AI is flipping the network traffic model

Most mobile networks were built for a downlink-heavy world. AI is flipping that assumption.

As AI moves from smartphones into glasses, earbuds, autonomous systems and other connected devices, the network is increasingly being asked to take in the world around us — not just deliver information back to us. Video, audio, sensor data and other contextual inputs can flow upstream to AI systems for processing, often continuously and in real time.

That changes more than the volume of traffic. It changes its direction, intensity and timing. And as AI experiences become more immersive and autonomous, uplink performance could become a critical factor in whether they work as intended.

End-user study results – AI/AR enabled Practical Task Helper

In a recent AI/AR glasses end-user study at Ericsson, we tested user preference on a service concept called AI Practical Task Helper. This service is designed to help users learn to complete unfamiliar tasks, such as repairing a toaster, through contextual, real-time, step-by-step guidance. We tested four levels for service:

• Level 1 – Reactive (command-response): The user asks, “How do I fix this?” and the system recites the steps via audio.
• Level 2 – Guided (prompted by the system, the user decides): For users who are visual learners, the system suggests a diagram. A floating schematic appears in the user’s view upon confirmation.
• Level 3 – Autonomous (the system acts, the user can override): As the user finishes step one, the system automatically identifies the next bolt and anchors a digital arrow precisely to that specific screw on the table.
• Level 4 – Immersive (seamless, always on): Anticipating a mistake before the user makes it, the system flashes red. An AI avatar guides the user’s hands, never obscuring the view of the tool.

The study results (shown in Figure 1) indicate that Level 3, representing an autonomous service, is the preferred service level for the Practical Task Helper. Specifically, 33% of users expressed a preference for Level 3, while only around 20% preferred the other service levels on average. At this level, the AI agent is expected to act on the user’s behalf by leveraging rich contextual information captured through AI/AR glasses. This allows the agent to provide a proactive, context-aware service rather than relying on a reactive approach with limited or no contextual awareness.

The impact of AI-based applications on traffic growth

Context-aware AI-based applications like Practical Task Helper drive traffic growth in various ways. One obvious impact comes from feeding AI processes information about the real world, as well as how these applications split compute between the device and the cloud. Many of the new AI experiences rely on cloud-based inference, where video streams, multimodal inputs or sensor data are sent upstream for processing.

In this context, traffic is no longer dominated by downloading content; instead, users and devices are increasingly sending rich streams of data into the network so the cloud can interpret scenes, generate outputs and return results in real time. At the same time, downlink (DL) traffic continues to grow through the delivery of generated media, textures, responses and other AI-produced content. The result is a more interactive traffic pattern with greater intensity in both directions.

Today’s per-device UL requirements are often in the 1-5 Mbps range – relatively low compared to DL – but cell-edge users frequently struggle to achieve even these moderate bitrates. Figure 2 shows that 40 percent of 5G networks already fall short of the minimum UL requirement for AI services.

Line chart comparing minimum uplink throughput in 5G networks across Europe, Asia, North America, and Oceania. All regions show decreasing percentages as throughput increases. A marked 5 Mbps threshold intersects the curves at about 60%, indicating that roughly 40% of 5G networks fall below the minimum uplink requirement for AI services.

Figure 2 – UL throughput across global regions

Scale presents an even bigger challenge. As the number of AI-capable devices grows and more users adopt video-heavy, sensor-rich and always-on AI services, the aggregate UL load will rise quickly. This combination of traffic growth and device growth is what increasingly pushes overall UL throughput and makes UL performance a factor of growing importance for network quality and differentiation.

The always-on capabilities of AI agents must also be considered

Whereas humans are limited in how much time we have to consume content, agents that assist humans can be active 24/7. In the longer term, AI technology will likely create additional users – “AI personas” – that are independent actors who work alongside humans and have similar communication needs. Agents working on our behalf in work tasks, robots, drones and autonomous vehicles are early examples of this. As AI technology continuously improves services for humans – through personalization, for example – it is likely to increase the amount of time that humans spend using those services.

Ericsson’s consumer research findings indicate that users expect AI agents to save time in everyday tasks and that much of this saved time is likely to be redirected toward other online activities. This suggests not only an increase in overall usage, but also a wider mix of services contributing to network traffic.

To understand what this means for networks in practice, it’s helpful to look at the specific use cases driving that growth.

Four common AI use cases impacting future networks

Table 1 presents four of the most common AI use cases – live AI, AR video calling, agentic AI and on-device AI – in terms of devices, experience needs, compute models and network impact.

Table 1 – Comparing AI use cases across devices, compute and network impact

While all four of these use cases contribute to traffic growth, it is the shift toward cloud-based compute in live AI and agentic AI that most directly drives uplink (UL) demand — and that shift is largely determined not by design choice, but by the physical constraints of the devices themselves.

Device evolution trends – implications for the use cases

The device landscape is changing rapidly, but not predictably. A useful way to frame it is in terms of “intelligence density” – that is, how much useful AI a device can deliver per gram, per cubic centimeter and per watt. This directly determines:

1. which experiences can be always-on
2. which can be truly interactive
3. which will still require the edge or cloud. 

For the use cases in Table 1, the split between local execution and offloaded execution is therefore not a design preference – it is largely set by silicon capability, memory bandwidth, thermals and form-factor constraints. That split is also what drives the network footprint: offload increases UL demand (sensor/media upload and prompts) and can introduce tighter latency coupling between UL and DL.

Across these device trends, the common theme is that future AI experiences will be shaped by the practical limits of memory bandwidth, sustained thermals and battery, as well as the physical constraints of each device category, such as optics and form factor in AI/AR glasses. Figure 3 illustrates that hardware is growing, but thermal and form-factor limits cap what can run locally.

The following numbers were used as input:

The key takeaway is that NPU performance grows the fastest, while memory bandwidth remains a critical system constraint.

As devices move toward lighter and more compact designs, intelligence will become more distributed across the device, companion hardware and the edge or cloud. That makes device evolution directly relevant to network evolution: the harder it is to sustain large AI models locally, the more important UL capacity and efficient offload become. Figure 4 illustrates GenAI model size versus on-device capacity, showing that advanced AI models outpace local hardware, making cloud offload and UL essential.

Horizontal bar chart showing the relationship between AI model size and reasoning capability. Smaller models running on today’s devices provide basic to good reasoning, while larger models with more parameters deliver strong and state-of-the-art reasoning. A note indicates that future devices in five years are expected to support more advanced on-device AI models.

Figure 4 – GenAI model size versus on-device capacity

AI-enabled use cases and continued mobile broadband demand drive traffic growth

By 2031, mobile networks are likely to carry substantially more traffic than they do today, but the change is not only about volume. It is also about the mix of services behind that growth. The projected increase comes from a combination of continued mobile broadband demand and a wider set of AI-enabled use cases operating over public wide area networks.

Ericsson Mobility Report predicts that within five years, UL traffic will grow significantly — in a medium adoption scenario, a three-fold increase is expected. GenAI applications alone account for nearly a third of all UL traffic. At the same time, we cannot dismiss the DL data; depending on the GenAI adoption scenario, 89, 86 or 80 percent of the traffic is still transmitted in the DL direction. The UL growth, however, will push UL/DL asymmetry well beyond today’s levels, as you can see in Figure 5.

Ericsson has been investing in UL solutions ahead of this shift. Beyond site upgrades, future growth will require more spectrum. Given the above, the first wave of network improvements should concentrate on site upgrades but looking further the elevated user experience requirements together with increased traffic demand, the network will inevitably reach a point where site software and hardware improvements will not suffice. Additional wide area spectrum in the mid- and centimetric-band spectrum is then required to add capacity. The additional capacity will address increasing user experience expectations for DL applications and for UL heavy use cases such as autonomous vehicles, AI-assisted surveillance and AI/AR glasses.

RAN enhancements and standardization considerations for AI traffic

Standardization plays an important role in making networks AI-ready. Building on the XR work already done in 3GPP, the next step is to evolve radio access network (RAN) support for traffic that is more UL-heavy, bursty and latency-sensitive, while keeping solutions broad enough to remain relevant as AI services continue to change. Improving RAN traffic awareness will help the network to utilize the radio resources/spectrum more efficiently, thereby enabling it to serve the increasing traffic volumes introduced by AI traffic.

Uplink as a strategic priority

The Practical Task Helper study offers a small but telling signal: when given the choice, users gravitate toward AI experiences that are richer, more contextual and more continuous. Multiply that preference across millions of devices, always-on agents and an expanding universe of AI wearables, and the network implications become clear. UL can no longer be treated as an afterthought. Operators and device vendors that recognize this shift early will have the most to gain.

Read more

• Boosting UL performance for 5G monetization
• AI powered smart glasses promises and issues
• Three augmented reality trends and the future of AR
• How AI-powered devices will drive the shift to uplink-heavy networks
• AI will drive future uplink – Ericsson Mobility Report