We use cookies to improve your browsing experience. By continuing to use the website you allow Senza Fili to deploy cookies, as detailed in our privacy policy.
Introduction
AI is everywhere and nowhere in the RAN. It is pervasive and yet elusive. We know that eventually AI will permeate the entire RAN, but we are far from this. Wide commercial deployments are still a rarity, with AI technology selectively used for specific tasks. AI can bring huge cost savings, and higher performance and power efficiency mostly through automation, but it is not yet clear what the integration path will be, and everybody in the industry (not just operators) struggles to find the right strategy. How aggressive should the transition to AI RAN be? How should we treat the legacy infrastructure? How should we change operations, organizational structure and culture to integrate AI in the RAN and, more generally, in mobile networks? How disruptive will the introduction of AI be? What’s the best way to prepare for AI and its attending challenges and risks? How should we change our approach to security, sovereignty and privacy? Will AI deliver the boost in revenues operators have been unable to get with 4G and 5G, and are not too confident that 6G will deliver?
AI will have a transformative impact on the RAN, and yet the function of the RAN will not change. The challenges will be the same that we have today, but they will be amplified. Security, for example, has always been a major concern, but AI increases the risk of security breaches. On a positive side, it also protects networks more effectively from security threats.
AI will increase complexity, but it also enables more streamlined operations and higher automation. Operators will be able to launch new services and benefit from greater efficiency, but it is unclear how fast the transition can (or should) be.
There are still many open questions that have to be addressed to chart a successful path to AI-native, fully automated RAN. There is no prescriptive roadmap that works for everyone. Each ecosystem player will have to find its own answers based on its specific context and requirements.
As always, I followed announcements and attended analysts events in preparation for this report, but the main source for what I present here is countless informal conversations I had with those actively working on AI in the RAN—from startups to tier-one vendors, operators worldwide, researchers and experts—which were informed by my graduate research in and teaching AI and cognitive science, at a time when AI was a mostly academic endeavor and which gave me some practical appreciation of the complexity and power of AI. The result is a report that presents a somewhat tangential perspective on AI RAN, centered on how to develop an AI RAN strategy, how to identify and address the unique challenges and opportunities, and how to get to a ROI-positive AI-RAN. 
The multiple roles of AI in the RAN
Let’s be honest: it can be confusing to keep control over the multiple incarnations of AI in the RAN, or elsewhere. As we are still in the early phase of AI RAN, when we still have to fully understand its costs and benefits and how we should deploy it, even the definition of these concepts is often a matter of debate.
Yet they all point to the multiple dimensions in which AI can get into the RAN, the choices and tradeoffs that are available, and the type and functionality of AI models that are best suited for specific goals.
Making AI invisible
New technologies generate hype and, in the case of AI, it is tempting to think that the hype is justified, and more justified than the hype surrounding 5G. Yet, hype does not help to understand the potential of the technology or guide us in deploying it successfully. This is the case in AI, where the mix of fear and awe that it inspires adds an emotional complexity to the assessment of its benefits and risks.
As a result, AI captures most of the limelight, and the AI RAN discussion is about AI more than it is about the RAN. We should not lose sight of the fact that we want to use AI to improve the RAN, rather than finding a way to use AI in the RAN. AI is a powerful and essential tool to enable RAN transformation, and we need to maximize its benefits. Instead, the dominant narrative appears to be predicated on a FOMO and centered on how to maximize the role of AI and speed up its adoption.
A more nuanced and selective perspective is slowly emerging among some vendors and operators, and it will be crucial to the success of AI in the RAN and, more importantly, to the evolution towards an autonomous, more efficient RAN.
AI is not always necessary or worth the cost or effort. We should not use AI for tasks for which we have an effective deterministic (e.g., a well-tested if-then set of rules) model. It would add unnecessary cost and complexity, and no benefit. For other tasks, AI may not have reached sufficient maturity to add any advantage and instead may increase risk.
At the same time, introducing AI is typically not a Boolean decision, but a gradual integration into a complex RAN environment with multiple legacy elements and an operational and organizational structure that may not be ready to adapt to full-fledged AI models.
Reducing the focus on AI and redirecting the attention to identifying the best evolution path for the RAN and, subsequently, the role that AI can play in that, will help ensure a smooth and successful transition to AI RAN.
This is not to deny that AI will eventually permeate all elements of the RAN and the rest of the network, and it will have a massive impact on how we deploy and run networks. On the contrary, its pervasiveness means that AI will not be a differentiating factor. Every vendor and every operator will continue to expand the use of AI. What will set them apart is the path they choose and the ability to successfully integrate AI into their products and networks.
AI will have a transformative impact on the RAN, and yet the function of the RAN will not change. The challenges will be the same that we have today, but they will be amplified. Security, for example, has always been a major concern, but AI increases the risk of security breaches. On a positive side, it also protects networks more effectively from security threats.
AI will increase complexity, but it also enables more streamlined operations and higher automation. Operators will be able to launch new services and benefit from greater efficiency, but it is unclear how fast the transition can (or should) be.
There are still many open questions that have to be addressed to chart a successful path to AI-native, fully automated RAN. There is no prescriptive roadmap that works for everyone. Each ecosystem player will have to find its own answers based on its specific context and requirements.
As always, I followed announcements and attended analysts events in preparation for this report, but the main source for what I present here is countless informal conversations I had with those actively working on AI in the RAN—from startups to tier-one vendors, operators worldwide, researchers and experts—which were informed by my graduate research in and teaching AI and cognitive science, at a time when AI was a mostly academic endeavor and which gave me some practical appreciation of the complexity and power of AI. The result is a report that presents a somewhat tangential perspective on AI RAN, centered on how to develop an AI RAN strategy, how to identify and address the unique challenges and opportunities, and how to get to a ROI-positive AI-RAN. 
Lessons from Open RAN
A rigid AI-centric view of the evolution of the RAN may also have a polarizing impact that limits innovation and effective deployments.
Open RAN offers a valuable lesson. Operators have created Open RAN to encourage multi-vendor deployments, and continue to support openness in the RAN. Understandably, some vendors have resisted the transition to a more open RAN, as this may reduce their end-to-end control over the RAN and revenues.
Despite this, there is an undeniable trend towards more openness in the RAN, and almost all vendors have come to accept it and develop products that work in multi-vendor environments, even though they prefer to control the end-to-end network as much as they can, and in most cases they still do.
Yet market estimates and forecasts for Open RAN are invariably disappointing, with commercial deployments mostly confined to greenfield operators or to small portions of the network of brownfield operators.
What many perceive as a lack of success or failure of Open RAN is the result of a too rigid and demanding definition of what Open RAN is, which does not fit into what operators need and want to deploy.
Operators have been gradually increasing openness in their networks, driven by their needs, financial and organizational resources, and commercial availability. In most cases, these changes do not neatly fit into the Open RAN definition, but they are gradually, organically and sustainably driving RAN evolution.
We should avoid to replicate this approach with AI RAN. If we force feed AI into the RAN and define a path for AI adoption that is not sufficiently flexible to meet operators’ requirements and resources, we may end up with disappointment over what we may perceive a limited adoption of AI RAN. AI is a technology with a huge transformative potential, but it could backfire if operators adopt it to meet industry pressures rather than their own RAN transformation strategy.Â
Doing the same, differently
How is the AI RAN going to be different from today’s RAN? The functionality will remain the same, with some added features that today are too computationally intensive to implement or are not yet available (e.g., those that will be available with 6G). Yet, the most promising potential of AI comes not from doing the same things more efficiently, but from doing them in fundamentally different ways. This is an aspect that is often glossed over, but that carries new challenges and risks that make the transition to AI a very delicate process that each operator has to manage carefully, avoiding herd thinking and finding their own path. Â
AI models harness a compute power that is unmatched by the human workforce, and, by itself, improves network efficiency. Humans have strict short-term memory limitations (5-8 items at a time) and that sharply limits their ability to optimize systems with a large number of interacting parameters. They are also unable to operate continuously in real time, with no sleep or time off. Hiring more staff would not be sufficient, nor financially sustainable, to remedy these limitations. Not only does AI improve the efficiency of individual RAN elements, but it can optimize performance across network elements and sites continuously, based on real-time traffic load and resource availability. When interacting with humans, AI also provides limitless patience, availability, and willingness to adapt to the needs and capabilities of its interlocutors.
Impressive as this is, this is a brute-force extension of today’s RAN: more of the same. The more disruptive innovation from AI comes from performing the same tasks differently from humans. Our experience and expertise make up for our limitations in number crunching and keep the networks running. Telecom AI models do not try to duplicate our learning processes—or indeed our “intelligence”. At present, it would be more difficult to do so (we still have a limited grasp of how human intelligence works) and not desirable. Having an AI model with the same short-term memory limitations as humans would not take us to AI RAN that is much better than our existing RAN.
Instead, AI learning and inferencing use techniques developed for models that can digest massive amounts of data and produce different outcomes from humans. They can show us alternative approaches and solutions that we were not able to see before, and that may urge us to reevaluate how to deploy and run networks.
They may also be flawed, and have undesirable performance, reliability and security impacts. To protect the RAN, we need to adopt operational and organizational models that allow us to manage appropriately the behavior of AI models.
AI amplification: The price of a dynamic RAN
AI models have an inexhaustible stamina in absorbing data, finding problems and solutions, devising new tricks to increase efficiency, protecting networks from security threats, and assisting customers and staff. The increased activity creates an amplification problem.
AI and automation will make networks more dynamic and capable of adapting to requirements and resources in real time. But to do so, AI models will trigger many more network adjustments than we are currently used to, and this may increase the number of erroneous interventions even though the rate of error is lower than for human-driven ones.Â
Without AI-driven automation, operators do not have the capability to aggressively optimize the RAN and instead use overprovisioning to ensure reliability and good performance.
Without the capillary network monitoring that AI can add, operators may not be aware of issues that may not be crippling but nevertheless reduce the performance.
Optimizing the RAN in real time is one of the biggest benefits that AI brings, but it comes with increased complexity and, at least initially, as the models and humans supervising them are still learning, with a higher potential for suboptimal or wrong decisions.
It is imperative to proceed with caution to fine tune the AI models and keep the humans in the loop until they reach sufficient trust in the models.
This is a two-way process. Humans have to check the output of the models, but the models have to become transparent and let the humans understand how they operate and how they get to their output. In turn, this allows humans to take corrective action and, in the process, to gain a better insight and higher trust in the models. AIOps is gaining traction to address these issues by reframing how operations work in an AI-driven RAN.
The path to AI RAN
AI will make the RAN more efficient and less reliant on direct human intervention, improve the economics and user experience, mostly by enabling automation. At the same time, an automated RAN will be more complex and vulnerable, adding new security and reliability requirements. As we noted, it achieves this by injecting more compute capabilities and by fundamentally changing RAN operations.
The introduction of AI in the RAN and the final transition to a AI-native, fully automated RAN have to carefully managed to protect the RAN from the vulnerabilities and amplification issues AI introduces and to ensure that AI delivers the desired outcomes in cost-effective ways. It is not a one-time upgrade; it is a protracted process that requires investment and new hardware and software, but also an organizational and cultural reorganization, with three partly overlapping components:
Â
- Getting ready: AI uses historical, current and real-time network data and documentation that is available but has not been extensively used in the past. We need to ensure that this data is reliable, accurate and accessible to AI models. We also have to feed the data at the right granularity level or give the models the ability to zoom to the appropriate level for a given task. Too much detail is likely to limit the ability to generalize. Insufficient data may prevent the model from learning or inferencing correctly.
Â
- Gradual integration: AI has to be smoothly integrated into the existing RAN, and initially the adoption has to target specific and limited functions where AI can bring a provable short-term benefit and can be added to the legacy infrastructure. AI deployment can then expand as the operator becomes comfortable with an increasing role of AI and higher automation. This is a phase of intensive exploration and learning:
- Selection of AI models and commercially available that are best suited the operator’s needs
- Integration of AI tools within the RAN and with already deployed AI tools
- Training, testing and refining AI models
- Organizational change to adapt to AI-driven automation
Â
- Pulling it all together: The initial piecemeal addition of AI functionality throughout the RAN may lead to unexpected conflicts or undesirable interactions. Suggested actions from AI agents with different remits may create conflicts that orchestration has to resolve. Maximizing the efficiency of the overall RAN or network or meeting specific performance targets requires more than the optimization of individual elements or functions. For instance, minimizing power consumption may have a negative impact on user experience, and tradeoffs are necessary to resolve issues like these. Similarly, optimizing the performance at one cell site may reduce the overall performance in the surrounding area, and multi-site optimization can lead to greater RAN efficiency. These are not not new, AI-specific challenges, but they become more consequential in an automated RAN that is inherently more dynamic, and they have to be addressed to unleash the full potential of AI.
Going AI native
The endgame is to get to an AI-native RAN, where AI permeates all functions natively, starting from Layer 1. This requires a long-term, massive transformation of the RAN, and we may need to wait for 6G to get to that stage for commercial availability. Assuming a deployment timeline comparable to 5G, we should expect several years before reaching large commercial deployments.
We cannot wait for 6G and AI-native RAN to get AI in the RAN. Yet, the AI-native RAN is increasingly presented as an urgent strategic priority, rather than a long-term goal. This risks shifting the efforts of AI adoption in the RAN away from what is available and cost-effective today.
A more realistic approach that yields measurable benefits in the short term and can be integrated in the existing RAN is to layer AI on top of the existing infrastructure, initially focusing on low-hanging-fruit use cases such as power efficiency, management of radio resources, or fault management, where AI can be bolted onto the legacy RAN and no rip-and-replace is required. This keeps the cost of the addition of AI low and reduces risks and complexity.
The transition to AI-native RAN will be gradual and driven more by operator’s life-cycle requirements than commercial availability. Similarly to virtualization, operators will shift to AI-native elements gradually, depending on their need to replace existing infrastructure or install new one, and their readiness to move to an AI-ready solution.
Layering AI from the top
AI may be everywhere in the RAN, as I stated at the beginning, but true AI-native RAN is more than a decade away. How do we reconcile that? There are two dimensions to consider.
From bolt-on to native AI. For each RAN element, AI can be bolted onto what is already there. This requires limited investment and effort, carries lower risk, and enables operators to explore how AI can improve over a RAN mostly run by humans. (AI and ML models have been used for quite some time already, so AI is not new to vendors or operators, but the extent of its potential has expanded.) Many vendor solutions promise to do so, on the correct assumption that most operators are reluctant to replace their infrastructure to add AI capabilities. The downside is that this approach limits AI capabilities and benefits, and is likely to be less energy efficient because it is not optimized for AI.
When operators are ready for a lifecycle refresh or network extension, they may switch to a solution with a tighter, more efficient integration of AI, and eventually to an AI-native solution. At this point, operators will be ready to integrate the solution at a deeper level, with the additional investment and operational change required.
From AIOps and energy management to PHY/MAC AI-optimized silicon. Use cases such as AIOps and energy management are among the first ones to be deployed because they are relatively easy to integrate in the RAN today (i.e., no need to wait for 6G) can unleash the compute power of AI on massive data sets that are available, and address pressing operators needs to cut costs and increase network efficiency in the short term. At the same time, these use cases give staff exposure to how AI is changing not only the RAN, but also their jobs and the organizational structure.
Do we need and can we afford AI?
The AI train has left the station. I do not know of any RAN vendor that is staying out of AI. The depth of AI integration and the benefits of AI vary across solutions and vendors, but AI has an increasing role that goes beyond the marketing pitch that was most prominent one in past years. Equally, operators are all committed to deploying AI, even though their readiness and speed of adoption vary greatly. Surveys capture the mood around the adoption of AI, but give little insight into how AI is being deployed and the challenges and benefits it brings.
But do we need AI, or are we following an unavoidable hype trend? AI is a hugely transformative technology and it would be foolish to ignore it. Most importantly, it is not an option; it is a necessity to manage the complexity of today’s and future networks and to optimize them to get the economics right and extract all the benefits of new telecom technologies. Deploying the latest beam or power management solutions without AI greatly limits their efficiency and with it the soundness of the capex investment. It is inconceivable to get to L5 fully autonomous networks without a major AI contribution, but we need AI well before we get there. We need it today to progress to L3 or L4. The open question is not whether AI is for real, but how to best integrate it.
A frequent objection to AI is that it is a shiny object that operators cannot afford in the current challenging environment. This is the case for a GPU-dominated AI RAN, but that’s not the plan for operators in the short-to-mid term.
Depending on how AI is introduced, the capex investment may be limited, but AI requires an internal operational effort for integration that may be demanding of financial and human resources. More importantly, AI makes sense to deploy only in cases with demonstrable benefits, and that clears the ROI hurdle.
The multiple flavors of AI
AI fundamentally differs from traditional deterministic models because it does not need instructions or established expertise. It probabilistically extracts patterns from massive data sets to generate the desired output. Because it does not follow the if-then algorithms of deterministic models, it is much more powerful in finding novel or more accurate answers, but it is also more prone to mistakes. This is an acceptable tradeoff, but one we have to get used to, by building guardrails to reduce the likelihood of mistakes and biases, and learning to spot them.
Nevertheless, even the best AI model adds complexity, cost and risks over a deterministic model, and we should keep AI to perform tasks for which we do not have effective deterministic algorithms. This is the case for most tasks in the RAN and in many cases, an AI model may show us that the deterministic model we have used in the past could be improved by moving to AI.
The same data-driven, probabilistic approach of AI, however, works over a large variety of model architectures, with different capabilities and suited for different tasks. We need this diversity and flexibility to harness AI to help us across use cases.
Initially, LLMs were often presented as the next-generation technology, replacing traditional AI/ML models. Today, many treat agentic AI as the new frontier that will solve all the problems where AI/ML and LLMs models failed.
AI and ML models that use historical and real-time network data to optimize and automate the network will retain a central role in the AI RAN. They are operationally efficient but not transparent, and often leave humans wondering how they got to their results and if they can trust them.
This is where genAI and telco small language models (SLMs) come in. Unlike the hallucination-prone LLMs we use for our less mission-critical daily work and personal
queries, telco SLMs are more accurate and reliable and have lower compute and power requirements because they are sharply limited to network data that should have been verified and cleansed. These language models create the interfaces humans need to gain access to information about the network, to get visibility into AI/ML models, and, ultimately, to get the support they need to complete their tasks. This is especially valuable for junior staff who may not have the experience and expertise needed to execute a task.
Agentic AI is the latest flavor of AI, and it is likely to be followed by new ones. It is difficult to escape the anthropomorphizing urge to see agents as human replicas, arguing with each other and complaining about their orchestration agent boss. They have a well-defined scope of action and help humans or other AI models in specific tasks. Language models give us information; agents help us do things and find solutions to specific problems. They are highly complementary and will often be used in a collaborative environment.
AIOps is an example of how different AI models come together to transform the way we operate the RAN and gradually move towards autonomous networks. AI/ML models are the foundation for applications such as energy optimization, fault and anomaly detection, or traffic and resource management. Language models combined with RAG can assist humans in gathering and managing the output of these models. Agents can then build on this to take further action and execute tasks. On top of this, AIOps can use digital twins to explore what-if scenarios that project and assess the impact of changes in the existing network in a virtual environment, and to avoid costly mistakes.Â
What makes a RAN an AI RAN?
There are many ways to answer this question, but it may be best to not try to answer it at all.
In the current market environment, where almost every vendor product has AI built in, it is difficult to decide what counts as AI RAN. A solution may be AI-ready, but the operator can decide whether to use the AI functionality and to what extent to do so. Does it mean that any RAN that includes an AI-ready element is an AI-RAN? Or do we require more than AI readiness? Is it sufficient if AI is used only for one use case? This creates a useless definition rabbit hole that does not benefit anybody, with the possible exception of some vendors.
Similarly, defining AI RAN as being AI native, pushes AI RAN into a long-term adoption timeline, and ignores the fact that AI is rapidly taking hold in the RAN along different depths and dimensions, through a layering approach.
A strict definition of AI RAN may introduce similar risks to the Open RAN overly strict definition. It creates polarization in the industry and adversarial narratives (i.e., AI RAN will never work versus AI RAN is everywhere) that fail to capture the transformative, but gradual contribution of the technology.
My suggestion is to avoid spending time defining what AI RAN is, and to be careful in assessing the market size and revenue opportunity for AI RAN. Dell’Oro forecast does exactly this. It predicts that AI RAN will account for about a third of RAN capex (and that’s not in addition to the overall capex forecast) by 2029, but it uses a flexible definition of AI RAN that includes bolt-on, non-AI-native AI solutions. Over the next decade, AI RAN may capture most of the RAN market by this definition and, at that point, we may not need a separate forecast for AI RAN.
Does it all add up? Monetization and ROI
At least half of the hype on AI RAN comes from the monetization opportunity it is expected to unleash—and that sounds ominously similar to the one presented for 5G and, before that, for 4G that failed to materialize.
ROI. The case for a positive ROI for AI in the RAN is easy to make, as long as operators choose wisely how, where, when and for what to integrate AI. AI can increase network efficiency and thus reduce costs, and this can justify the investment. A gradual adoption of AI is also a strategic move within a longer-term RAN transformation to autonomous networks that requires deep and ongoing organizational change, even though it may reduce the initial ROI attractiveness because of the upfront costs.
A positive ROI is the primary driver for AI RAN, and one that the operator can control by choosing the appropriate adoption strategy.
In addition, AI adoption among users and enterprises enables and drives new services or applications, or enriches existing ones, creating more demanding performance requirements, e.g., for capacity, latency, reliability and security. AI will help operators meet the new requirements by increasing network efficiency, instead of investing in non-AI infrastructure (e.g., adding new cells) which is more expensive.
Monetization. It is not clear if and to what extent operators will be able to monetize AI directly by offering new services. For instance, AI is going to be a crucial enabler for sensing (ISAC), and will widen the opportunity for drones and network APIs, but AI will not be the differentiating feature of the service. Operators have to sell a drone service to get revenues, not AI.
Enterprises will be heavy consumers of AI, and increasingly physical AI, and some of this use of AI will require connectivity over public mobile networks. But it is still unclear how much the role of mobile operators will extend, if at all, beyond that of providers of connectivity.
The AI-on-RAN monetization opportunity from leasing unused GPU capacity is the most problematic:
- In the short term, it would require operators to buy GPUs they do not yet need to lease capacity. It is doubtful that this could be a sound business model unless it is part of edge AI enterprise services, not specifically tied to the RAN.
- The RAN has extremely tight reliability and performance requirements, and operators will be cautious in opening up precious and vital GPU resources to third parties, especially in a competitive market where revenues will be constrained.
- Leasing spare capacity assumes that there will be demand outside RAN peak hours. But it is likely that third-party demand is correlated in time and peaks at similar times, and operators will not be able to meet third-party requests unless they overprovision the RAN, which is an expensive proposition.
Takeaways
- Use AI where you need it. Not everywhere (yet).
- AI does things differently. We need to learn to do things differently, too.
- Integrate AI gradually, layering it from the top.
- Choose the AI tools you need wisely to get a positive ROI.
- No hurry to get AI-native. No need to deploy GPUs to have an AI RAN initially.
- Expect cost savings. Be wary of too optimistic monetization opportunities.