5G networks and AI (Artificial Intelligence) are mutually enabling. On one hand, AI/ML (Machine Learning) techniques are being embedded into the 5G network operations to optimize performance, known as “AI for Network”. On the other hand, 5G's ultra-fast and low-latency capabilities are enabling new AI applications (Network for AI) by providing services such as network slices, edge computing, and high-bandwidth for distributed AI/ML workloads. This blog will examine both aspects in depth and conclude with a summary of each.
AI for Network refers to using Artificial Intelligence and Machine Learning to make the network itself smarter. In the 5G RAN (Radio Access Network), Core, and MANO (Management and Orchestration) domains, ML models ingest operational data to predict conditions or recommend actions. Key use cases include: traffic prediction, RAN resource optimization, fault detection, and energy saving capabilities.
Figure 1. AI for Network
AI/ML functions require a large amount of data, with the 3GPP specifying data collection and controllers to enable this, highlighting that efficient data management is key to guaranteeing good performance of AI/ML. In practice, raw or pre-processed data will be collated at network elements or analytics functions, such as the NWDAF (Network and Data Analytics Function), and models will be trained. Once trained, inference and decision-making can be deployed across the aforementioned domains.
In terms of benefits, operators will see higher automation and faster optimization with vendors, just as Ericsson reports that AI is fundamental in automating RAN operations and reducing the time-to-market for new features. As an example, Ericsson’s AI-based RAN redesign notably improved live network performance, with carrier aggregation connections growing by 30%, secondary cell data throughput by 22%, and downlink cell throughput overall by 4.3%, all achieved in minutes instead of months. Another example within the RAN that Ericsson offers is the implementation of AI-guided cell shutoff/wake-up policies, which have been estimated to cut RAN power consumption by approximately 12% annually.
However, challenges still remain. ML models require quality data and sufficient training, meaning that terabytes of data will be required in order to optimally train and utilize AI/ML models. Limitations of data may exist as operators may be cautious to share data due to privacy concerns and vendor lock-in. Another limitation is the concept of dynamics; models trained on past data may not generalize to sudden changes or new services being implemented. Finally, computational overhead and lifecycle management (training, version control, retraining) must be addressed.
The other side of the story is Network for AI. This is about how 5G networks are able to enable AI/ML applications. As previously discussed, 5G can support ultra-low latency and high bandwidth, making it well-suited to support demanding AI workloads.
Traditionally, mobile networks have been optimized for downlink traffic as traditional user traffic often consists of streaming videos, downloading files, and browsing websites. With a shift towards AI being available on devices, it is expected that there will be an increase in uplink data due to processing occurring on the device, and then there will be a requirement for cloud synchronization or uploading. Error! Reference source not found. represents the different types of AI traffic a 5G network may encounter and the strain they can place on the uplink.
Figure 2. AI Impact
Therefore, 5G networks need to anticipate this traffic through a Network for AI approach, which refers to the design and optimization of the 5G network to efficiently support AI workloads and the AI-based services running over the 5G network. Key objectives include enabling low-latency communications to support real-time AI, providing high uplink throughput for transmitting large volumes of sensor and video data to edge or cloud processing platforms, and integrating edge computing resources to reduce communication delays. Collectively, these capabilities allow the 5G system to act as an enabling infrastructure for advanced services such as the ones highlighted in Error! Reference source not found..
Figure 3. Network forAI
In summary, AI for Networks is about embedding AI/ML inside the 5G network to run more efficiently, while Network for AI is about evolving 5G connectivity to enable external AI applications. Industry demos and studies show large potential gains (throughput, energy, flexibility) but also underscore integration and data challenges.
If you’re interested in exploring what AI for Network features are supported in Release 19, our new course “Release 19 5G AI/ML Enhancements” is now available.