It’s 2040. How are new devices helping decade-old 6G networks keep up with AI demand?

What you should know:

• AI applications and traffic patterns will evolve faster than mobile network standards, making adaptability a core 6G design consideration.
• 6G can make QoS behavior more dynamic while preserving the established flow-based QoS principles: core-network session management and traffic classification paired with RAN resource management and packet treatment.
• Device context, alternative QoS profiles, real-time adaptation and performance-driven protocols can work together to meet evolving application requirements while adhering to operator-defined policies.

AI experiences increasingly distribute intelligence across the compute continuum — devices, the edge and cloud — making connectivity a key factor in enabling the user experience. A mobile assistant may begin with a text request, inspect a live camera view, exchange audio and coordinate with cloud services within one interaction. To the subscriber, it is one task. To the mobile communication system, its traffic profile and priority may change several times during one session.

Operators may seek to support that variability without reserving peak resources continuously, exposing unrestricted network controls or requiring separate integration with every application provider. Rather than anticipate every application that may emerge, 6G can build on established QoS principles while enabling increasingly capable devices and networks to adapt as application requirements change to deliver greater efficiency and open new service opportunities.

Design 6G for applications still taking shape

Compute and connectivity have converged for years. Extended reality (XR) showed how rendering could be split between a device and the cloud, making network performance part of the experience. Distributed AI extends that model across the compute continuum so intelligence runs where it makes the most sense.

The challenge is designing for applications still taking shape. AI applications already run over 5G and Wi-Fi, but their capabilities, traffic patterns and performance requirements will continue evolving before and after commercial 6G systems arrive. Because an early 6G implementation may remain deployed for years, the question is not whether AI can use the system as a reliable transport, but how efficiently the system can scale and adapt as new applications and service requirements emerge.

A robust approach is to make the initial system configurable: devices can adapt to local conditions, while network schedulers can evolve without requiring a new standardized procedure for every application.

Understanding diverse AI traffic patterns

Beyond improving network operations, AI also generates traffic whose network impact must be understood.

3GPP is considering how AI and machine learning use cases may shape 6G system requirements. Deployed applications provide observable traffic patterns, emerging agentic applications indicate how autonomous interactions could change traffic and physical AI remains harder to characterize as deployments and test beds develop. Future applications may also distribute inference across devices, edge infrastructure and the cloud, while human-machine-agent interactions create new combinations of traffic, timing and coordination.

As a general rule, there is no single AI traffic profile. For example, text may use little data, video understanding can drive substantial uplink demand, and multimodal experiences may alternate among text, audio and video. Which applications scale will shape future system requirements, especially uplink coverage and capacity, so operators need a way to serve the emerging demand without designing the entire network around an uncertain peak.

AI also changes the latency equation. Inference currently can dominate time to first token, but faster models and compute will make communication delay a more significant factor in determining the user experience. Efficiently transporting traffic bursts, including allocating radio resources when a device knows a burst is urgent, may therefore matter more to whether interactive AI feels immediate and dependable. It may also affect whether autonomous tasks performed by agents on the subscriber’s behalf can be completed reliably.

Physical AI adds further uncertainty as robots balance local processing, safety requirements and coordination. Across these cases, demand varies by application, modality and time, which calls for a more dynamic approach to QoS.

Balance control and responsiveness in 6G QoS

5G QoS establishes a useful division of responsibilities. The core network translates subscriptions and service agreements into policy, authorization and charging rules. The RAN applies those rules through packet treatment and radio scheduling.

6G can preserve that framework while making execution more responsive. Guaranteed bitrate mechanisms suit stable demand, but video, gaming and AI applications can change bitrate, modality or performance need within the same flow. Dynamic QoS can direct resources where they may improve the experience instead of provisioning every interaction for its possible peak. This could enable the intelligent 6G system to use resources more efficiently by optimizing performance across users and applications.

The mechanisms can remain within an existing PDU session and preserve backhaul QoS mapping where appropriate. The architectural shift is therefore not from network control to device control. It is from fixed execution to adaptation within network-defined policy.

Leverage device context for adaptive QoS

Dynamic QoS requires information the network may not directly observe. A device can combine traffic classification, buffer state, radio conditions, application requirements and sensor context with subsystem monitoring. That intelligence can range from simple rules on constrained IoT devices to richer perception and reasoning on smartphones, vehicles and robots. It can allow selected device-side functions such as buffering, reordering and packet discarding to adapt within network-defined limits.

QoS establishment and mobility remain network-coordinated, while selected protocol parameters can vary within operator-defined ranges. When local adaptation is insufficient, the device can initiate network procedures to coordinate QoS, mobility and resource allocation. A service tier might permit greater flexibility, and the network can narrow or remove that flexibility if a device behaves outside expectations.

Device awareness also makes differentiated treatment conditional. A device can remain on best effort service while performance is sufficient, then request added treatment only when application needs and network conditions indicate that it would materially improve service. Local traffic identification reduces dependence on every application provider integrating with network APIs, while application-provided information can strengthen the decision when available.

Application-aware QoS for 6G systems

Dynamic applications create two different QoS problems. The traffic can change in kind, or its performance demand can change in degree. A single interaction may move among text, audio and video while the subscriber expects it to continue seamlessly.

When traffic type changes, the multimodal interaction described earlier may move from text to video within the same QoS flow. Because IP-flow filtering alone may not distinguish among text, audio and video modes within that flow, alternative QoS profiles (AQPs) could let the device or network select more appropriate treatment. With enhanced AQP switching, local coordination between the device and RAN could switch among core-authorized profiles faster than repeatedly modifying QoS through the core, without requiring a separate network slice or PDU session.

When performance demand changes, the traffic type may remain the same while its resource needs rise or fall. For example, a video or cloud-gaming stream may move from a simple scene to one with far more motion and visual detail, requiring substantially more throughput to maintain the same quality. Guaranteed bitrate reservations can over allocate resources or miss temporary peaks. Real-time QoS adaptation could instead define an allowable envelope across throughput, latency and error rate. In addition to the network monitoring the downlink buffers, the device could signal changing needs from its local knowledge buffer or application state, and the scheduler could respond within policy.

Adaptable radio protocols for responsive AI

Dynamic QoS changes how the system treats traffic. A third enhancement, the performance-driven protocol stack, makes selected radio protocol behaviors adaptable within operator-defined limits.

Fixed timers, retransmission rules and other deterministic parameters are straightforward to specify and test, but one value may not perform best in every condition. A device can use immediate radio, traffic and application context to adjust selected behavior without changing network policy.

Together, these mechanisms create a multi-layer adaptation framework that operates at different layers and timescales moving from behavior-driven to a performance-driven protocol stack. The device could adapt selected protocol parameters within RRC-configured ranges, the RAN could manage near-real-time selection among NAS-configured alternative QoS profiles, and the RAN adjusts radio resources using device feedback, with each behavior remaining subject to network authorization and policy.

Adaptable schedulers and device-side traffic awareness can therefore support future use cases without requiring a standardized feature or application-provider integration for each one. That flexibility matters when an early 6G release remains deployed as traffic patterns and service models evolve. Operators gain room to respond through software and policy, while subscribers depend less on assumptions made when the system was first designed.

Design 6G to rise with demand

Not every future AI application will need a new network mechanism. Many will work over Wi-Fi, 5G or 6G. But when the system can make a meaningful difference, 6G should be ready.

We cannot know what every AI application will ask of the network. But we can design 6G so that each new generation of devices makes the system more aware, responsive and capable. Years after deployment, the system will not simply support the future. It could be designed to adapt as requirements evolve.