Every mobile call, video stream and chatbot query that reaches a phone passes through a radio access network, or RAN: the base stations, antennas and software that talk to devices over the air. For decades the intelligence in that network was baked into each vendor’s equipment. Open RAN changes the arrangement by defining standard interfaces and adding a programmable layer, the RAN Intelligent Controller, where operators can run machine-learning models written by anyone. This explainer covers how that layer works, what the AI programs on it do, and how the idea connects to 3GPP standards, the AI-RAN movement and the 6G framework.
Why the radio network needs a control layer
A cell site makes thousands of small decisions every second: which user gets which slice of spectrum, which cell a moving phone should hand over to, when a lightly used carrier can sleep to save power. Traditional networks make these decisions with fixed algorithms tuned by engineers, often per vendor, and adjusted through occasional drive tests and configuration changes.
Machine learning is attractive here because conditions vary by hour, weather, event and neighborhood. A model that learns patterns in traffic and interference can adapt faster than a hand-tuned rule. The obstacle is access. If the control logic is locked inside one vendor’s product, an operator cannot add a third party’s optimization without that vendor’s cooperation. Open RAN, defined through specifications from the O-RAN ALLIANCE, is meant to open that door through standardized interfaces between radio, baseband and management components.
The two control loops
O-RAN’s architecture places intelligence in two controllers that differ mainly in timescale.
The Non-Real-Time RIC sits in the service management and orchestration layer. Its control loops run on a timescale of seconds or longer. It is where models are trained and updated using data collected over hours or days, and where policies are set, for example “prioritize energy savings in this district overnight.”
The Near-Real-Time RIC sits closer to the radio and works on a loop of roughly ten milliseconds to one second. It runs inference and makes decisions that adjust live behavior, such as steering traffic between cells or tuning resource allocation for groups of users.
The two connect through defined interfaces. The A1 interface carries policies and guidance from the non-real-time layer down to the near-real-time one. The E2 interface lets the near-real-time controller collect measurements from and send control to the distributed radio units. The O1 interface handles management and data collection. We summarize these from the O-RAN architecture; the exact definitions and current versions are in the specifications published by the O-RAN ALLIANCE, which releases updates in batches.
xApps and rApps: the AI programs
Applications on these controllers have names that reflect where they run. An rApp runs on the non-real-time controller. An xApp runs on the near-real-time controller. Because the interfaces are standard, in principle an operator can buy a radio from one company, a controller from another and an optimization app from a third.
Typical examples include:
- Traffic steering: moving users between frequency bands or cells to balance load.
- Energy saving: predicting low-traffic periods and putting carriers or radio components into sleep modes, then waking them before demand returns.
- Anomaly detection: spotting a cell whose performance indicators drift from their normal pattern, which can reveal faults before customers complain. The same logic appears in our explainer on predictive maintenance.
- Interference and coverage optimization: adjusting antenna parameters based on measurements from many cells at once.
The models behind these range from classical forecasting and gradient-boosted trees to reinforcement learning, in which an agent learns a control policy by trial and reward. Our glossary entry on reinforcement learning explains the basic idea. In a live network, trial and error is risky, which is why training and testing often happen in simulation or on logged data in the non-real-time layer before a policy is deployed.
Guardrails: conflicts and trust
Opening the control layer creates new problems. Two apps from different vendors might both try to change the same parameter in opposite directions, so the architecture needs mechanisms to detect and resolve conflicts. A model that performs well on last month’s traffic can degrade after a stadium event or a software update, so operators need monitoring, rollback and clear ownership when an app misbehaves. Security also matters: a programmable controller with access to the live network is a sensitive target, and the interfaces need authentication and isolation. These concerns, not the algorithms themselves, are a large part of what operators weigh when they decide how far to open their networks.
How standards and AI-RAN fit in
The O-RAN ALLIANCE defines the open architecture, but the radio air interface itself is specified by 3GPP. 3GPP’s Release 18 work program lists artificial intelligence and machine learning among its themes, and its Release 19 work continues in that direction. Within the standards, AI/ML is studied both for network management and for the air interface itself, for example in how devices and base stations exchange channel information. The 3GPP pages list the work items; we have not reviewed the individual specifications.
A separate movement, the AI-RAN Alliance, argues that radio processing and AI workloads should share the same computing hardware at the cell site or in regional data centers, so that capacity sold for radio during busy hours can run AI inference during quiet ones. Whether that economic model works depends on utilization numbers that operators have not broadly published, so treat it as a proposal rather than a settled practice.
Where 6G comes in
In November 2023 the International Telecommunication Union approved Recommendation ITU-R M.2160, the framework for what it calls IMT-2030, the family of technologies commonly referred to as 6G. The framework is a statement of goals and usage scenarios rather than a technical specification, and detailed standards are being developed through 3GPP. AI is widely expected to be built into 6G designs from the start, instead of added later as it was in 5G. Commercial 6G is not expected to arrive for several years, so for now the practical action is in 5G networks and Open RAN deployments.
What to watch next
- New O-RAN specification releases and whether they add features for AI model management.
- 3GPP progress on AI/ML in the air interface and in network management.
- Whether operators publish measured energy savings from rApps and xApps in live networks.
- How the AI-RAN proposal to share hardware between radio and AI workloads fares in trials.
- Early 6G study items that treat AI as a native part of the design.
Key takeaways
- The RAN Intelligent Controller is a programmable layer that lets operators run third-party optimization software on radio networks.
- The near-real-time controller acts in roughly 10 milliseconds to one second; the non-real-time controller works over seconds or longer.
- xApps and rApps are the applications; common uses are traffic steering, energy saving and anomaly detection.
- Conflicts between apps, model drift and security are the main practical risks of opening the control layer.
- 3GPP defines the air interface, ITU sets the 6G framework, and Open RAN is about the architecture around them.
Frequently asked questions
What is the difference between an xApp and an rApp?
An xApp runs on the near-real-time controller and reacts quickly; an rApp runs on the non-real-time controller and works on slower policies and model training.
Is Open RAN the same as AI-RAN?
No. Open RAN is about open interfaces and multi-vendor networks. AI-RAN is a separate effort to run AI and radio workloads on shared computing infrastructure.
Will 6G be built on AI?
The ITU framework for 2030 and beyond points toward AI playing a larger role, but detailed 6G standards are still being developed, so specifics remain open.
