Summary Bullets:
• Ericsson introduced a new RAN software package called “AI in RAN,” which includes RAN-based training models and agentic AI that may enable greater automation over time
• The software package includes six new performance-enhancing features leveraging AI, some of which are downloadable this month while others become available later this year.
Ericsson took another step in the commercialization of using AI to enhance the radio access network (RAN) this week by introducing a new RAN software package the vendor calls “AI in RAN.” It includes:
• Telco-grade AI models designed to run in the RAN in real time
• Continuous learning using scalable, high-quality data
• Agentic AI support to enable RAN operations and automation
By training AI models within the RAN, mobile operators can make use of the full depth of data at each RAN site, where a wide variety of local conditions and site-specific factors can affect the user experience and RAN efficiency in ways that may be too complex for the human mind to process and manage in real time. In the long run, this ongoing learning, reinforcement and refinement may enable new levels of network automation, as AI agents take on more work.
The new offering’s specific RAN features include:
• AI-Native Scheduler for Link Adaptation, which boosts throughput amid poor radio conditions
• AI-Powered Macro Positioning, which increases the accuracy of pinpointing users
• AI-Managed Beamforming, which increases throughput and reduces call drops and signaling interference
• AI-Powered Multi-Layer Coordination, which predicts where users are and moves traffic to the best spectrum layer for their service type, increasing spectral efficiency
• Performance Management Event Schema Files, which allows AI agents to read performance management data in place of humans, and
• Augmented Observability, which uses better data (e.g., signal strength) to yield more valuable insights
Some of the above features were previewed in this year’s first quarter, when Ericsson distinguished its AI RAN position from those of other vendors by introducing new Massive MIMO radio chips, commercially available later this year, that include “neural network accelerators” using tensor pools to perform AI inference computing. (A tensor is, as Ericsson puts it, “a generalized matrix representation,” or an array of multi-dimensional data that enables processing tasks analogous to the parallel computing done by graphics processing units, or GPUs. Conceptually, tensors can be thought of as three-dimensional cubes of data, as opposed to two-dimensional matrices.) With neural network accelerators in the radio itself, some AI tasks can take place directly in the radio and avoid the need to consume even fronthaul capacity or baseband compute resources.
The first AI-in-RAN features become available this month, with additions later this year. Ericsson is emphasizing the fact that the new package is a software upgrade that doesn’t require replacing radios; but as the broader vision Ericsson is describing involves new chips in the radio, some hardware evolution will ultimately be part of the process. At the same time, the fact that some software features have already become available helps Ericsson leverage its differentiation among competitors, including Nokia. Nokia’s approach to AI RAN is reliant on GPUs (which Ericsson is open to but doesn’t currently require), and its first GPU-driven base stations won’t be available for trials until late this year. Ericsson’s move this week lets its customers know they don’t need to wait to start upgrading their RANs with AI, and they don’t need to start by deploying new hardware.