The emerging impact of generative AI on telecom networks by 2030

The development of Large Language Models has not yet translated into a major traffic disruption. However, the increasing automation of requests, driven by the rise of agents and connected machines, may influence Internet traffic patterns in the future. This could give networks a more important role in enabling advanced Generative AI use cases.

This change will be more qualitative than quantitative: more uplink, more latency-sensitive flows, more machine-originated interactions. These flows will carry value and could be eligible for better monetization for telcos, through differentiated quality of service.

From back-end tool to interaction layer

In just two years, Large Language Models (LLMs) have reached more than one billion users. This creates a paradox: Generative Artificial Intelligence (GenAI ) inference demand has grown exponentially, but its current impact on access-network traffic remains limited compared with video. This article examines how Generative AI may influence telecom networks in the coming years.

First impact on GEN AI on networks

OpenAI reported more than 2.5 billion consumer messages per day by July 2025, which likely corresponds to roughly 1–4 trillion visible consumer tokens per day for ChatGPT. Extrapolated over a year, this remains very small compared with total global telecom traffic. Ericsson estimates GenAI’s current impact on mobile traffic at around 0.06%, which is consistent with the idea that text-based interactions remain marginal in volume. Even image generation, while growing quickly, does not materially change the overall picture at network scale.

Both Ericsson and Nokia indicate in their traffic reports that consumer video streaming and social media still dominate traffic volume. Artificial Intelligence is emerging and appears rather as an incremental driver than a replacement of existing traffic categories.

However, some early changes can be noticed:

• Slight increase in uplink and session interactivity (voice, multimodal prompts, context syncing).
• Fewer page loads or API calls per task, but each interaction is heavier in compute, context and retrieval.
• Early signs of machine-originated traffic (agents, copilots calling APIs). Even if a large share of web activity is already non-human through crawlers, scrapers and bots, the next shift is traffic increasingly driven by execution rather than browsing.

Nevertheless, this first phase remains mostly human-centric.

When interactions stop being user-driven: Agents x Robots

The first game changer is the rise of agents, which will take more autonomous decisions and amplify human actions. We can expect assistants to suggest actions without being asked and perform some actions autonomously on behalf of users. To throw a few examples:

• Search information by querying multiple websites, where LLM-driven crawling can multiply hits
• Auto-publish at scale: agents can generate articles, product pages, posts, and potentially media assets

Even if the number of tokens generated by an LLM query is increasing, most token exchange stays within data-center environments. For telecom networks, the main effect is less the tokens themselves than the multiplication of requests, retrieval loops, API calls and machine-to-machine sessions. Nokia expects agentic-AI WAN traffic to rise strongly, with machine-to-machine traffic becoming a meaningful new category reaching 537 EB/month in 2034 (10 times compared to 2025).

The second game changer is the rise of robotics after a long period of relative stagnation. IFR reports more than 542,000 industrial robots installed in 2024 and a global operational stock of around 4.66 million units. Professional service robots are also expanding, with almost 200,000 units sold in 2024 and a robots-as-a-Service fleet growing by 31%.

Robotic systems generate and adapt content continuously on a larger scale. Connected cars are an example of a first generation of such robots. Current wide-area offloaded traffic is still constrained by architecture, costs and local processing, but the trend points toward more continuous exchange of video and control data across networks.

Agents combined with robotics will provide connected devices with capabilities to take decisions, such as applications using cameras to interpret the environment. Ericsson also points to video-based AI assistants, smartphone camera assistance and smart glasses as possible drivers of uplink growth. This has the potential to increase payloads of multimedia files and real-time context flows.

The non-human share of sessions can therefore increase significantly.

Consequences for the network : a new case for Low latency and edge

Ericsson’s Mobility Report and Nokia’s traffic forecast both suggest that the real change is not only more traffic volume, but a change in traffic patterns.

First, uplink becomes more important. If devices are continuously sending data such as video, context or sensor input, the traditional imbalance between download and upload starts to shift. This may be triggered by automated prompts, high-resolution telemetry, automated assistants and agent feedback loops. Session density will also increase, involving many more interactions per user or task and more machine-to-machine traffic.

Second, latency matters more. Some of these interactions, especially when they involve real-world environments, need to happen quickly. We can imagine capture of images that need to be analyzed and augmented in real-time for applications ranging from field technicians to consumer assistants. We can also imagine conversations that are instantly translated, summarized and enriched with accurate responses for interactions between agents, customers or participants in meetings.

This leads to the third consequence, which is the distribution of tasks within the network.

The majority of the work is still borne by Hyperscalers data centers , which will continue to concentrate computing power. But a single interaction may increasingly involve some processing on the device (user context, small local model, privacy filtering, first-token or draft generation), some at the edge or in private cloud (session continuity, regional cache, speech or vision enhancement, private RAG, vector search, sensitive document processing), and some in the cloud (coding assistants, batch generation, back-office automation).

The requirement to optimize latency for critical applications will lead to the optimization of the inference performance. This could increase the relevance of telecom operators, provided they can expose edge, latency, locality and quality-of-service capabilities as programmable services

Conclusion: towards a traffic pattern transformation?

By 2030, machines may account for a large share of interactions, while humans still dominate traffic volume and 70% of global WAN traffic will still come from non-AI sources according to Ericsson. Nokia projects AI traffic to become a major component of WAN growth by 2034, with strong expansion in both user-driven and agentic machine-to-machine flows. The real change is therefore not simply a bandwidth explosion, but a transformation in traffic nature: more uplink, more latency-sensitive flows, more inter-data-center traffic, and a shift from browsing to execution. This will affect the way networks are designed and may open new monetization opportunities for telcos.