At present, as mobile communications evolve beyond 5G toward 6G, AI-RAN has emerged as one of the most prominent and critical keywords. Its significance lies not merely in the idea of “adding AI to RAN,” but in the much broader ambition of redesigning the radio access network itself around AI.
In this article, we organize the concept of AI-RAN and, through analysis of real-world demos and showcases, explore what AI-RAN is and what it aims to achieve.
<AI-RAN: The Evolution of Mobile Networks – The Infiltration of AI into Base Stations / AI Taking Over Mobile Networks>
The radio access network (RAN) of mobile operators has evolved from D-RAN → C-RAN → vRAN → O-RAN, and is now progressing toward AI-RAN, where AI is applied directly to the RAN.
The AI-RAN Alliance, formed in 2024 under the leadership of SoftBank (Masayoshi Son) of Japan and NVIDIA (Jensen Huang) of the United States, defines the concept and use cases of AI-RAN. At events such as MWC 2024 and MWC 2025, the alliance and its ecosystem partners have demonstrated AI-RAN solutions, continuously proving the potential of AI-RAN through live demos and showcases.
The AI-RAN Alliance was founded by 11 global telecommunications and IT companies, including NVIDIA, Arm, SoftBank, T-Mobile, Nokia, Ericsson, Samsung Electronics, Microsoft, and DeepSig. As of January 2026, the Alliance has grown to 109 member companies, including all three major mobile operators (SK Telecom, KT, LGU+) in Korea.
The core objective of the AI-RAN Alliance is to maximize wireless communication performance by integrating AI into the RAN (base stations) (AI for RAN), and to create new business value based on AI applications (AI and RAN, AI on RAN).
The Alliance defines three key pillars of AI-RAN—AI for RAN, AI on RAN, and AI and RAN. Let us take a closer look at each of these concepts.
|
|
AI-RAN |
||||||||||||||
|
AI for RAN Spectral Efficiency |
AI and RAN Asset Utilization |
AI on RAN New Applications |
|||||||||||||
|
Concept
What is This?
|
Objective: To improve and optimize RAN performance by leveraging AI/ML (machine learning). Instead of traditional rule-based algorithms used in conventional RANs, AI/ML models based on data-driven learning and inference are embedded into base stations to enhance their performance.
Key examples: AI/ML-based channel estimation, interference management, MAC scheduling, beamforming optimization, and power control.
Impact: Improved communication quality for UEs and reduced network investment (CAPEX) and operational costs (OPEX) for mobile operators.
|
Objective: To increase resource utilization by allowing RAN workloads and AI workloads to share a single computing resource (RAN servers), optimizing the usage of GPU and memory resources.
Key examples: During periods of low traffic— such as late at night or daytime hours in residential (bedtown) areas—RAN server resources are leased to external enterprises for AI computing (AI training or AI inference). This enables the creation of new revenue streams by monetizing idle RAN server resources.
Impact: Resource efficiency is maximized by reallocating GPU resources to AI computing when mobile traffic is low. Flexible distribution of idle resources helps reduce the overall total cost of ownership (TCO).
|
Objective: To create new revenue streams by running AI application services on the RAN and selling them to enterprises or consumers.
Key examples (primarily AI inferencing): Real- time video analytics, autonomous driving, remote surgery assistance, real-time robot control, AR, LLMs, enterprise RAG, and agentic AI.
Impact: Because AI services are provided from RAN servers to devices directly connected to the RAN, ultra-low-latency AI services can be delivered.
|
||||||||||||
|
AI-RAN Alliance and SoftBank's AI-RAN Concept
|
|
|
|
||||||||||||
|
|
|
|
|||||||||||||
|
|
|
|
|||||||||||||
|
NVIDIA’s AI-RAN concept
|
• AI-algorithms for RAN performance improvement • Site specific learning & optimizations for spectral efficiency
|
• Common infrastructure to run AI and RAN workloads • Orchestration & dynamic workload distribution between AI and RAN
|
• Network-differentiated connectivity • Local breakout of AI traffic for enhanced Quality of experience
|
||||||||||||
|
AI RAN Alliance's MWC 2024 Demo
Demos that represent-atively showcases the concepts of AI for RAN, AI and RAN, and AI on RAN. |
Impact: A 25% improvement in uplink throughput compared to conventional RAN by applying AI-based channel interpolation.
Demo participants: SoftBank, NVIDIA, Fujitsu
|
The SMO monitors workloads generated on the vRAN side and communicates with the Cluster Agent. to the cloud, and the cloud sends additional workloads that can utilize those resources (i.e., inference workloads are dispatched to the corresponding cluster). User queries are sent to an AI inference application running on the RAN server, which processes the request and returns a response. Communication between the user and the AI application, as well as between the Cluster Agent and the cloud, takes place over the Internet.
Demo participants:: NVIDIA, Radisys, Aarna Networks, Supermicro
|
A demo in which a video stream is received from a 5G camera over a 5G network and used to detect objects (pedestrians, cars, trucks, and buses) in real time. The AI application that would normally run on a MEC server is instead deployed directly on the RAN server.
Demo participants:: NVIDIA, Radisys, Fujitsu, Supermicro
|
||||||||||||
|
Let’s explore more AI-RAN demo and develop- ment use cases
|
Impact: More than a 30% improvement in smartphone uplink throughput at the cell edge. Expanded base station coverage also reduces network investment costs for mobile operators.
Demo participants: NVIDIA, Samsung, Keysight
Impact: In high-interference, low-SNR conditions, uplink throughput is improved by 20–50% compared to conventional L1 without AI.
Demo participants: SoftBank, NVIDIA, Fujitsu
Neural network–based encoders, receivers, and decoders replace conventional pilots and modulation schemes while preserving a 5G-NR- compatible CP/DFT-OFDM slot structure.
This demo demonstrates that AI/ML-based wireless interface design can improve radio performance, increase spectral efficiency, and seamlessly integrate with existing 5G networks.
These capabilities are essential for future AI-RAN deployments.
Demo participants: DeepSig, NVIDIA
|
AI and RAN Northeastern AutoRAN Demo – MWC 2025 AI-RAN Coexistence
Demo participants: Keysight and Northeastern University Open6G
|
Demo participants: ARM, Tannera, Effinet, Phluido
Ultra-low-latency LLM: SoftBank has developed an AI (an ultra-low-latency LLM) that generates robot motions in real time based on sensor information mounted on the robot, developed using NVIDIA AI Enterprise. Significance: This demonstrates that an LLM running at the mobile network edge, outside the robot itself, can control robot motion in real time.
SoftBank has developed an enterprise-grade RAG application equipped with multiple technologies that improve the usability and accuracy of RAG.
By integrating enterprise data, the system can generate responses based on the latest internal information, enabling generative AI to handle tasks specialized for corporate operations.
All data is processed in a closed environment on AITRAS edge AI servers, allowing the application to operate in a more secure environment than cloud services on the public Internet.
|
||||||||||||
The pilot network validates an intelligent security service that uses AI-powered CCTV to detect intrusions, wrong-way movements, and abnormal behaviors in real time within restricted areas or security blind spots at the airport, and automatically links these detections to an alert system.
At the same time, AI algorithms improve signal quality between 5G cameras and base stations, enabling stable, uninterrupted transmission of high-resolution, high-volume video streams.
Sources (Korean)
|
|||||||||||||||
|
|
Register | About Us | Sponsor Channel Service (Korean Market) | Advertise | Contact Us | LinkedIn | Our Korean Site
|
|||
|
HQ: 201, Hanho Building 33, Nonhyeon-ro 64-gil, Gangnam-gu, Seoul, 06228, Rep. of KOREA Netmanias USA: 5214 36th Ave. NE Seattle, WA 98105 USA Netmanias was founded in year 2002, and provides in-depth analysis and overview of 4G & 5G mobile, IoT, SDN/NFV, Gigabit Internet and video streaming technologies as well as trends in the evolution of Korean operators’ networks and services Copyright ⓒ 2002-2023 NMC Consulting Group. All rights reserved. |
||||
AI-RAN has become one of the most important and well-known keywords as mobile communications move from 5G to 6G. snowrider