AI-Native 6G: From Network to Intelligent Fabric

6G aims not for faster communication but for AI-native intelligent infrastructure. The goal is a network that embeds intelligence at every layer from device to cloud, understanding user intent, optimizing itself, and verifying security and reliability.

While previous generations expanded speed and capacity, 6G presupposes AI fabrics that combine communication and computing. Operating this requires real-time optimization, workload distribution across multiple systems and regions, and governance that satisfies operational, regulatory, and safety requirements.

The core is Network Language Models(NLMs). NLM extends the existing LLM concept into the network domain, handling high-frequency telemetry time series, topology graphs, event sequences such as alarms and state transitions, and structured data such as policies and parameters together.

NLM development proceeds in stages.

• Make general-purpose LLMs smaller and more efficient using compression techniques such as pruning, quantization, and knowledge distillation.
• Fine-tune on network-related corpora to support configuration generation, log analysis, alarm correlation, and troubleshooting.
• Extend into a multimodal architecture combining a time encoder, graph encoder, and structured data encoder.
• Add reinforcement learning and policy constraints to ensure safety and executability, and use federated learning or distributed training to protect data sovereignty and confidentiality.
• Formal verification based on automated reasoning is also needed to ensure reliability.

An NLM matured in this way accumulates deeper domain knowledge about network data. For example, it learns BGP convergence patterns, 3GPP signaling flows, the impact of configuration changes on state, and interdependencies across RAN-core-transport.

The network intelligence fabric proposed by AWS is a distributed reasoning system that combines this with RAG, a graph DB holding real-time topology state, a knowledge graph holding protocol semantics, and an agent-to-agent communication framework. The goal is to enable optimization across multiple domains and operators while maintaining guardrails through a policy enforcement layer.

The deployment roadmap is presented in 4 stages.

• Stage 1: Implement closed-loop automation on top of EMSes through digital twins.
• Stage 2: Open cross-domain control with standardized interfaces, turning it into a programmable system.
• Stage 3: Introduce federated NLM and autonomous agents to conduct multi-operator collaboration within governance boundaries.
• Stage 4: Achieve fully autonomous resource orchestration through context-aware reasoning, dynamic service discovery, and flexible agent federation.

The ultimate goal is hyper-composed networks. This dynamically combines compute, storage, networking, data, and AI resources according to purpose, time, and location, and requires 10 principles: model-driven abstraction, sense-discern-infer-decide-act control, context reasoning, collaborative intelligence, fluid agent associations, dynamic discovery, adaptive protocol evolution, multidomain federation, repeatable patterns, and fractal emergence.