The value of telecom AI becomes clear when its decisions can be evaluated against a working network. Teams need access to live measurements, meaningful control functions, repeatable tasks, and a way to inspect what an agent changes.
Built on MX-PDK AI-RAN, this solution combines two live O-RAN cells, telecom agents, developer tools, and 5G edge nodes. It gives AI researchers, telecom engineers, and application developers a shared environment to build, compare, and improve AI workflows using real network feedback.
Build or evaluate agent automation, multi-cell AI-for-RAN, or edge AI applications. MX-PDK AI-RAN brings together the live network and the development tools needed to study their interaction.
Value for your team
- Shorten the path to a useful experiment: begin with integrated network access and agent workflows.
- Make AI decisions reviewable: inspect the measurements, proposed actions, and observed results.
- Choose models on evidence: compare task quality, latency, and token usage using the same evaluation suite.
- Preserve engineering investment: reuse agents, applications, blueprints, and datasets as the platform expands.
One year of software updates and technical support is included with the software license. Training and custom integration can be scoped to the programme.
What this solution brings together
- A network with observable multi-cell behavior. Three O-Cloud nodes, two synchronized indoor O-RUs, a 5G Core, near-RT and non-RT RICs, and five Ubuntu-based Pictel 5G edge nodes support mobility, traffic, and application experiments. Emulated configurations extend experimentation to up to four gNBs and sixteen soft UEs; the included live radio footprint remains two cells.
- Reusable observability and automation. The included blueprints connect agents to network, infrastructure, log, and metric data, and to supported RIC, SMO, and API actions. Network-changing actions remain subject to explicit user approval.
- A development and evaluation toolkit. xApp and rApp SDKs and the Container Development Kit (CDK) share the same foundation as O-RAN. The included BAT Agent DevKit adds agent development, testing, packaging, benchmarking, and reuse. A2A supports agent coordination; MCP provides access to tools, data, and control APIs.
- Choice of models and execution environments. Configure supported local, on-premises, or remote model endpoints per agent. Model endpoints, subscriptions, and optional local GPU capacity are scoped separately to the actual workload.
Two complementary innovation tracks
| Track | What your team evaluates | Customer value |
|---|---|---|
| AI-for-RAN | Agents and applications that investigate, automate, or help optimize network behavior through telemetry and supported controls | Determine whether an AI workflow improves the operational task and the measured network outcome. |
| AI-on-RAN | AI and application workloads running alongside RAN infrastructure, including applications connected through the 5G edge nodes | Understand application performance, resource needs, and placement under realistic connectivity conditions. |
Agent workflows coordinate with the RIC and SMO. Their decision time is evaluated separately from radio control timing and end-to-end application latency.
More use cases you can unlock
The included network, agent blueprints, and developer tools provide a foundation for the following projects. Application-specific agents, policies, and integrations are developed or scoped for the use case.
| Related use case | What it helps you establish | Starting scope |
|---|---|---|
| AI-assisted service and slice assurance | Correlate application symptoms with network telemetry, then evaluate proposed policy changes against service targets. | Observability and Automation Blueprints are included; define the service metrics, permitted actions, and supported slice/QoS controls. |
| Traffic and interference analysis | Compare AI-assisted diagnosis or control with a baseline under changing load and radio conditions. | Two live cells and RIC access are included; models, algorithms, and supported control functions are selected for the experiment. |
| Connected robotics and video analytics | Measure how mobility, traffic demand, and compute placement affect an AI-enabled application. | Use the five included 5G edge nodes. Cameras, robots, application software, and any required local GPU are additional project scope. |
| TN/NTN agent evaluation | Test whether an agent responds usefully to the conditions represented by supported terrestrial or NTN scenarios. | Supported TN/NTN experimentation and emulated capacity up to four gNBs and sixteen soft UEs; the topology must fit the selected profile. Live NTN requires separate radio and integration scoping. |
| Energy-aware automation | Explore policies that balance application performance, infrastructure utilization, and measured node power. | Add optional energy visibility/PDU control; develop and evaluate policies with the included automation tools. Digital Twin validation requires an NDT extension or CAMPUS. |
From AI idea to validated behavior
- Choose a bounded task. Define the operational request, allowed tools, network actions, baseline, and success criteria.
- Observe and propose. Let the workflow gather evidence and prepare its action plan. Inspect the supporting measurements, tool results, and proposed changes.
- Approve and measure. Apply approved actions through supported interfaces and compare the resulting network and application behavior with the baseline.
- Benchmark and reuse. Repeat under different traffic conditions, models, or tools. Retain the task suite, agent configuration, results, and datasets for future evaluation.
For a load-balancing experiment, assess per-cell load, per-UE throughput, handover outcomes, application continuity, and the agent’s task completion time. Model response quality alone is insufficient to establish network value.
FAQs
1️⃣ Is a local GPU included?
A local GPU is optional in MX-PDK AI-RAN. Compute requirements depend on the models and applications you choose. Model endpoints, subscriptions, and optional GPU resources are scoped separately. The overall configuration is quoted to your requirements.2️⃣ Does the package include Digital Twin validation?
It includes the Observability and Automation Blueprints. Digital Twin-backed optimization requires the Network Digital Twin capability available through MX-PDK CAMPUS and MX-DT.3️⃣ Can we bring our own models, agents, and applications?
Yes. The platform supports configurable model endpoints, custom agents, MCP tools, API adapters, xApps, and rApps. Compatibility, compute requirements, and the permitted actions are defined for the selected workflow.4️⃣ Can this be extended beyond the indoor lab?
Additional compute and outdoor O-RUs can be scoped for a larger trial. MX-PDK CAMPUS provides the next configuration for site coverage, enterprise infrastructure, GPU capacity, and Digital Twin workflows.Ready to put your telecom AI to the test?
Share one operational task or application, the models you want to evaluate, your data-hosting requirements, and the outcomes that matter. BubbleRAN can define a live demonstration and an evaluation configuration around them.
Find the right AI-RAN configuration · Explore MX-PDK AI-RAN · Explore the TelcoFabric ecosystem.