Upgrade Wave: A Network Event Horizon

According to Counterpoint Research’s Global Smartphone Installed Base Forecast, the global smartphone installed base is expected to reach 4.4 billion by 2028. Currently, it is mostly running on conventional silicon, with no dedicated AI processing and little to no on-device inference capability.

Counterpoint data shows GenAI-capable smartphones accounted for just over one-third of all shipments in 2025. Going by current trajectories, this shipment share will cross the 50% mark in 2027, the point at which GenAI becomes the default rather than the exception in new devices. As that happens, the app ecosystem is likely to reach sufficient scale to accelerate the move from mostly text-based interactions toward multimodal and agentic applications. 

GenAI Smartphone Penetration and Mobile Network Demand Evolution

In the midst of this transition, key considerations for the industry and operator network planning include:

• Upgrade wave is not uniform: It tracks income levels, market maturity and device price points. AI silicon is already arriving in premium devices in developed markets. The mass-market transition, where volumes are largest, will come later and could be further delayed, with memory chip shortages likely to impact supply over the medium term.
• Changes in per-device network footprint: As new GenAI-capable smartphones replace the legacy ones, AI-capable devices will generate more frequent, more data-intensive interactions. This means more uplink, more bidirectional traffic and more cloud roundtrips for complex tasks. Multiplied across hundreds of millions of new devices annually, the cumulative network demand implications are significant.

Smartphones are the highest-volume AI device category today and the primary driver of measurable consumer mobile traffic. But they are the floor of the network planning challenge, not the ceiling. Smart glasses already generate 40% uplink traffic, while industrial robots require five times the uplink capacity of conventional devices. Agentic AI is creating a new category of Agent-to-Agent traffic – machine talking to machine – with characteristics that bear little resemblance to conventional mobile data. The smartphone upgrade cycle is the entry point. What follows it is more demanding, not less.

The device transition is already priced into chipmaker roadmaps but does not seem to be accounted for in many operator network plans.

The Window is Open for Now

Most operators are using AI to optimise their networks, but few are adapting their networks for AI. The opportunities span infrastructure, spectrum, pricing and edge compute, and in each case, the chance for first-mover advantage remains. How operators respond in the next 24 months will also shape their starting position for 6G.

AI-Ready Network: Gaps, Implications and 6G Horizon

WRC-27 and the 2025-2028 smartphone upgrade cycle are the two data points operators should be following closely.

Recommendations

• Reconfigure for AI traffic behavior, not volume: Unlike videos, which are predominantly downlink, AI interactions are uplink-intensive, bi-directional and latency-sensitive. Videos took a decade to reshape mobile networks and although AI traffic is smaller today, it is growing and is different from day one.
• Smartphone upgrade cycle a leading indicator: By the time AI traffic is visible on operator dashboards, the GenAI smartphone upgrade cycle causing it will already be advanced. Counterpoint Research's shipment data shows the transition to AI-capable silicon is accelerating.
• Build 5G SA foundations first: Network slicing, guaranteed uplink QoS and API-exposed capabilities all require a mature 5G standalone core. Operators still on non-standalone architectures will find the higher-value moves significantly harder to execute.
• Position edge compute before hyperscalers do: Most AI inference today routes to hyperscaler cloud – operators carry the traffic but don’t own the compute relationship or the margin. Ericsson recommends moving inference to the cell site, co-locating network functions with AI applications at the far edge. Operators with edge infrastructure can intercept latency-sensitive AI workloads before they hit the core network. The window to establish that position ahead of AWS, Azure and Google is open but narrowing.
• Monetise network performance: Volume pricing models leave AI traffic value uncaptured. AI applications need guaranteed uplink performance and deterministic latency, and not just data allowances. The commercial mechanism already exists. Both Ericsson's Aduna and Nokia's Network as Code are built on the GSMA Open Gateway's CAMARA standard, exposing Quality on Demand APIs that let operators charge for guaranteed network performance. China Telecom's QoD API has logged 5 billion cumulative calls and $500 million in revenue for gaming, proving the model works. AI traffic is the logical next step.