0

Networks optimized for delivering content to people may need to be rebalanced for machines sending the physical world to AI

For most of the mobile broadband era, network traffic has been overwhelmingly about moving information toward the user. Streaming video, social media and cloud applications made downlink capacity the dominant network planning problem, and mobile networks evolved accordingly. AI could begin to invert that assumption.

As artificial intelligence moves from text-based applications toward multimodal agents, smart glasses, connected vehicles and eventually robotics, devices increasingly become producers of data rather than primarily consumers of it. Cameras capture video, microphones capture audio, sensors describe the physical environment. That information moves upstream for inference and reasoning, with a comparatively small response potentially coming back in the other direction.

The implication is that AI changes the directionality, timing and performance requirements of that traffic. At Mobile Future Forward, T-Mobile President of Technology John Saw offered an important qualification that this transition has not happened at scale yet.

“We haven’t seen much impact yet” from AI on network usage, Saw said, with most of the effect currently concentrated in transport networks and around large data centers. But, he added, “I think it’s coming.” His expectation is that the network impact becomes more significant as AI workloads distribute toward the edge, one reason T-Mobile has emphasized 5G-Advanced and greater uplink capabilities. 

For more from Saw, read my note, “The T-Mobile US kinetic token thesis and its role in the AI economy.” 

The data increasingly points in the same direction. Ericsson’s June 2026 Mobility Report analyzed traffic growth across 55 service providers during 2025. At 43 of them, uplink traffic grew faster than downlink traffic; at 17, uplink growth was more than 1.5 times the downlink growth rate. Ericsson attributes the current shift primarily to decidedly non-futuristic applications, including communications and collaboration, user-generated content and cloud storage. That creates the baseline onto which AI could add a much different class of traffic.

From consuming the internet to sensing the world

Today’s network remains highly asymmetric. Ericsson has previously measured uplink at only around 8% of total traffic across a sample of networks, while conventional video traffic was approximately 97% downlink and 3% uplink.

Even current generative AI is noticeably more symmetrical. Ericsson measured gen AI traffic in a mature mobile market at 74% downlink and 26% uplink, compared with a roughly 90:10 split for overall traffic. Some applications, including DeepSeek and Microsoft Copilot in Ericsson’s measurements, were approximately 50:50. But there is an important reality check in that gen AI represented only 0.06% of overall network traffic in that study.

The potentially disruptive change comes when AI acquires eyes, ears and other sensors. Ericsson estimates that AI/AR smart glasses performing cloud-based inference could require between 1 Mbps and 10 Mbps of uplink throughput, with some applications producing a downlink-to-uplink ratio of roughly 1:8 as cameras, audio and sensor data stream toward cloud AI systems. Its medium AI adoption scenario has uplink traffic reaching three times 2025 levels by 2031; under a high-adoption scenario, the increase reaches fivefold.

That compares with Ericsson’s forecast for overall mobile data traffic to grow by about 2.2 times between 2025 and 2031. In other words, the issue is that networks will need to carry more data and the traffic mix could become materially less asymmetric.

Counterpoint Research reaches a similar conclusion. It argues that multimodal AI interactions will increasingly require devices to send images, voice, video and contextual data upstream before receiving an inference result, while smart glasses, industrial systems and machine-to-machine agents expand the range of devices generating this traffic.

Physical AI makes the problem harder

Ericsson Head of Advanced Technology Mathias Riback pushed the argument into physical AI at Mobile Future Forward. Robotics and humanoid systems face fundamental constraints around onboard compute and battery life. Offloading some computation into the network could allow those systems to run larger models while consuming less local power, but doing so creates simultaneous requirements around connectivity, compute and latency.

Senior Vice President of AT&T Business Product Shawn Hakl described the same architectural transition from the enterprise side. “In the AI world we are seeing on average people using six to 12 models and those models could live anywhere,” Hakl said.

In this case, the application architecture stops resembling the familiar model of a user requesting content from a server. Sensors, agents, models and data stores may be distributed across devices, enterprise locations, edge infrastructure and multiple clouds. Hakl described cases where the traffic flow effectively reverses and large volumes of data move upstream while a relatively small instruction comes back. Video and audio already exhibit the pattern; generalized robotics could dramatically amplify it.

This also strengthens the case for edge compute. Hakl identified latency and data volume as fundamental determinants of where workloads ultimately run. As more content is generated dynamically at the edge rather than simply retrieved from storage, aggregation and processing closer to the source becomes increasingly logical.

The network-planning problem arrives before the traffic does

None of this means robots and AI glasses are about to overwhelm mobile networks. GSMA estimates that direct generative AI remains only around 0.2% of mobile network traffic today and explicitly cautions that operators could overprovision capacity in anticipation of AI growth that has yet to materialize. But network architecture is necessarily built ahead of demand.

The practical question for operators is whether the emerging workload mix is sufficiently different to change investment priorities in spectrum, radio configuration, transport, edge compute and 5G-Advanced capabilities. There is increasingly evidence that it is.

Mobile networks were optimized for an internet in which humans downloaded the digital world. Physical AI points toward a different model where millions, maybe even billions, of intelligent endpoints continuously describing the physical world upstream. If that transition occurs at scale, the next major mobile traffic problem will be figuring out what happens when those bits start flowing the other way.