AI Network Standardisation Moves Towards AI-Native 6G

Artificial Intelligence has been used in mobile networks for many years, particularly for optimisation, automation, analytics and network management. With 5G-Advanced and now 6G, however, the discussion is changing. AI is increasingly being considered not simply as a tool that can be added to the network, but as something that could influence how the network itself is designed.

A useful overview of this transition was presented by Hye-Young Lee from the Telecommunications Technology Association (TTA) during the Mobile Korea 2025 Regional Workshop in October 2025. Her presentation, 'AI Networks Standardization – Current Status and Future Trends', looks at how AI-related work is developing across 3GPP, O-RAN and the wider telecoms ecosystem.

One of the most important points in the presentation is the evolution from AI-assisted networks towards AI-native networks.

In the early stages of the IMT-2030 discussion, AI was one of several important capabilities associated with 6G, alongside areas such as sensing and computing. The emergence of initiatives such as the AI-RAN Alliance and the Global Telco AI Alliance during 2024 helped push AI much further towards the centre of the industry discussion.

The presentation describes two complementary tracks. On one side are standards organisations such as 3GPP, ITU and O-RAN, dealing with areas such as openness, virtualisation, interfaces, data and network intelligence. On the other are industry initiatives looking at how these capabilities can actually be implemented and commercialised. The AI-RAN Alliance, for example, introduced the now familiar concepts of AI-for-RAN, AI-and-RAN and AI-on-RAN.

The 3GPP work did not suddenly begin with 6G. Network intelligence has been evolving progressively through 5G and 5G-Advanced.

The 5G Core introduced the Network Data Analytics Function, or NWDAF, while subsequent releases expanded analytics, automation and AI/ML support across network management and the RAN. AI/ML-related work now spans several 3GPP groups. SA2 looks at architectural enablers for network automation and data collection and exposure, SA5 addresses management and analytics, RAN3 considers RAN optimisation and intelligence, while RAN1 and RAN2 have been studying AI/ML for the radio interface.

The use cases have also become progressively more ambitious. Earlier work concentrated on areas such as energy saving, load balancing and mobility optimisation. 5G-Advanced expanded this into areas including coverage and capacity optimisation, network slicing and more sophisticated coordination between network functions.

For 6G, the ambition is broader again.

The presentation highlights three areas of early 3GPP study. SA1 was considering new 6G use cases and service requirements, including AI agents and AI-related services. Radio studies were looking at efficient transfer of data for AI/ML and the possibilities for AI-native radio architectures. Work on the 6G system and core network was considering both 'AI for Network' and 'Network for AI', together with the role of computing resources.

This distinction is important.

AI for Network is about using AI to improve the operation of the communications network itself. This could include optimisation, automation, prediction, resource allocation, mobility management, energy saving and many other functions.

Network for AI, on the other hand, asks how the communications network should evolve to support distributed AI workloads, AI agents, model training and inference, edge computing and increasingly AI-intensive applications.

There is potentially a third dimension as well: the network becoming a platform through which AI capabilities can be exposed as services.

This is why one of the strongest messages from the presentation is that AI should not simply be treated as another add-on feature for 6G. The speaker argues that AI should become a design principle for 6G.

O-RAN is following a similar path. Its architecture already provides mechanisms such as the Non-Real-Time RIC, Near-Real-Time RIC, rApps and xApps through which data-driven optimisation and AI/ML can be introduced into the RAN. The presentation points towards further evolution of this architecture for AI-native and cloud-native 6G RAN, including real-time resource management, AI function integration, data transfer and model management.

Ultimately, the ambition is to move towards networks with much higher levels of autonomy. But achieving this requires more than increasingly capable AI models.

Standards become important for making AI work across multi-vendor networks. The presentation identifies four particularly important areas: interoperability through common interfaces and models, predictable performance, accessible and standardised management of network data, and sustainability including AI model lifecycle management and energy efficiency.

The full presentation is embedded below and is worth watching for the slides showing how the different standardisation activities fit together.

What Has Changed Since the Presentation?

The presentation was delivered in October 2025. As of August 2026, some of the activities it describes have progressed significantly.

Perhaps the most important development at the ITU level is that Artificial Intelligence and Communication, or AIAC, is now explicitly one of the six IMT-2030 usage scenarios.

In February 2026, ITU-R Working Party 5D agreed the draft minimum technical performance requirements for IMT-2030. The framework contains 20 performance requirements and is based on six usage scenarios: Immersive Communication, Hyper Reliable and Low-Latency Communication, Massive Communication, Ubiquitous Connectivity, Artificial Intelligence and Communication, and Integrated Sensing and Communication. Formal approval of the performance requirements by ITU-R Study Group 5 is expected in December 2026.

That represents an important progression from simply describing AI as an important attribute of 6G. AI and communication now form an explicit part of the framework against which candidate IMT-2030 radio technologies will eventually be evaluated.

3GPP has also moved forward.

Release 19 was frozen in December 2025, while Release 20 and Release 21 are currently open. 3GPP continues to describe Release 20 as primarily the study phase for 6G, with Release 21 marking the official start of normative 6G specification work.

On the service side, the Release 20 TR 22.870, Study on 6G Use Cases and Service Requirements, moved under change control in March 2026. More importantly, the new Release 21 TS 22.270, 6G Service Requirements, has already been created. A first draft appeared in March and version 0.2.0 was uploaded in June 2026.

There has also been an important milestone on the radio side. TR 38.914, Study on 6G Scenarios and Requirements, moved under change control in June 2026. This first 6G RAN study establishes the scenarios, radio capabilities, deployment assumptions and requirements that will feed subsequent detailed radio studies and eventually the Release 21 specifications.

The AI discussion within the 6G architecture work is also becoming much more concrete.

3GPP SA2 contributions in 2026 show that 'Network for AI' is no longer just a high-level concept. The Release 20 6G architecture study includes proposals covering AI-agent discovery and communication, AI-agent identity and authorisation, characteristics of Generative AI traffic, support for AI model training and inference, capability exposure to AI applications and even AI services delivered through the 6G Core. These are study-stage proposals rather than agreed 6G specifications, but they show the direction in which the architectural discussion is moving.

Meanwhile, AI/ML work inherited from 5G-Advanced continues to establish many of the foundations needed for 6G. 3GPP describes efficient and valid data collection as a key requirement for model training and inference, with Release 18 already introducing data-collection and signalling enhancements for AI/ML in NG-RAN.

O-RAN has moved forward as well.

The O-RAN Alliance completed O-RAN Release 5 in June 2026. Among its features are enhancements to the AI RAN framework, including AI/ML workflow services that combine the Non-RT RIC and Near-RT RIC for AI model development and training. Release 5 also adds support relevant to AI-based Massive MIMO optimisation, network energy saving and AI/ML security controls.

O-RAN-R006 is now under development and will be the first O-RAN release to include specific 6G study features and study items. O-RAN has identified six tracks for its 6G work covering use cases and requirements, architecture, fronthaul, management and automation, O-Cloud evolution and security.

The AI-RAN Alliance has also evolved considerably since the TTA presentation. By February 2026 it had grown to 132 members and was demonstrating 33 AI-RAN innovations at MWC 2026. It has since published a more detailed AI-RAN architecture, component definitions and orchestration frameworks covering the sharing of infrastructure between RAN and AI workloads.

It is important, however, to distinguish this ecosystem work from formal global standardisation. Organisations such as the AI-RAN Alliance can develop architectures, blueprints, reference implementations and implementation experience, while 3GPP, O-RAN and ITU address different parts of the interoperable specifications and international standards framework.

There is another important point that can easily get lost when discussing an “AI-native” network.

AI standardisation does not necessarily mean standardising the AI models themselves.

3GPP has previously made clear that it does not currently intend to specify particular AI/ML models. Instead, much of the value of standardisation lies around the model: how data is collected and exchanged, how capabilities are exposed, how training and inference functions interact, how models are deployed and managed, how network and device functions coordinate, and how performance remains measurable and controllable.

This distinction may ultimately be one of the most important aspects of AI-native 6G.

A future network could allow vendors and operators considerable freedom to innovate with different models and algorithms while standardising the interfaces, data structures, procedures, lifecycle management and performance expectations required for those AI implementations to coexist.

The question for 6G therefore may no longer be whether mobile networks will use AI. They clearly will.

The much harder standardisation question is which parts of an AI-native network need to be interoperable, observable, controllable and globally standardised, while still leaving enough freedom for AI technology itself to continue evolving rapidly.

The work now underway in ITU, 3GPP, O-RAN and the wider AI-RAN ecosystem is beginning to provide the answer.

Related Posts: 

Posted by on

Comments

Post a Comment