Implications of GenAI for uplink, downlink and network planning

Emergence of AI-native multimodal use cases

AI and GenAI are rapidly transforming both consumer and enterprise domains, enabling entirely new classes of AI-native, multimodal applications, which process multiple types of data input simultaneously.

AI is driving hyper-personalized experiences through GenAI-powered smartphones. Users are benefiting from faster access to relevant information or easier content creation. As a result, the technology can boost user engagement and retention. Smart and AR glasses – requiring the offloading of compute-intensive tasks such as large language models (LLMs) or Gaussian splatting1 to the cloud – enable rich AR experiences. Significantly, the emergence of AI agents and multimodal LLMs is ushering in a new era of intelligent assistants that rely on uplink-heavy video and voice processing to deliver real-time, context-aware interactions.

Together, these innovations are causing a foundational shift toward AI-native platforms and ecosystems, where intelligent, multimodal interfaces redefine how humans and machines interact. All of these emerging use cases require high-performance connectivity.

Rise of personalized AI agents

At the intersection of these emerging use cases, a new class of highly personalized virtual and physical AI agents is gaining momentum. These AI agents – consumed on smartphones, via AR glasses and other wearables, or embodied through a companion droid – represent a fundamental shift in human-computer interaction.

An AI agent can serve as a highly personalized assistant for consumers, offering services like proactive scheduling, real-time language translation, immersive navigation, adaptive learning and content curation across devices such as smartphones, AR glasses, or laptops. In the enterprise context, AI agents can automate workflows, manage routine communications, assist with knowledge retrieval and support frontline employees with real-time recommendations, effectively acting as a smart interface between those employees and complex systems.

On the other hand, a physical AI agent, such as a service robot or an autonomous droid, can assist in enterprise environments. They can take on roles in logistics, surveillance, warehouse operations, or even customer services, where they can handle repetitive, dangerous, or time-sensitive tasks while continuously learning from and adapting to these environments. For consumers, they can provide physical support paired with sentient behavior for various tasks.

A critical distinction is emerging between on-demand AI agents (which are invoked by the user) and always-on AI agents (that proactively assist autonomously). The personalized AI agent of 2030 will have a pervasive presence, embedded in our devices, environments, and interactions. Different types of AI agents present unique challenges in terms of compute demands, latency sensitivity and network resource consumption. Importantly, proactive AI agents will consume more resources and also need to be carefully managed to ensure user privacy and safety.

Convergence of networks, devices and content

The rise of AI agents and AI-native applications is driven by the commercial alignment of three core enablers: networks, devices and content. 

Modern 5G networks are evolving beyond mobile broadband to deliver differentiated capabilities, such as deterministic latency, high uplink performance, improved handovers and ultra-reliable edge access. APIs now enable developers to tap into these capabilities.

Devices have also matured to support AI-native experiences. Smartphones are evolving into multimodal AI terminals with enhanced sensors and dedicated AI accelerators. Meanwhile, AR devices are advancing along two distinct paths: Sleek camera-enabled smart glasses for casual consumption, or more immersive, full-stack systems designed for rich, continuous interaction. These devices act as sensor ingestion hubs and delivery endpoints for AI agents, enabling contextual, always-available assistance.

Finally, the surge in GenAI capabilities allows content – whether text, images, or immersive environments – to be generated at unprecedented scale and quality. Multimodal media, including 2D, 3D and volumetric content (via, for example, techniques for rendering 3D imagery such as neural radiance fields2 or Gaussian splatting), are unlocking new engagement formats. These formats are particularly suited for immersive AR environments and physical AI agents, allowing them to reason about and operate within complex, spatially rich contexts. 

Shifting traffic characteristics

As AI and GenAI become increasingly integrated into personalized and immersive experiences, the nature of network traffic is undergoing a fundamental transformation. AI-native workloads are introducing new traffic dynamics that are more bidirectional, context-sensitive, and therefore uplink-intensive:

• Personalized content with mainly downlink-centric growth: GenAI enables at-scale creation of hyper-personalized content – from entertainment to education – to be tailored in real time for each user. This increases engagement, retention and consumption, something already seen today. However, despite increasing the load on downlink, the traffic impact is manageable with the current 5G spectrum.
• Immersive interactions with mainly uplink-centric growth: Where AI truly starts to reshape traffic is in real-time, immersive interactions, especially those involving multimodal assistants or AI agents embedded in AR experiences. These systems rely on a constant uplink for video streams, sensor data, and conversational cues, followed by contextual inference and real-time adaptation. The shift from cloud-based AI to on-device or edge-executed GenAI amplifies this trend by lowering latency, but it still requires continuous uplink for personalization and environmental awareness. AI agents may also pull additional information from various other sources, which also increases traffic toward the AI agents.
• Semantic compression technologies: Emerging avatar-based communication represents a new approach to traffic optimization by transmitting high-level semantic data instead of video. If done on a device, this can result in significantly reduced data rates, particularly in controlled environments. However, their broader traffic impact is expected to remain limited in the near term as adoption will likely be confined to closed ecosystems and enterprise use cases.

These emerging classes of bidirectional traffic include real-time queries, streamed context, inference inputs and outputs, as well as orchestration commands. This leads to a new traffic profile that differs sharply from traditional patterns, in terms of volume, peak versus average characteristics, latency requirements, as well as packet size and frequency.

Not all GenAI traffic is equal

As GenAI continues its rapid integration into everyday applications, it’s important to recognize that not all GenAI-powered experiences will have a meaningful impact on network traffic. While nearly all future apps, from productivity tools to creative platforms, will incorporate GenAI in some form either on the device or in the cloud, only a subset will drive mobile traffic growth: Namely, those that enjoy a wide-scale adoption and require access to the cloud or content at high data rates. To understand impact, it is safe to disregard:

• High-rate but low-adoption applications, such as professional video editing or cloud-based 3D rendering. These are bandwidth-intensive but niche in usage, and therefore capable of being absorbed by peak-dimensioning of networks.
• High-adoption but low-rate GenAI applications, such as text-based chatbots that are ubiquitous but lightweight in their data demands.
• Low-adoption and low-rate applications, for example, occasional real-time audio translation.

Significant network impact will stem from applications that are both data-intensive and widely adopted, including:

• Video-based AI assistants that use real-time video feeds for interaction, requiring constant uplink/downlink flow and semantic understanding which can unlikely be provided by a GenAI model on the device.
• Immersive gaming or gamified environments powered by sophisticated GenAI-driven characters and environments, potentially combining multi-user streaming with dynamic, procedural and volumetric content generation.

These categories stand out as potential drivers of net-new traffic, particularly when experienced through AR devices with always-on assistants. Their data intensity comes from content rendering as well as from continuous AI inference and environment interaction, creating persistent uplink and downlink demands.

These new high-rate, high-adoption applications are likely to define the next wave of traffic drivers and subsequently impact spectrum requirements, network planning, investment, and ecosystem alignment.

Although uplink-heavy video calls and downlink-heavy media streaming might be consumed on AR glasses in the future, it is not clear yet if it will constitute net-new traffic or substitution and so it is not considered here.

Uplink, downlink and future spectrum implications

To obtain the actual global averages of usage minutes, and thus uplink and downlink rate requirements, Pareto and power distribution laws are applied. Specifically, it is assumed that of the population that will own a smart or AR headset in the future, 20 percent are power users, such as developers, heavy gamers or influencers, and 80 percent are median users with light usage for navigation and occasional use of an AI assistant. Furthermore, it is assumed that power users are using the device for 100 minutes per day over 5G, while the median users use them for 10 minutes per day.

Based on the assumption that at some time in the future, adoption of AR headsets will reach 20 percent, average usage will amount to approximately 5.6 minutes per day. A medium-quality AI agent implementation at 0.7 Mbps uplink and 2 Mbps downlink would cause an approximate increase in uplink by 47 percent and downlink by 14 percent. This increase will require proper network dimensioning and optimization, uplink improvements and more spectrum.

AR device uptake is predicted to grow at pace, though predictions still differ significantly. The most optimistic predictions indicate that about 20 percent of the US adult population will likely own an AR device by 2028, while other predictions put this some years later. Is it important to note that the growth is predicted to continue for years to come, meaning a 20 percent population uptake is a rather modest prediction long term. The traffic impact analysis looks at adoption rates of 10, 20 and 40 percent.

The resulting growth in uplink and downlink requirements for different quality AI assistants are shown in Figure 16. Projected growth in requirements is substantial; however, differentiated connectivity with slicing has the potential to improve spectral efficiency, enabling high-quality connectivity for these and other high-bandwidth, conversational applications.

AR is likely to incentivize hands-free and mobile usage – therefore, it may happen that more traffic will be consumed over 5G compared to indoor only Wi-Fi. In summary, the quantified uplink and downlink traffic increases due to the example application of an AI assistant application running on AR glasses using
5G. Note that the analysis provided focuses on the average rate increase, rather than deriving detailed distributions or specific regional variations. It is also based on the population average, rather than peak requirements.