Ericsson is pushing AI-RAN from lab promise into operational reality. The company says operators need models that prove gains at network scale. For carriers, that means better coverage, capacity, energy discipline, and service quality.
Gabriel Foglander, who leads strategic RAN leadership at Ericsson, outlined the company’s strategy, positioning AI-RAN as a step toward embedding AI across every domain of the network.
“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.
The message is clear. Telco AI cannot behave like a general cloud chatbot. In radio networks, decisions must arrive fast and predictably. Some actions must happen in less than one millisecond. That leaves little margin for unstable or oversized models.
Ericsson wants telco-grade AI models to be specialized, compact, and energy efficient. They must run on existing hardware wherever possible. That matters because radio sites already carry major cost pressure. New hardware everywhere would slow adoption and weaken investment returns.
At the same time, the opportunity is meaningful. Ericsson sees AI improving multilayer coordination across spectrum bands. In simple terms, networks can guide users between coverage layers. They can do this without flooding the system with signaling. That could improve throughput and reduce service interruptions.
The company also expects new traffic patterns. AI applications will use more uplink capacity than today’s mobile broadband. Robots, sensors, cameras, and human-machine tools will send richer data upstream. These services may also need steadier latency. Latency means the delay between request and response.
This shift aligns with standalone 5G networks and slicing. Slicing lets operators reserve virtual network lanes for specific needs. A carrier could tune one slice for AI uploads and stable response times. That creates a possible path to premium enterprise services. Still, demand must justify the added operational focus.
Data remains a central challenge. Foglander said Ericsson will not use one training method for every RAN function. Some models can learn from each deployed cell. That approach suits local coverage prediction. Other models need broader training to work across many network conditions.
Energy use also sits under scrutiny. AI inference means running a trained model to make decisions. At national scale, networks could run trillions of these operations daily. Ericsson argues existing equipment can handle many tasks with minimal extra power. If accurate, that would ease sustainability concerns.
Yet operators should avoid viewing AI-RAN as instant transformation. Radio algorithms already reflect decades of tuning and field experience. AI must improve those systems, not merely decorate them. Measurable gains will matter more than marketing language.
Ericsson says about 15 customers are testing or deploying live AI-RAN systems. The vendor uses ultra-low-latency software models for its basebands. Its newer radios also include neural processor chips.
“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.
Foglander urged operators to start learning now. “There is really no substitute for practical experience,” he said. That view will resonate with engineers facing real network limits. AI-RAN may advance quickly, but field knowledge will decide its value.
