Skip navigation
Previous searches
    Suggested searches
      ""

      AI RAN

      Networks tuned by AI, and built for the performance AI applications need

      AI RAN explained

      AI RAN covers every point where artificial intelligence (AI) acts on the radio access network (RAN). At one end, the baseband software at each site decides how every transmission is sent, within microseconds. At the other, automation applications known as rApps tune thousands of cells from a central management layer, over seconds to weeks. Each model runs where its timescale and data allow: compact models on the radio hardware for real-time decisions, larger centralized models for network-wide prediction and configuration. 

      This is the foundation for autonomous, intent-driven operations, where the network thinks, learns and adapts in real time toward an outcome the communication service provider (CSP) has defined. 

      ""

      Why AI RAN matters now

      Mobile traffic is turning uplink-heavy

      AI applications send far more data up to the network than today’s apps and need predictable latency. A pair of smart glasses streaming video to an AI assistant needs up to 1.4 Mbps of uplink per session, and an autonomous robot up to 150 Mbps, which changes the balance between uplink and downlink traffic. 

      Traditional rule-based algorithms have peaked 

      RAN software is close to its theoretical limits, and models that learn from live network data now outperform fixed rules. AI-native link adaptation, which sets each transmission’s data rate, delivers close to 10 percent more traffic from the same spectrum, improving spectral efficiency. 

      AI RAN has to decide in microseconds, trillions of times a day

      The scheduler in one 5G cell runs more than 7 million times an hour, and a CSP’s network handles trillions of transmissions a day, each a potential inference (a decision by a trained model on live data). The lower RAN layers need each answer within tens of microseconds, which only real-time models at the radio site can meet.

      It adds capacity without new spectrum or hardware 

      AI RAN adds capacity from spectrum already licensed and sites already built. The cost side is inference: models run at thousands of sites, many times a millisecond, so their compute and energy footprint, not the one-off training, decides the business case. 

      ""
      Already familiar with AI RAN and exploring how to deploy it?
      Go straight to deployment considerations and Ericsson's AI RAN offering.
      Jump to deployment

      AI RAN architecture

      AI RAN takes three modes, defined by where the model runs and what it acts on: AI for RAN applies models to the network from the management layer, AI in RAN embeds them in the radio software, and AI on RAN applies the network to AI workloads from devices. “AI and RAN” is a fourth concept from the AI-RAN Alliance in which AI and RAN workloads share the same compute, with an orchestrator allocating resources between them. 

      ""

      AI for RAN

      Models that run outside the RAN software in the management layer, and feed predictions and control inputs back to the network. They work over seconds to weeks on tasks such as configuration, optimization and fault detection: an rApp, for example, can predict traffic and move users between frequency layers before cells congest. This mode is live in commercial networks today.

      AI in RAN 

      Models embedded in the RAN software, in the layers that handle each transmission, deciding within microseconds to milliseconds. Link adaptation, scheduling and beamforming are the first functions to move from fixed rules to trained models, and a model can also predict whether a device is within coverage of another frequency layer, so handovers happen faster with fewer drops. This mode is in commercial software and live-network trials, including coverage prediction with Optus and link adaptation on SoftBank’s commercial 5G network.

      AI on RAN

      The network carrying AI workloads for devices, with the uplink capacity, guaranteed quality of service and low latency they need. Compute at the network edge lets a device offload processing it cannot do on board: a robot handles light decisions itself and hands heavier ones to nearby edge compute, with the network adapting in real time to the load. This mode is in proof of concept, demonstrated with SoftBank for network-enabled physical AI. 

      AI RAN versus traditional RAN

      Unlike traditional RAN software, which runs on rules set in advance and changed by hand, AI entered the RAN by augmenting those rules with machine learning, then replacing single functions with trained models, and now designing functions around AI from the start. The difference is in the software, not the radios: the same sites and spectrum perform differently once trained models replace fixed rules. 

        Traditional RAN AI RAN
      How features work Rules set from offline analysis and static parameters Trained models decide on live data in real time
      How the network is tuned Manual configuration, cell by cell Models trained on network data and retrained locally; the CSP sets the intent
      How it adapts to change Rules stay fixed as signal quality and interference change Decisions track changing conditions, and models keep learning in operation
      Where the intelligence sits Added on top of the network software Built into the radio software and the management layer

      Centralized and distributed AI

      Where a model runs is set by how fast it has to act. A typical CSP has 5 to 10 central sites and thousands of radio sites, and the control loop spans microseconds at the radio to weeks in the management layer. Tasks that work over seconds or longer, such as network-wide optimization and healing, belong centrally. Tasks that must complete in microseconds, such as scheduling and link adaptation, can only run at the site. 

      The two layers work together. Distributed models act on each transmission and report what they see; centralized applications, the rApps, use that data to configure and assure thousands of cells at once. Placing compute too close to the antenna is costly, so the split is as much an economic decision as a technical one. 

      ""

      Central sites

      Five to ten data center locations per CSP host the management layer and its rApps. Here, AI handles prediction, configuration, optimization and fault detection across the whole network, in non-real time.

      Edge data centers

      Regional and local sites sit between the two. They offer lower latency than a central site, though not the microsecond level the radio needs, and can process AI on RAN workloads from devices. Manufacturing, healthcare and gaming need between 50 milliseconds and one second.

      Radio sites

      Thousands of sites run the compressed models that act on every transmission, in the baseband and the radio units themselves. The benefit of each model has to outweigh the cost and complexity of running it at that scale. 

      Hardware and software

      AI in the RAN runs on two kinds of hardware: silicon designed for radio processing, and general-purpose servers. The industry has taken two routes to adding AI: extra processing boards dedicated to AI, which add footprint at every site, or evolving the existing hardware to run AI functions in place. Either way, models and hardware have to be designed together, because a model tuned to its processor runs faster and uses less energy. 

      Network build icon
      Purpose-built silicon

      Application-specific chips built for radio processing carry hundreds of signal-processing cores, and the matrix math of beamforming is the same math a neural network runs on. With on-chip memory and instructions for inference, the same board can run RAN and AI workloads concurrently at microsecond latency, without extra rack space.

      Cloud Icon
      Cloud and open hardware

      Cloud RAN runs the same RAN functions on standard x86 servers from the wider ecosystem, with vector and matrix extensions and selected accelerators handling the heaviest functions. Graphics processors remain the workhorse for training models; inference needs lower latency and tighter cost control, which these extensions can supply. Software portability lets a CSP choose the platform per site without changing the functions it runs. 

      Benefits and value

      Applying AI to the network raises what existing sites, spectrum and energy budgets can deliver; building the network for AI turns that performance into services AI applications can run on. 

      AI Icon
      More capacity from the same spectrum

      Models predict channel conditions, traffic and demand per user and set each transmission’s rate, beam and timing, where fixed rules settle for a safe average. The result is higher throughput, more users served at high load and fewer dropped connections, without new spectrum or hardware. 

      Power Icon
      Lower energy per site

      Models forecast traffic and cell usage and switch radio functions such as multiple-antenna transmission on and off to match, instead of leaving them on for peak load. Energy use falls while performance holds, and each model’s own consumption is weighed against its savings. 

      Smart-automated-operations icon
      Operations that run on intent

      Centralized applications learn the network’s normal behavior, flag anomalies such as sleeping cells and act before performance degrades. The CSP defines the outcome, and the network configures, optimizes and heals itself toward it, cutting manual planning and tuning across thousands of sites. 

      Network icon
      A network ready for AI services

      Guaranteed levels of throughput and latency, more uplink capacity and more accurate device positioning let a CSP offer connectivity with predictable performance to consumers and enterprises. That is the basis for new services, from assistants on smart glasses to robots that offload processing to the network. 

      Use cases and applications

      Each mode of AI RAN produces its own kind of use case: AI for RAN changes how the network is run, AI in RAN changes how each radio performs, and AI on RAN changes what the network can carry.  

      Running the network with AI for RAN

      AI for RAN
      ""

      Intent-based operations with AI agents

      The CSP states an outcome, such as a latency target for a service area, and AI agents in the management layer plan and carry out the configuration changes to reach it, with a human approving business-critical actions.

      ""

      Network-wide energy optimization

      Centralized features measure, predict and optimize energy consumption across the network, putting radio capacity to sleep where traffic allows and restoring it before demand returns, without lowering performance.

      ""

      Traffic prediction and optimization

      An rApp forecasts traffic per cell and moves users across frequency layers ahead of congestion, raising cell throughput in both directions from the management layer, with no change to the radio software.

      Read more

      New services on the network with AI on RAN

      AI on RAN
      ""

      Smart glasses and AR/VR

      Glasses stream video, audio and spatial data to an assistant in the cloud for translation, navigation and overlays, which needs steady uplink and a round trip of about 200 milliseconds.

      ""

      Personalized AI agents

      Always-on assistants on wearables stream contextual data to the cloud and respond in real time, creating a persistent uplink load the network has to carry at predictable quality.

      ""

      Autonomous vehicles, drones and robots

      Machines stream telemetry and video, receive control data, and offload heavier decisions to edge compute, within a round trip of about 100 milliseconds and up to 150 Mbps of uplink.

      Read more

      Real-time radio performance with AI in RAN

      AI in RAN
      ""

      AI-native link adaptation

      A model trained on live network data sets the data rate of every transmission as signal quality and interference change, lifting throughput most where conditions are poor, such as at cell edges.

      ""

      Mobility optimization

      Models predict the best handover target and whether a device is within another frequency layer’s coverage, so handovers complete faster, fewer connections drop and device batteries last longer.

      ""

      AI-powered energy saving in radios

      Models in the radio software put multiple-antenna functions into sleep modes between bursts of traffic and bring them back as demand rises, with no loss of performance.

      Read more

      • Press release: Bell Canada and Ericsson field test AI-native link adaptation
      • Press release: T-Mobile and Ericsson take AI-native scheduling into commercial trials
      • Press release: Optus and Ericsson trial AI to boost 5G downlink
      • News: AT&T and Ericsson run AI-native software on Cloud RAN with Intel Xeon 6 

      Enabling technologies

      No single technology makes AI RAN work. It is the result of models, network domains and services working together across the whole network, from the radio and its compute to the management layer and the systems that sell the service. 

      Telecom AI

      Telecom AI applies the right model in the right place across the whole network, from business support systems to the core and the radios. AI RAN is its radio domain, where the same data strategy, model types and agentic operations run at the tightest timescales in the network.

      Learn more about telecom AI

      Autonomous networks

      Autonomous networks replace rule-based automation with closed loops that define an intent, observe the network, decide and act, with rising levels of autonomy. In the RAN domain, intent-aware functions adjust in real time while rApps and AI agents handle the slower, network-wide loop.

      Learn more about autonomous networks 

      5G Advanced

      5G Advanced is the stage of 5G in which networks become programmable and service-aware, adjusting their behavior to a requested outcome. It is also where the first AI-native radio functions, such as link adaptation, arrive as standard RAN software features.

      Learn more about 5G Advanced

      6G

      6G is being designed with AI as a native part of the architecture, with 3GPP work on a common AI and machine learning framework for the radio interface already under way. AI RAN in 5G is where the models, data pipelines and compute that 6G will depend on are being proven in live networks.

      Learn more about 6G 

      Differentiated connectivity

      Differentiated connectivity lets a CSP offer several levels of connectivity, each with deterministic performance, on one network. It is how networks for AI become a service: slicing and priority handling give an AI application the throughput, latency and reliability it needs, and adapt as its workload changes.

      Learn more about differentiated connectivity

      Beyond the buzzwords

      AI RAN within the wider network evolution

      Placeholder for network evolution content, pending campaign alignment. Two paragraphs placing AI RAN inside the high-performing programmable network story.

      Explore network evolution
      ""

      Deploying AI in RAN

      Deploying AI in the RAN comes down to four decisions: where each model runs, which AI technique fits each function, what hardware carries the real-time workloads, and how the radio software and the central applications divide the work. Taken separately they lead to over-dimensioned sites or models that miss their latency targets; taken together they set where the return on each model is highest. 

      The path to AI-native RAN

      AI has entered RAN software in three stages, and each raises what a feature can do and what the hardware and data behind it must supply. A network can run features from all three at once; the aim should be to evolve the functions that gain most from real-time decisions to become AI native. 

      ""
      1
      Rule-based with machine learning

      The feature keeps its rules, and machine learning tunes the parameters from network data instead of an engineer setting them from offline analysis. This is where AI in the RAN began, in 4G, with mobility optimization, sleep modes for multiple-antenna radios and detection of sleeping cells.

      2
      AI-powered

      A trained model takes over part of the function, typically prediction, and the surrounding logic acts on its output. Performance improves because the decision now follows live conditions, but the function is still designed around the original rules.

      3
      AI-native

      The function is designed around AI from the start, and a model replaces the algorithm rather than assisting it. This stage needs real-time data, more storage and processing at the site, and deep telecom expertise to keep complexity and cost in check. Our first AI-native radio functions, link adaptation among them, run in commercial networks today.

      Choosing the right AI technology

      No single AI technique suits every radio function. We match the technique to the function, and the data available decides which techniques are possible: a model is only as good as the quality and timeliness of the data it learns from, and real-time radio functions need data at the same resolution. 

      Generative AI

      Large language models support network operations, software development and the people who build applications on the network. They work with operators in a conversational loop, with a person in charge of business-critical actions.

      Agentic AI

      Agents sense, plan and act toward an intent, alone or in coordination, without step-by-step instructions. In the RAN they make the autonomous decisions that intent-driven operation depends on.

      Digital twins

      A digital twin mirrors the live network so that models can be trained and tested against real conditions without touching it. We use twins to introduce new functions without degrading performance indicators.

      Reinforcement learning

      The model learns by trying actions in the control loop and keeping those that improve the outcome, which suits radio resource allocation where conditions change constantly. In first deployments, our reinforcement learning models delivered up to 20 percent higher downlink throughput and 10 percent higher spectral efficiency.

      Neural networks

      Neural networks carry the real-time inference inside the radio software, where their matrix operations run on the same signal-processing cores that handle beamforming. Compressed networks meet microsecond deadlines that larger models cannot, even on powerful general-purpose processors.

      Hardware for AI RAN

      Models need storage and processing at thousands of sites, so the hardware decision sets both the cost and ceiling of an AI RAN deployment.  

      Read the report: How to deploy AI in mobile networks 

      Standards and ecosystem collaboration

      AI RAN can scale only when models, data and compute interoperate across vendors. No supplier can set that on its own, which is why we work with industry peers and partners so that CSPs can deploy AI in the RAN without being tied to one vendor’s hardware, models or data formats. 

      ""

      Standards

      We are a leading contributor to the 3GPP work that decides which radio functions AI may take over and how models are trained, delivered and monitored across vendors. A Release 18 framework and Release 19 use cases already follow this path, and we are committed to the same work for an AI-native 6G, together with the O-RAN Alliance, ETSI and TM Forum on the automation layers around it.

      Industry alliances

      Our work with industry alliances includes the AI-RAN Alliance, where we are a founding member and sit in the working groups on how AI and RAN share compute, and the AI-RAN Innovation Center in Bellevue, where we run RAN software on accelerated computing in over-the-air trials with CSPs. We also keep our automation platform open, so a CSP can run its own rApps, or another vendor’s, on the same management layer as ours.

      Featured insights

      ""

      Beyond the buzz, the value of AI RAN

      Melike Erol-Kantarci, Ericsson, and Leonard Lee, neXt Curve, on what AI in the RAN delivers in practice.

      Watch the video
      ""

      AI in RAN, the inference tsunami is here

      Why RAN inference runs to trillions of decisions a day, and why radio silicon is built for it.

      Read the blog post
      ""

      AI in RAN

       How AI in the radio network improves throughput, handover, energy use and anomaly detection.

      Read the report

      Ericsson AI in RAN solutions

      Ericsson AI in RAN is a software subscription that brings telco-grade AI models into basebands and radios, starting with scheduling, beamforming, multi-layer coordination and positioning. It runs on Ericsson RAN Compute and Ericsson Silicon radios, and on partner platforms through Cloud RAN, with no additional hardware. AI-ready radios with neural network accelerators and rApps on the Ericsson Intelligent Automation Platform complete the offering. 

      Explore the AI RAN offering
      ""

      Related offerings

      ""

      AI-ready Radios

      Massive MIMO and remote radios with neural network accelerators for on-site inference.

      monopoly

      Ericsson rApps

      Automation applications for the network-wide loop, on an open management layer.

      ""

      Ericsson Intelligent Automation Platform

      The open, multivendor management and orchestration platform that hosts rApps