without affecting performance. Such systems mark a clear change in how networks operate, continuously sensing, predicting and optimising in real time.

Energy and spectral efficiency have become the two most visible outcomes of AI. Radios that once transmitted at full power through quiet hours now scale down auto­matically. Predictive algorithms replace static rules with learning-based control, reducing unnecessary energy consumption. In some operator deployments, early versions of these models delivered around 20 per cent savings. As predictive control matures, the improvement has multiplied several times over. The same approach is improving spectrum use. AI-based channel-estimation models extract more ­capacity from existing bandwidth, driving spectral efficiency by roughly 10-15 per cent.

Automation and monetisation

AI’s reach in telecom depends on how intelligently compute is distributed across layers. Centralised cloud RANs manage large-scale training and analytics, while edge and radio sites handle low-latency inference closer to the user. GPUs anchor data centre workloads; CPUs manage aggregation; low-power systems on chips and field programmable gate arrays operate at the edge, each chosen for its balance of latency, energy and cost. This distributed compute is redefining value.

During off-peak hours, idle edge resources can be reused for AI inferencing or training-as-a-service, turning spare ­capacity into revenue. These emerging concepts highlight the evolution of infrastructure from a cost centre into an on-demand platform. Intelligent RAN extends this logic through dynamic resource allocation. Physical resource blocks and network slices can now adjust in real time to match application priorities, offering differentiated connectivity for enterprises, mission-critical systems or premium users. While the technical groundwork is in place, policy frameworks will determine how far this flexibility can extend.

At the network scale, automation is becoming the default mode of operation. Classification algorithms, ranging from decision trees and random forests to long short-term memory (LSTM) predictors, are steering multiple radio access technology (multi-RAT) traffic between long term evolution (LTE) and 5G dynamically, maintaining consistent performance even during large events or traffic surges. Real-time prediction is replacing manual configuration as the standard for responsiveness. The next leap is happening in system integration. Pairing radios with distributed units, validating interoperability and running full compliance tests across bands can take months. Large language model-based tools are beginning to automate these tasks such as generating test cases, verifying configurations and flagging inconsistencies across O-RAN interfaces. While these methods are still being explored, they signal what the next phase of automation will look like: networks that not only optimise themselves but also build and validate themselves faster.

Expanding role of sensing

AI is transforming how networks perceive the world around them. Integrated communication and sensing technology now allows radios to act as sensors, detecting motion, vehicles and even drones within their coverage areas. This sensory cap­ability extends the value of telecom infrastructure beyond communication, turning networks into sources of environmental and operational intelligence. Early trials suggest how this could unlock new avenues of data monetisation. Sensing-enabled networks can generate real-time insights such as traffic flow analytics, drone movement detection and industrial safety alerts. These applications are still in early stages, but they reveal a shift in purpose: networks are no longer just carriers of data; they are becoming systems that sense and respond to their surroundings.

As sensing expands what networks can perceive, intelligence is also reshaping how they operate. AI-driven safety systems are being used to monitor field operations, detecting whether engineers are wearing helmets or harnesses before climbing towers. By automating compliance checks, these systems prevent accidents and bring accountability into network operations. The same intelligence that enhances safety is also driving sustainability. Energy efficiency is now being built directly into network architecture. Predictive models can shut down radios or air-conditioning units during idle hours, cutting both operational costs and carbon emissions. What started as simple rule-based energy saving has matured into predictive, self-learning control.

Sustainability discussions now extend well beyond energy use. Operators are looking at the entire life cycle, from the embodied carbon in compute hardware to e-waste management and cooling demands at edge sites. Some have started developing energy-rating systems for network hardware, similar to consumer appliance labels, as part of their long-term 6G road maps. The goal is to make efficiency measurable and comparable across the ecosystem. These initiatives mark a broader change in industry mindset. Intelligence is being built in as a guiding principle. Sustainable design, efficient compute and ethical AI are becoming inseparable components of performance and profitability.

The road ahead

The future of intelligent RAN depends as much on people and processes as on technology. As networks shift from reactive to predictive operation, teams must adapt to working alongside automation. This means trusting AI systems to make decisions while maintaining oversight. Model governance, validation and continuous learning cycles must mature into full machine learning operations to ensure the safe transition to AI systems. Security, privacy and data ownership remain central. As network data becomes the foundation for AI training, protecting it will be critical to maintaining trust. Equally important is interoperability. Open standards and common certification processes will decide how quickly innovation moves from concept to nationwide roll-out. Telecom’s next revolution will be defined by a new operating model that is open, predictive, efficient and sustainable. For this, standardisation provides the foundation, while AI provides the intelligence and sustainability gives it direction. The networks of the future will not just connect people; they will sense, predict and continuously improve themselves in the process.

Based on discussions from the session “5G Evolution Through Intelligent RAN” at the India Mobile Congress 2025