A modern network should not simply use AI to make predictions. It should be able to understand what is happening in the network, make decisions, take actions, and verify whether those actions have improved the network.
AI-native RAN depends on more than the quality of an AI model. A useful decision must connect four essentiel elements: Measurements, which provide information about current network conditions, such as traffic load, signal strength and latency; Supported Controls, which define the network parameters and mechanisms that AI can modify to influence network behavior; Execution Constraints, which represent the rules, requirements and operational limitations that AI must respect when making decisions; and Evidence of Effect, which consists of feedback and performance measurements used to verify whether the actions taken by AI have achieved the desired improvements, such as reducing congestion or improving network performance.
Each paradigm adds a reusable foundation for network intelligence. These paradigms describe cumulative architectural capabilities. They coexist within a network, with different levels of integration and operational maturity.
| # | Paradigm | Capability | Narrative |
|---|---|---|---|
| 1 | Software-defined | Flexibility | Change network behavior through software. |
| 2 | Cloud-Native | Elasticity and lifecycle automation | On-demand and Self-Service network at scale. |
| 3 | Open & Programmable | Interoperability, Observability, and Control | Expose capabilities. Observe and Steer behavior. |
| 4 | AI-Native | Sensing, learning, and reasoning | Understand context. Learn and adapt. |
| 5 | Autonomous | Intent-driven coordination & actions | Pursue goal within constraints. |
| 6 | Self-Synthesizing (R&D) | Capability generation & composition | Create capabilities where and when needed. |
1. Software-defined: behavior becomes changeable
Software makes network functions easier to develop, configure, and extend. Engineers can experiment with protocol implementations, scheduling policies, slicing, and control logic while retaining the constraints imposed by the underlying hardware. Open CU/DU implementations strengthen this foundation. Access to the central and distributed unit software allows developers to inspect behavior, add measurements, and expose new controls. Duranta/OpenAirInterface and OCUDU contribute to this open software ecosystem. The FlexRAN research illustrates the connection between accessible implementation and programmable RAN behavior.
Reusable foundation: Software execution platforms and compute resources build the reusable foundation.
What it makes possible? Software execution platforms and compute resources provide the foundation on which network functions can be developed, modified, and extended. This creates a flexible execution foundation that cloud-native technologies can subsequently manage, scale, and operate consistently across deployments.
2. Cloud-Native: operation becomes repeatable
Cloud-native design turns deployment, configuration, scaling, recovery, and upgrades into managed lifecycle operations. Declarative orchestration lets teams describe the required state and use common mechanisms to establish and maintain it. This enables on-demand, self-service networking within available capacity and supported configurations. RAN workloads still require careful placement, synchronization, acceleration, and timing. Automation makes those requirements part of a repeatable deployment process.
Reusable foundation: a cloud runtime and lifecycle orchestration.
What it makes possible? A cloud runtime and lifecycle orchestration provide a shared operational environment for deploying, managing, scaling, and maintaining network functions. Applications and controllers can build on a shared operational environment instead of recreating deployment procedures for every use case.
3. Open & Programmable: capabilities become accessible
Open interfaces expose what a network can observe and change. Telemetry supplies evidence; control interfaces provide supported actions; application environments host logic that uses both. SMO and RIC architectures connect these capabilities to management and control applications. Open source allows implementation changes, while interfaces allow integration across components. Interoperability requires agreement on identifiers, units, timing, and control semantics.
Reusable foundation: control interfaces, telemetry, and application hosting.
What it makes possible? Control interfaces, telemetry, and application hosting transform network capabilities into accessible resources that different applications can discover, observe, and use. This creates the interaction layer that allows AI systems to connect intelligence with real network observations and actions on which to base decisions.
4. AI-Native: context informs adaptation
An AI-native architecture connects measurements and operational context to learning, reasoning, and action. Models can recognize patterns or estimate outcomes. Agents can interpret requests, retrieve information, select tools, and coordinate workflows. Intelligence can contribute across radio, cloud, and service management. Its placement must respect execution timescales: management-layer reasoning can guide operations while local controllers handle decisions with tight radio deadlines.
Reusable foundation: context, models, tools, and agent services.
What it makes possible? Context, models, tools, and agent services provide the building blocks for reusable network intelligence. A capability developed for one workflow can support others when its assumptions and dependencies remain explicit.
5. Autonomous: goals coordinate actions
Autonomy connects individual decisions around a desired outcome. The operator defines a goal and constraints; the system observes conditions, chooses actions, checks their effects, and adapts within its authorized scope. Coordination matters when objectives compete. Energy reduction, capacity, and service assurance may demand different actions on the same resources. An autonomous workflow must reconcile those demands and recognize when the goal is infeasible.
Reusable foundation: goals, policies, and executable blueprints.
What it makes possible? Goals, policies, and executable blueprints capture operational objectives and the knowledge required to achieve them. This makes autonomy a repeatable capability that can be applied across different network scenarios while remaining within operator-defined constraints.
6. Self-Synthesizing: capabilities can evolve
Self-synthesizing networks explore what happens when existing actions cannot satisfy a new requirement. The system first tries to compose available capabilities. A confirmed gap may justify generating an adapter, workflow, or application, then validating it before registration and deployment. Generation cannot resolve insufficient physical resources or contradictory goals. The system must distinguish those limits from a missing capability.
Reusable foundation: a capability catalog and a generation–validation–deployment pipeline.
What it makes possible? A foundation for extending the network beyond its existing capabilities. This remains an R&D direction, extending reuse toward controlled capability evolution under externally defined goals and constraints.
An ecosystem built on reusable foundations
The evolution towards AI-native and autonomous RAN is not driven by a single technology or company. It requires an ecosystem where developers, technology providers, and operators contribute complementary capabilities. Software functions, cloud environments, open interfaces, AI models, agents, and operational workflows become more valuable when they can be shared, combined, and reused across different network scenarios.
This is where TelcoFabric, telco artifact portal led by BubbleRAN, comes into play. It provides an ecosystem channel for discovering and sharing applications, models, agents, and blueprints, helping turn individual contributions into building blocks for future network intelligence.
How BubbleRAN connects the paradigms
The six paradigms are not isolated technology trends, but complementary steps towards a more flexible, intelligent, and adaptive network. BubbleRAN brings these capabilities together through an evolving ecosystem of products, platforms, software, and intelligent services. At BubbleRAN, we are building this vision through two complementary product families: MX-PDK and Sphere.
MX-PDK provides integrated environments for developing, experimenting with, and deploying open and programmable RAN solutions, from laboratory setups to campus-scale deployments. Across its different editions, it brings together software-defined RAN, cloud-native infrastructure, open interfaces, and AI-RAN capabilities. Sphere product range extends these foundations towards intelligent network operations. With SMO-Sphere for orchestration and lifecycle management, RIC-Sphere for programmable RAN control, and Opti-Sphere for AI-driven operational intelligence and optimization, the Sphere portfolio connects network capabilities with increasingly intelligent and automated operations.
Together, these products provide a practical path from open and programmable RAN towards AI-native and autonomous network architectures. They also provide a foundation for exploring emerging research directions such as self-synthesizing networks.
BubbleRAN combines these technologies in the MX-PDK family and Sphere products. MX-PDK packages integrated environments from LAB to CAMPUS; Sphere products provide orchestration, programmable control, and operational optimization. The same technologies support several paradigms.
| Paradigm | BubbleRAN technologies | Related BubbleRAN products |
|---|---|---|
| Software-defined | 5G SA/NSA Rel. 16/17; FR1 TN/NTN and FR2 TN, according to the supported stack and radio configuration. | MX-PDK range |
| Cloud-Native | Telco-optimized O-Cloud, SMO/OAM from Day 0 to Day 2+, declarative lifecycle orchestration, CDK, and on-premises GPU resources where configured. | MX-PDK range; SMO-Sphere |
| Open & Programmable | Near-RT and Non-RT RIC, xApps/rApps and SDKs, E2/A1/O1/R1 interfaces, TELCOFabric catalog, O-RAN 7.2 fronthaul and PTP synchronization; dashboards, CLI, and APIs provide access to operations. | MX-PDK range, RIC-Sphere and SMO-Sphere |
| AI-Native | AI controller and agent services, observability/automation agents, BubbleRAN Agentic Toolkit BAT ADK/CLI and catalog, A2A/MCP, NVIDIA NIM and NeMo Agent Toolkit (NAT), and chatbot access. | MX-PDK AI-RAN and MX-PDK CAMPUS, Opti-Sphere |
| Autonomous | Data-driven optimization, Network Digital Twin (NDT), network replicas, RAN mirroring, channel models, NDT Scenario Kit, scenario generation, synthetic data, what-if validation, and bounded control with outcome monitoring; PDU metering/control supports energy workflows. | MX-PDK CAMPUS, Opti-Sphere, RIC-Sphere and SMO-Sphere |
| Self-Synthesizing (R&D) | SSN / Symmorphosis: capability catalogs, composition, and controlled generation–validation–deployment pipelines. | MX-PDK range |
From reusable capabilities to measurable improvements
Consider a practical example: users experience degraded uplink performance in a 5G cell. Network software provides radio measurements and supported controls, while cloud infrastructure supplies the execution environment. Open interfaces make these capabilities accessible to AI agents, which correlate network information, investigate potential causes, and evaluate possible improvements. A reusable workflow coordinates diagnosis and execution while respecting operational constraints.
The loop closes by checking new network measurements against the original performance objective. Executing an action does not necessarily mean achieving the desired result; only observing network behavior can demonstrate an actual improvement. The same workflow can potentially be reused in other deployments, provided its configurations, interfaces, and effects are validated. An example of this direction is MX-AI’s published research, connecting SMO-layer agents to a live Duranta/OpenAirInterface and FlexRIC testbed. Ultimately, reusable foundations turn individual technological capabilities into shared intelligence that can be applied, evaluated, and extended across network deployments.