Introduction
Open RAN has generated considerable excitement, promising a true transformation in network flexibility and openness. But has it truly lived up to the hype? While Open RAN has laid the groundwork, the reality is that managing these complex, disaggregated networks is proving to be a significant challenge. Traditional methods simply can't keep pace with the dynamic demands of 5G and beyond. The answer lies in intelligent automation.
AI-RAN, by embedding artificial intelligence and machine learning directly into the RAN, offers a fundamental shift in how we approach network management. This shift is further amplified by the emergence of efficient and application-focused models like DeepSeek and Qwen, which are reshaping the AI landscape in telecom. These advancements, coupled with the growing importance of hyperscaler partnerships and data sharing, are opening up new possibilities for AI deployment within the RAN. This article explores how AI-RAN can revolutionize key RAN functions, from intelligent beamforming to optimized energy consumption. We'll examine how these next-generation LLMs can be strategically integrated into the intelligent RAN architecture, leveraging hyperscaler data and specialized AI/ML services. Our focus will be on three critical use cases: Massive MIMO Beam Management, Coverage and Capacity Optimization, and End-User Energy Efficiency.
Furthermore, we'll discuss the crucial role of integrating AI-RAN with existing OSS/BSS systems, ensuring robust multi-domain service assurance, and strategically exposing network capabilities through standardized APIs.
AI-RAN: A Paradigm Shift in Radio Access Networks
Open RAN has laid the foundation for a more flexible and open RAN architecture. However, the increasing complexity of modern networks and the relentless demand for real-time optimization necessitate a more intelligent approach. AI-RAN addresses this need by embedding AI/ML directly into the RAN, enabling dynamic resource allocation, automated network management, and personalized user experiences. This article details how AI-RAN can significantly improve the performance and efficiency of key RAN functions, exploring the strategic placement of LLMs within the RIC and the crucial role of lightweight AI/ML models for dApps at the DU level.
AI-RAN: Unleashing the Power of Intelligence
The true power of AI/ML in the RAN extends beyond just large models. For truly real-time, network-specific functions, lightweight, specialized AI/ML models deployed as dApps are essential. These models, designed for efficiency and speed, can handle tasks where the computational overhead and latency of LLMs would be prohibitive. This combination of powerful LLMs for higher-level RIC functions and agile, lightweight models for distributed tasks creates a robust and adaptable AI-RAN architecture.
First Use Case: Massive MIMO Beam Management
a) Current Open-RAN Limitation
Traditional beamforming relies on pre-defined algorithms and limited real-time feedback, posing challenges for Open RAN. The limited real-time capabilities of Open RAN make it difficult to adapt to dynamic user distributions and varying interference conditions.
b) AI-RAN Improvements:
The current RIC approach, with its Nr-RT and Non-RT components, faces challenges like complexity, limited AI/ML integration, interoperability issues, scalability concerns, and security vulnerabilities. AI-RAN directly addresses these limitations by leveraging the latest advancements in AI/ML, built with AI-native considerations.
AI-RAN does :
Improve network performance through native AI/ML-driven analysis of network data and real-time decision-making.
Enhance user experience by leveraging next-gen LLMs to personalize network parameters.
Simplify network management by automating tasks, optimizing operational costs
Finally, enable/accelerate new services creation like network slicing in collaboration with OSS/BSS and Network Exposure layers.
c) Open Interfaces and 5G Requirements:
3GPP provides a framework of APIs and interface references, enabling AI-RAN implementation. Open APIs (3GPP, TM Forum, O-RAN Alliance, Camara, IETF, etc.) enable AI-RAN interaction with OSS/BSS and Data Lakes for data retrieval, configuration updates, and policy enforcement. Within Open RAN, the O1 interface is used for performance data collection, the A1 interface for policy control, and the O2 interface for control and adjustment of O-Cloud lower layers. 5G requirements emphasize eMBB for high throughput and URLLC for sub-4ms latency.
d) Collaboration with Hyperscalers for Data Lakes:
Distributed data lakes, instead of a single central repository, offer advantages like multi-cloud and edge computing strategies. This enables the development of tailored AI/ML analytics for Open RAN with AI-RAN. Data retrieval, utilizing cloud-native principles like Kafka or Amazon Kinesis, can draw from telecom OSS or hyperscaler data lakes, ensuring real-time optimization and tangible end-to-end network performance improvements.
Distributing the data also cuts down on latency for AI/ML processing at the edge, which is absolutely critical for making real-time beam adjustments. Plus, it boosts resilience – if one data lake hiccups, the others can pick up the slack. In fact, a multi-hyperscaler collaboration strategy could be key to avoiding lock-in and ensuring access to the best data and services from various providers.
Real-time optimization is absolutely essential for Massive MIMO beam management. User locations and channel conditions can change in the blink of an eye, especially with mobile users. AI-RAN, by tapping into these hyperscaler data lakes and doing AI inference at the edge, can adapt to these changes dynamically. The result? Optimal beamforming, minimal interference, and a superior user experience, especially for very demanding real-time applications like AR/VR and streaming of video or gaming contents.
e) DeepSeek/Qwen and the Changing Landscape:
Both DeepSeek & Qwen represent a real shift in the AI landscape and especially for AI-RAN :
First, Reduced footprint where these models are natively designed to be more efficient, leaner in computational in power and memory for both training and inference. This makes them more efficient suitable for distributed deployments and especially in resource-constrained environments such as Far Edge or Hyperscaler edge.
DeepSeek and Qwen represent a significant leap forward in the AI landscape, particularly for AI-RAN. Their reduced footprint is a game-changer. It means they're far more efficient in terms of computational power and memory usage for both training and inference. This is absolutely critical for distributed deployments, especially in resource-constrained environments like the far edge or even within the hyperscaler's edge infrastructure. Their improved performance offers better accuracy and faster processing compared to earlier, more cumbersome LLMs or general-purpose AI models. This efficiency and performance boost makes them well-suited for integration within the RIC architecture.
For instance, alongside the Non-RT (Non Real Time) RIC, DeepSeek or Qwen can play a role in:
Training and Optimization : These LLMs could eventually be used to train and refine the AI/ML models used for real-time beam management. Their apparent ability to process massive datasets could be relevant to analyzing historical network data, user mobility information and environmental factors that could affect beamforming strategies.
Features definition : Next-Gen LLMs can assist in identifying and extracting the most relevant features from the vast amount of available data. Just an assumption...
Policy Optimization : These next-gen LLMs can help developing and optimizing policies that define beamforming decisions. For instance, they can assist in defining rules for balancing strength of signal, interference mitigation and energy efficiency as explained in the third use case.
Now, at the (Near Real-Time) NR-RT RIC, where decisions are made quickly, the next-Gen LLMs could be deployed for:
Analyzing real-time network conditions and user context to make more informed beamforming decisions. This contextual awareness could ne user location, application type and interference levels, etc..
Anomaly detection, LLMs can be used to fast-detect anomalies in network behavior that indicate a need for beam adjustments where a drop in signal quality might change beamforming vectors.
Predictive beam steering, this one is very exciting, by leveraging user mobility patterns and data gathered from the converged datalakes. LLMs can help predict user movements and proactively adjust beams to ensure connectivity consistency.
Regardless of how powerful and learning efficiency that they provide, these next-Gen LLMs are, it is important to consider that the most sensitive tasks at the DU level require highly specialized, lightweight AI/ML models deployed as dApps is still essential. Such dApps (Decentralized Applications) will be optimized for specific beamforming algorithms and would be running in the Dataplane, which might offer extreme low-latency processing compatible with next-generation 5G services such as URLLC for true Real-Time requirements. These dApps can scale both vertically and horizontally and offer flexible inclusion in CI/CD/CT management. Additionally, they ensure security and trust through transparent execution and API exposure for consumption.
This hybrid approach of Next-Gen LLMs and dApps combines the best of both worlds : The power of LLMs with the agility and specialized AI/ML at the edge. Impacting directly training costs for edge models, pushing for the bulk of intensive training centrally while edge models can be fine-tuned with small datasets. Reduce training costs of LLMs by using pre-trained LLMs and fine-tuning for specific RAN tasks that avoids huge training cost.
Now, the impact on inference management could also be interesting, in my view: Lower inference costs at edge as these models are smaller and more specialized then require less computational power and memory for execution. Faster inference at the edge as specialized models are designed for speedy execution. They can indeed perform inference much faster than general purpose approaches.
Second Use Case: Coverage & Capacity Optimization
a) Current O-RAN Limitations:
Optimizing coverage and capacity often involves manual tuning and reactive adjustments, inefficient and slow to adapt to changing 5G service demands.
b) AI-RAN Improvements:
AI (xApps/rApps or DU dApps) enables dynamic resource allocation by analyzing traffic patterns, user density, and application requirements (FWA, eMBB, URLLC) to allocate radio resources (frequency bands, time slots). AI-RAN improves interference management by learning and anticipating patterns, optimizing resource allocation. It enhances self-organizing networks (SON) by automating network optimization tasks (cell planning, parameter tuning, fault management), improving operational costs and network performance. Option 2 split is preferred here.
c) Open Interfaces and 5G Requirements:
O1, A1, and O2 interfaces are crucial for policy control, AI-driven optimization, and control over O-Cloud lower layers. AI-RAN is crucial for FWA (SME focus) and eMBB, maximizing coverage and throughput by dynamically adjusting resources.
d) Hyperscaler Collaboration:
Hyperscaler’s datalakes can provide crucial data information such as location, demographic information and other relevant data or already crunched KPIs for optimizing coverage and capacity, such as :
Location Data : Population density, foot traffic patterns, this could allow AI-RAN to determine how how coverage should be prioritized.
Demographic data : Age, Income, tech usage can help predict service usage patterns and to tailor capacity planning accordingly
Application Usage Data : Appreciating which applications are popular and how they are used allows AI-RAN to make pinpointed decisions to allocate resources for specific service approaches.
Mobility patterns : Hyperscalers can provide valuable information about how people move. Which would help predicting traffic locations and optimizing handovers between mobile cells.
It is rather important to avoid the risk of being locked-in when engaged with a single hyperscaler and this on multiple accounts Flexibility, Cost-optimization, Resilience on a per target services and innovation as not all hyperscalers move at the same pace. Therefore, to approach a multi-cloud collaboration, several options can be entertained: Hybrid-Cloud, Multi-Cloud Management Platform through OSS, Edge Orchestration capabilities and Network API exposure. Open APIs and industry standards would be crucial to ensure interoperability between different cloud platforms notably for API Aggregation/Marketplaces and also for API Brokerage.
A multi-cloud approach to hyperscaler collaboration offers several advantages. It avoids vendor lock-in, provides access to a wider range of data and services, and allows operators to choose the best cloud platforms for specific tasks. However, managing data across multiple clouds can be complex. This is where the concept of converging data lakes becomes crucial.
A converging data lake isn't necessarily a single physical entity but rather a logical framework that allows AI-RAN to access and integrate data from various hyperscaler clouds and on-premise data stores. This framework should:
Provide a unified view of the data: Regardless of where the data resides, AI-RAN should be able to query and access it through a consistent interface.
Handle data transformation and normalization: Data from different sources may be in different formats. The converging data lake framework should be able to transform and normalize the data so that it can be used effectively by AI/ML models.
Ensure data governance and security: Data privacy, access control, and compliance with regulations are paramount, especially when dealing with sensitive user information.
3GPP’s NWDAF (Network Data Analytics Function) might also play a key role in AI RAN. However, a distributed approach for NWDAF would prove more interesting than a centralized one with significant benefits for predictive analytics about the 5G core therefore important for the RAN:
Reduced latency for real-time analysis of network data that could reveal crucial for many AI-RAN functions such as dynamic resource allocation and interference management.
Improved scalability: A distributed NWDAF can handle the massive amounts of data more efficiently than a centralized NWDAF
In my view, the power of Cloud & Hyperscaler data, multi-cloud collaboration, converged data lakes and edge-based NWDAF NFs could lead to better network performance, reduced optional costs and improved user experiences. And for Telcos which are leveraging NaaS, this approach could be applicable too.
e) DeepSeek/Qwen and the Changing Landscape:
Same as previous previously addressed in the first use case, DeepSeek and Qwen, with their efficient architecture and focus on practical applications, offer significant potential for enhancing Coverage & Capacity Optimization in AI-RAN. Their reduced footprint and improved performance make them suitable for deployment within the RIC architecture, both Non-RT and Near-RT.
In the Non-RT (Non Real-Time) RIC, next-Gen LLMs can help in :
Network planning and Optimization: DS & Qwen can analyze massive datasets of historical network performance data, user traffic habits, and external factors as potentially the demographic factors but also weather patterns where and if applicable (like our crazy snow storm this weekend here in Eastern Canada).
Establish capacity planning, where LLMs can forecast future capacity demands based on historical trend data, projected growth/decline and anticipated service usage based on converged datalakes.
Parameter tuning: Similarly to SON previously, LLMs can assist in automatically tuning network parameters to optimize coverage and capacity.
Policy Optimization: These next-Gen LLMs can help define and optimize policies for resource allocation and interference management. For instance, service priorities, QoS requirements and cost constraints could be enhanced to maximize network efficiency and operational costs.
In the Nr-RT (Near Real-Time) RIC (This is very exciting !!!), next-Gen LLMs can be used for:
Dynamic Resource Allocation, by leveraging real-time traffic patterns, user locations, and application requirements then dynamically allocate radio resources such as frequency bands and time slots across cells and users then avoid overcrowded radio cells and harmonious distribution of resources.
Interference management, LLMs could analyze real-time interference levels and adjust network parameters to improve the quality of signal.
Handover optimization, mobility is quite important and these LLMs might anticipate user mobility patterns and real-time network conditions to optimize or force handover decisions for better user experiences.
Faster anomaly detection, LLMs could indeed spot anomalies in network behavior that might indicate a capacity issue or bottleneck/coverage issue. A particular increase of dropped sessions for a given cell could trigger the LLM to investigate an to recommend corrective actions.
I don’t want to repeat myself as per the first use case but LLMs provide higher-level intelligence and optimization, specialized, lightweight AI/ML models deployed as dApps at the DU level are still essential for the most time-sensitive tasks. The LLMs in the RIC could work in concert with these dApps, providing the strategic guidance and optimization while the dApps handle the immediate, low-level control actions needed for real-time coverage and capacity management."
Third Use Case: End-User Energy Efficiency
a) Current O-RAN Limitations:
Traditional power-saving mechanisms are static, not adapting to individual user behavior and application needs.
b) AI-RAN Improvements:
AI-RAN enables personalized power saving by learning user patterns and dynamically adjusting power-saving parameters. It offers application-aware power management by analyzing traffic characteristics and optimizing consumption. It also improves network-level energy efficiency by optimizing cell sleeping and other mechanisms based on traffic and user activity.
c) Open Interfaces:
The O1 interface provides data on user activity and device characteristics, the A1 interface enforces power-saving policies, and the O2 interface retrieves O-Cloud KPIs for end-to-end service energy consumption. A more centralized split is suitable here.
d) Hyperscaler Collaboration:
For End-User Energy Efficiency, hyperscaler collaboration can provide significant insights into user behavior and application usage, which could benefit AI-RAN to fluctuate power consumption, data such as:
Application usage patterns: Which apps are used and how often ? and How data is averagely consumed ?
Device characteristics : battery capacity, processing power, screen and resolution size.
User Mobility patterns : Where do users typically roam during the day ? Do they generally in good or bad reception areas ?
Using such data, AI-RAN can establish power-saving strategies, optimize application-oriented power consumption, predict and adapt to user needs and optimize network energy consumption and efficiency.
As discussed before, a multi-cloud approach provides flexibility and avoids vendor lock-in. This simply means that data and services provided by multiple cloud providers can help in establishing specialized analytic KPIs on distinct models and services that are strategic to your respective B2B customers. Therefore, a converging Datalake framework is crucial to provide a unified view of data coming from different sources to better handle data transformation and ensuring data governance and security.
In terms of standards, 3GPP specification like TS 28.554 & 28.552 define both the interfaces and data models for network data analytics. These can certainly for ensuring interoperability between AI-RAN components and the converging Datalakes. More importantly, it will ensure interoperability of data sharing between partners of a MOCN/MORAN strategy.
Similarly for the previous use case, Edge NWDAF could play a crucial role in End-User Energy Efficiency, notably for:
Enabling real-time reporting of energy consumption of the 5G core services including E2E network slices or network slice subnet segments
Improving predictive analytics on energy consumption, forecasting future energy demands and pinpointing potential energy consumption.
Identifying Network slices energy efficiency, as edge NWDAF can provide slice-energy specific reports to ensure that network slices are operated efficiently and sustainably.
At the end, such an orchestra of capabilities could help portray the energy footprint of given service infrastructure that is not only connected to its vertical benefits but could apply on an end-to-end perspective.
E)Next-Gen LLMs or dapps at du levels ?
Unless I don’t see something that you guys, I can’t find any ways that LLMs like DeepSeek and Qwen could benefit the real-time aspects of End-User Energy Efficiency at the DU level. As we've discussed, their strength lies in higher-level reasoning, optimization, and planning, not necessarily the millisecond-level decisions often required for power management at the edge. However, that doesn't mean LLMs are completely out of the picture for energy efficiency. But, I still need to search for some clear value but I see some options in User Behavior modeling, Energy-aware application prioritization but also for reporting and analysis in the context of Non-RT RIC.
However, dApps are really crucial, notably in:
Device-level power management, where sleep modes and adjusting transmit levels could be useful
Cell-Power used, dynamically adjusting cell parameters such as cell sleeping based on real-time traffic and user activity.
Implementing power-saving strategies at the application level (I am not entirely sure how this can be applicable but just an idea).
Similarly to other use cases, next-Gen LLMs used in the Non-RT RIC can provide a boost in the higher level intelligence for analysis and planning although dApps at the edge could handle the immediate, real-time actions or corrections.
Key Considerations Across Use Cases:
a) Integration with OSS/BSS Layers:
AI-RAN intelligent decisions, to be fully made operationally meaningful, need to be seamlessly integrated with existing OSS/BSS layered systems tasks like service provisioning & orchestrating, Billing & charging also @ Edge and customer relationship management. It should be noted that traditional OSS/BSS were not designed for the dynamic nature of AI-RAN but should be seen as a jump from SON (Self-Organized Networks) to dynamically orchestrated RAN networks.
Several considerations are important:
First, AI-driven orchestration, where AI-RAN dynamic nature necessitates an AI driven orchestration within the OSS/BSS layer. Because, AI-RAN should then considered as a sub-domain orchestration function where AI-RAN would be fused with the SMO therefore the OSS/BSS platform should be able to understand and respond to the real-time insights pushed by the AI-RAN significantly enhanced SMO.
Second, AI-RAN enhancing SMO means that SMO’s ability to manage resources efficient would be dramatically improved therefore the OSS/BSS layer must be on par with this new capability to make data-driven business decisions such as pricing optimization, target new customer segments or develop new services.
Third, monetizing the rise of edge computing, new SMO with AI-RAN capabilities might require edge-aware billing and charging capabilities in combination with network API exposure. In such a way service models would become more sophisticated enabling operators to implement usage-based billing and provide differentiated pricing and monitoring options and pricing options.
Finally, Integration with hyperscaler platforms would be unavoidable as they are being an integral part of telecom ecosystem. This does not only include data analytics exchange but also with hyperscaler-provided orchestration, management tools and exposure brokerage capabilities. As such, hyperscaler offerings represent an extended business/operational footprint of any Telco or Techco operators.
b) Service Assurance:
AI-driven network requires new approaches to monitoring, fault management, performance monitoring and operational troubleshooting.
Key advantages include:
AI-driven initiatives must provide proactive monitoring and explainable AI, ensuring an understanding of the reasoning behind decisions.
Automated fault management which encompasses fault detection, isolation, and recovery while minimizing service disruptions.
Real-time performance analytics with open access to a single pane of glass for multi-domain KPIs.
Multi-Domain Integration:
A Service Assurance solution must integrate with AI-RAN across all pertinent network domains—radio, transport, core, and cloud—to provide a comprehensive view of service quality. This integration requires standardized data modeling, compatible distributed datalakes, and appropriate communication methodologies for data convergence.
Correlation and Root-Cause Analysis:
AI-RAN insights should be correlated with data from other domains to provide a comprehensive view can enable more accurate and efficient root cause analysis.
Predictive Maintenance
AI-RAN with next-Gen LLMs can be used to predict equipment failures and schedule maintenance proactively, reducing downtime and improving network reliability. This could be crucial at edge/DU level, where access to equipment can be challenging.
Closed-Loop Optimization
Service Assurance could naturally be integrated with SMO-fueled AI-RAN to enable closed-loop optimization. Insights from service assurance can be used to automatically impact decisions at the Global service orchestration level or at low-layer such as SMO-based AI-RAN.
Integration with Multi-Domain Service orchestration
As hinted in the previous chapter, service assurance HAS to work fully integrated with multi-domain service orchestration.
c) Network/Service API Exposure (Camara/GSMA Open Gateway/TM Forum):
Challenge:
API Monetization Models means exploring different monetization models for your APIs. Consider tiered pricing, usage-based fees for B2B customers, or even revenue-sharing agreements with partners for B2B2X approaches. Think about how you can create value for third-party developers and incentivize them to use your APIs. Consider API Platforms for aggregating partner APIs, managing and securing API access and collaborating with hyperscalers for API brokerage (routing of API calls towards partners).
Final Considerations:
Data Governance and Security: With increased data sharing and collaboration, data governance and security become even more critical. Implement robust mechanisms for data privacy, access control, and compliance with regulations.
AI Model Lifecycle Management: Managing the lifecycle of AI models is essential. This includes versioning, testing, deployment, and monitoring of models (CI/CD/CT). Develop a process for updating and retraining models as network conditions change.
Skills and Expertise: Building and operating an AI-RAN requires specialized skills and expertise. Invest in training and development to ensure that your team has the necessary knowledge to manage these complex networks.
Collaboration and Ecosystem: AI-RAN is not something that can be built in isolation. Collaboration with hyperscalers, vendors, and other industry players is essential for success. Actively participate in the AI-RAN ecosystem to stay up-to-date on the latest developments and share best practices.
Conclusion
AI-RAN represents a significant evolution, moving beyond simply adding AI to Open RAN. It's a fundamental reimagining of how we design, build, and operate radio access networks. The advent of efficient and application-driven LLMs, such as DeepSeek and Qwen, alongside the power of hyperscaler data and collaboration, is accelerating this transformation, enabling the deployment of sophisticated AI models directly within the RAN infrastructure.
By intelligently integrating these LLMs with lightweight AI/ML models, we can create networks that are not only more efficient and performant but also capable of dynamically adapting to the ever-changing needs of users and applications. This article has examined how this approach can be applied to three key use cases, demonstrating the tangible benefits of AI-RAN. However, realizing the full potential of intelligent RANs requires careful consideration of several crucial factors. These include seamless OSS/BSS integration with AI-driven orchestration, robust multi-domain service assurance, and the strategic exposure of network capabilities through standardized APIs. Furthermore, data governance, security, AI model lifecycle management, and the cultivation of specialized expertise are essential for long-term success.
By proactively addressing these considerations, Telcos/Techcos can unlock the full promise of intelligent RANs and pave the way for a truly connected future.
Comments