Nokia: Report Points to AI-Driven Shift in Mobile Traffic

AI is now embedded across mobile devices, with applications ranging from scene recognition to voice interaction and content generation.

These tools rely on large language models and are already shaping how traffic moves across mobile networks.

The effect is visible in changing traffic patterns, particularly as uplink demand grows and latency becomes more critical in telecoms.

A Nokia analysis of more than 50 AI applications points to three developments: higher uplink traffic, overall data growth and increasing sensitivity to delay in conversational services such as chat and voice.

These trends raise a deeper question for operators: whether the radio access network (RAN) needs structural change to handle what comes next.

RAN design under pressure from AI traffic

Mobile networks traditionally support a wide mix of traffic types, including high-volume video and low-bandwidth messaging.

Telcos typically respond to rising demand by adding capacity while relying on best-effort delivery, which has proved resilient, yet history shows that capacity alone does not always solve new challenges.

The shift from circuit-switched voice to packet-based internet protocol traffic forces a redesign, as variable packet sizes replace predictable voice patterns.

The rise of the IoT also introduces small data packets and deep indoor coverage needs, prompting the development of LTE-M and NB-IoT, both narrowband technologies designed for efficiency and reach.

By contrast, video streaming and web-based services scale without major architectural change, as additional capacity meets demand. Current AI applications largely follow this model, meaning existing RAN designs are sufficient for now.

Physical AI introduces new network demands

The outlook changes with the emergence of Physical AI, embedded in machines such as robots and autonomous vehicles, enabling them to perceive and interact with the physical world in real time.

These systems often depend on continuous video input and rapid decision-making, which places new demands on the network.

Unlike traditional video streaming, Physical AI cannot rely on buffering. In Physical AI, video frames must arrive within strict time limits to remain useful. This makes latency a critical factor.

Maintaining consistent low latency using best-effort delivery requires substantial unused capacity.

“Physical AI introduces the possibility that large-volume uplink video with strict latency requirements will become a meaningful part of mobile traffic, creating both a design challenge and a monetization opportunity,” writes Harish Viswanathan,  Head of the Radio Systems Research Group at Nokia.

Delivering uplink video with around 20 milliseconds latency may need three to four times the average data rate provisioned. This is manageable for low-bandwidth services such as voice but becomes costly when applied to high-volume video streams.

As more devices connect, the required reserved capacity increases sharply. This limits scalability and raises costs, making Physical AI a distinct category of traffic that cannot rely on existing approaches.

Telco strategies for scaling low latency services

To address these pressures, telcos are likely to focus on changes across networks and service models. At the application level, not all data needs equal priority.

When video is processed by AI rather than viewed by humans, only essential information may need immediate delivery. This process, known as semantic communication, reduces the amount of data requiring strict latency guarantees.

Within the network, existing tools such as quality of service and network slicing are significant. Quality of service allows operators to prioritise certain types of traffic while network slicing enables separate virtual networks with defined performance characteristics.

Service models also evolve. Supporting low latency, high-volume traffic increases network costs, which means operators must move beyond best-effort pricing. 

Physical AI therefore pushes telecoms providers to rethink long-standing assumptions about traffic handling. While most AI applications fit within existing frameworks, the combination of high data rates and strict latency requirements introduces a new category of demand.

For the RAN, this means moving towards more programmable and flexible architectures. Operators need systems that can adapt to varying traffic types, prioritise critical data and maintain performance at scale.

As Physical AI use cases expand, from autonomous transport to industrial automation, the ability to manage low latency uplink traffic is key for network strategy.