Speaking during the RCRTech Telco AI Forum, Gabriel Foglander, head of strategic RAN leadership at Ericsson, described AI-RAN as a bridge toward a broader shift to AI embedded across network domains.
In sum – what to know:
Telco-grade AI – Ericsson says RAN models must be specialized, deterministic, energy-efficient, and compact enough to run on existing hardware.
Data drives scaling – Different RAN functions require different training strategies, from cell-specific learning to globally trained models designed for broad generalization.
Experience matters – Ericsson urges operators to deploy available AI-RAN models now, arguing that practical experience will be essential to developing more AI-native networks.
AI-RAN is moving from a technology concept toward practical deployment, with Ericsson arguing that operators need specialized, deterministic and energy-efficient AI models that can deliver measurable network gains at scale.
Speaking during the RCRTech Telco AI Forum, Gabriel Foglander, head of strategic RAN leadership at Ericsson, described AI-RAN as a bridge toward a broader shift to AI embedded across network domains.
“When we look at AI-RAN, we think this is really the bridge into something that is a paradigm shift for AI embedded into the networks, and we think this is going to influence multiple domains,” he said.
Foglander outlined three areas of AI-RAN development. The first is embedding AI into Ericsson products. The second involves using data produced by the RAN for agent AI, augmenting existing operations and practices to accelerate AI operations and network optimization. The third involves AI products that can unlock capacity, spectral efficiency, throughput, and user experience, while enabling AI-generated traffic.
AI-generated traffic will include physical AI, multimodal applications, human-to-machine, and machine-to-machine interactions. A common characteristic is that AI relies more heavily on uplink connectivity than current mobile broadband applications. AI traffic also requires more consistent latency to support interaction between equipment and AI models, the executive said.
Ericsson is applying AI across several RAN functions. In multilayer coordination, AI-based coverage prediction can enable smoother transitions between spectrum layers toward users without overloading control signaling.
The vendor is also developing machine-readable interfaces toward data and augmented observability, allowing data to be used by models and consumed northbound. Foglander described a scaled approach to structuring data and continuously evolving models.
The inference itself is designed to be highly energy efficient. In real networks, Foglander said there is very little additional, if any, energy consumption in existing equipment. Deploying AI on existing hardware also provides greater scalability than requiring new hardware.
Foglander also said AI needs to deliver outcome-driven gains on top of algorithms that have been tuned over 20–30 years. RAN models also operate within strict time budgets, with some responses required beneath millisecond windows and needing to be deterministic.
At network scale, AI could involve trillions of energy-efficient inference operations per day. Inference therefore needs to scale without creating a prohibitive impact on return on investment or unnecessary energy consumption.
Models also need to fit existing hardware so operators can leverage previous network investments. Compact models are consequently important, according to the executive.
“When we say that all these characteristics are combined, that’s what we consider to be a telco-grade AI model that can be applied in the raw,” he said.
Ericsson currently has around 15 customers involved in live AI-RAN deployments or trials. The company is using specialized ultra-low-latency models through portable software for Ericsson basebands, while neural processor chips are included in its latest radios.
The executive also said that Ericsson does not use one training strategy for every RAN function. Some models can train on data from each deployed cell, or “train on the job,” for context-aware problems such as coverage prediction.
Operators with standalone 5G networks already have technology that can tailor connectivity for higher upload speeds and deterministic latency. They can deploy a slice or similar configuration optimized for AI traffic today, providing insight into potential AI-traffic monetization.
Foglander stressed that operators need practical experience with embedded AI rather than waiting for the technology to mature fully. “There is really no substitute for practical experience,” he said.
The executive encouraged operators to deploy models already available, arguing that doing so will help determine how networks should be operated as AI takes on greater control.
