Sovereign AI for 6G: Towards the Future of AI-Native Networks

Swarna Bindu Chetty, David Grace, Simon Saunders, Paul Harris, Eirini Eleni Tsiropoulou, Tony Q. S. Quek, Hamed AhmadiarXiv 2025 · DOI · arXiv

Abstract

As 6G transitions from being cloud-centric to AI-native, Artificial Intelligence is poised to become the literal backbone of telecommunications infrastructure. This shift presents a critical challenge: preventing vital networks from becoming "black boxes" controlled by external entities. This paper proposes a strategic framework for Sovereign AI, ensuring national and operator-level control over AI development, data, and governance within the emerging 6G ecosystem.

The Evolution to AI-Native 6G

In previous generations, such as 4G and 5G, AI served primarily as an auxiliary tool for specific optimization tasks, like predicting channel behavior or managing energy consumption. 6G represents a paradigm shift; it is designed to be AI-native from the ground up. In this new architecture, AI will manage everything from real-time resource allocation to autonomous network troubleshooting.

However, the rise of Generative AI (GenAI) and Large Language Models (LLMs) introduces a "cost trap" and a "sovereignty risk." Building proprietary models is prohibitively expensive, yet relying on third-party providers risks ceding control over sensitive network data and potentially violating national regulations. Sovereign AI addresses this by building a 6G architecture that is secure, explainable, and aligned with local laws and values.

The Sovereign AI Framework

The proposed Sovereign AI ecosystem is an integrated framework designed to maintain control across multiple layers of the network stack, particularly within the Open Radio Access Network (O-RAN) architecture.

The multi-layered Sovereign AI ecosystem for 6G networks.

As illustrated in the framework, sovereignty is achieved through several interconnected layers:

  • Governance & Policy Layer: Establishing national AI strategies, regulatory frameworks, and ethical guidelines.
  • Core Infrastructure Layer: Developing sovereign compute clouds (GPUs/TPUs) and secure data lakes for curated, anonymized national telecom data.
  • AI Development & Training: Leveraging a skilled workforce and research labs to build domain-specific models.
  • AI Deployment & Operations: Integrating sovereign AI-based "xApps" and "rApps" within RAN Intelligent Controllers (RICs) to power troubleshooting agents and telecom copilots.
  • Talent Development: Creating a new generation of AI-literate telecom engineers through specialized curricula and public-private partnerships.

Technical Enablers and Large Telecom Models (LTMs)

To realize this vision, several technical enablers are required. Federated Learning (FL) allows models to be trained across distributed infrastructure without moving sensitive raw data. Synthetic Data Generation (SDG) creates high-fidelity training data to bypass privacy restrictions, while Explainable AI (XAI) ensures that autonomous network decisions are auditable and understood by human engineers.

Central to this approach is the development of Large Telecom Models (LTMs). Unlike general-purpose LLMs like GPT-4, LTMs are fine-tuned on telecom-native semantics—such as mobility patterns and spectrum allocation—reducing the risk of "hallucinations" in critical infrastructure. This shift also signals the rise of Agentic AI, where intelligent agents autonomously formulate multi-step strategies to meet high-level network goals.

Global Landscape Analysis

The approach to AI sovereignty varies significantly across the globe, driven by different national priorities and regulatory environments.

Global landscape analysis of Sovereign AI strategies.

As shown in the analysis above, the EU focuses heavily on digital sovereignty and strict regulation (GDPR), while India prioritizes self-reliance and local language support through platforms like Bhashini. The US emphasizes national security and economic competition through chip control, and the UK focuses on domestic compute power (AIRR) and pro-innovation frameworks.

Challenges and Future Outlook

Despite its necessity, Sovereign AI faces significant hurdles. The capital and operational expenditures required are massive, potentially leading to higher consumer prices unless shared-infrastructure models like Neutral Host Networks are adopted. Furthermore, a "dependency paradox" exists: even sovereign initiatives often rely on hardware and software frameworks from a few global dominant players, creating a dependency at a different level.

Future work must address the delicate balance between local control and the need for global 6G standards, ensuring that sovereign AI "islands" can communicate securely without compromising the network's integrity. Ultimately, Sovereign AI is not merely a regulatory hurdle but a foundational pillar for the resilient, autonomous networks of the future.

Summary updated

BibTeX
@misc{chetty2025sovereign,
  author = {Swarna Bindu Chetty and David Grace and Simon Saunders and Paul Harris and Eirini Eleni Tsiropoulou and Tony Q. S. Quek and Hamed Ahmadi},
  title = {{Sovereign AI for 6G: Towards the Future of AI-Native Networks}},
  year = {2025},
  eprint = {2509.06700},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.06700},
}
Figure