Ericsson is framing AI RAN as a two-sided transformation: networks must use AI to improve their own performance while also evolving to support a new generation of AI applications
In sum – what to know
AI reshapes networks – Ericsson’s Anders Söderlund says AI RAN involves both using AI to optimize networks and evolving networks to support AI applications.
Uplink becomes critical – Physical AI, smart glasses, and humanoids will require stronger and more consistent uplink performance, alongside low latency and reliable connectivity.
AI moves across the network – Ericsson sees AI models running across devices, cell sites, the edge, and centralized systems, with model size and placement determined by latency, power and performance requirements.
Ericsson is framing AI RAN as a two-sided transformation: networks must use AI to improve their own performance while also evolving to support a new generation of AI applications.
Speaking at RCR Wireless News’ virtual Intelligent RAN Forum 2026, Anders Söderlund, strategic product manager for Ericsson RAN Compute, described the two elements as “AI for networks” and “networks for AI.”
The first concept involves using AI models to enhance network performance and optimize operations. The second focuses on adapting networks to meet the requirements of applications such as physical AI, smart glasses, and humanoids.
“Applications such as physical AI or smart glasses will demand new capabilities from the network, such as a stronger uplink, for example,” Söderlund said.
AI for networks
On the network optimization side, Söderlund highlighted the need for energy-efficient AI inference across communications networks. With tens or hundreds of thousands of cell sites, he said, operators need to be highly efficient because there is limited room for additional power and space at individual sites.
“The way we do it at Ericsson is that we do hardware and software co-design to build really energy efficient systems that also support AI inference then,” he said.
Ericsson’s RAN compute platforms and silicon are designed to support both RAN workloads and AI inference. Söderlund said the vendor has supported AI-enhanced, rule-based algorithms since the 4G era and introduced RAN compute platforms capable of supporting such workloads more than a decade ago.
The latest platforms can run AI-native features in 5G. Söderlund cited examples including up to 25% higher spectral efficiency, five-times more accurate positioning, improved coverage prediction, and up to two times more high-traffic users.
Ericsson is also adding neural network accelerators to its massive MIMO radios and plans to incorporate them into future RAN compute platforms and Ericsson silicon. Söderlund said these developments will provide a foundation for 6G and AI-native RAN.
Networks for AI
The second part of the strategy focuses on what AI applications will require from networks. Söderlund said applications such as smart glasses and humanoids will require stronger uplinks. While 5G mid-band TDD provides substantial downlink capacity, he said TDD is less capable in the uplink, increasing the importance of FDD bands.
Ericsson has measured uplink performance across more than 100 networks globally. According to Söderlund, the measurements indicate that AI applications will require at least 5 Mbps of uplink capacity. “For that reason, we see that going forward, the uplink needs to be strengthened,” he said.
He also highlighted the importance of consistent low latency. AI applications could require latency-prioritized scheduling so that, for example, smart glasses users do not experience excessive delays.
Putting AI in the right place
Söderlund also described a distributed AI architecture spanning devices, cell sites, edge locations, and centralized network systems. The appropriate location for an AI model depends on factors including power consumption, model size, and latency requirements. Models running on smartphones need to remain small to preserve battery life, while AI inference at cell sites must be highly efficient and operate within microsecond-level latency requirements.
At centralized data centers, there is more time available for inference and larger models can be used. Söderlund noted that for some large language model applications, a response arriving within one, two or three seconds can typically be acceptable.
The result, he said, is an architecture in which AI is distributed throughout the network, with the model selected according to where it can meet the relevant requirements.
He concluded that stronger uplink capabilities will be necessary to support emerging AI applications and said the time has come for communications service providers to begin moving into the area and monetize AI.
Table of Contents
