As the telecommunications landscape shifts from laboratory testing to full-scale deployment, the integration of artificial intelligence within the Radio Access Network (RAN) has become the defining frontier for global carriers. Dominic Jainy, a seasoned expert in machine learning and network infrastructure, joins us to discuss how industry leaders are navigating this critical transition. With a focus on moving beyond the hype, our conversation explores the practicalities of real-time optimization, the economic hurdles of “AI worthiness,” and the architectural demands of managing distributed intelligence at the network’s edge.
Our discussion covers the measurable performance gains reported by major operators, specifically focusing on the trade-offs between computational cost and spectral efficiency. We examine the evolution of “AI-worthiness” as a gatekeeping metric for new deployments and how distributed architectures are requiring more disciplined lifecycle management. Finally, we delve into the strategic shift toward open, programmable platforms that allow for multi-vendor innovation and automated, closed-loop network remediation.
How are these AI-native link adaptation features currently transforming the baseline metrics for throughput and spectral efficiency in your live network environments?
We are seeing a profound shift in how the network breathes and adapts to user demand in real-time. In our recent deployments, AI-native link adaptation has moved from being a experimental feature to a core driver of performance, specifically showing significant improvements in how we manage spectral efficiency. By moving away from rigid, rule-based systems, the RAN can now predict channel conditions with much higher precision, which directly translates to a more robust downlink. For instance, in collaboration with partners like Ericsson, we have documented measurable gains in downlink spectral efficiency and overall coverage that simply weren’t achievable with traditional software. These aren’t just marginal upticks; they represent a fundamental change in how we utilize our most expensive asset—the spectrum—to ensure that the customer experience remains seamless even in high-density environments.
What specific challenges have you encountered when attempting to move these AI capabilities beyond small-scale trials into a distributed network architecture?
Scaling is where the true complexity of the modern network reveals itself, as the challenge shifts from proving the technology works to managing it across thousands of sites. When you move beyond a controlled pilot, you have to account for a massive variety of environmental variables and hardware configurations, which necessitates a transition toward open, programmable architectures. We have found that automation is the only way to survive this transition; without closed-loop execution and disciplined lifecycle management, the operational burden of updating models at the cell edge would be overwhelming. It is no longer just about the algorithm, but about how that algorithm is maintained, updated, and monitored across a distributed management and orchestration layer. This requires a level of organizational discipline that connects the data foundation directly to the site deployment teams, ensuring that every AI application is a stable, long-term addition to the infrastructure.
Could you elaborate on the concept of the “AI-worthiness” test and how it dictates which problems your team chooses to solve with machine learning?
The “AI-worthiness” test is our primary defense against “AI for the sake of AI,” ensuring that every deployment delivers a tangible return on the increased computational cost it introduces. Every time we integrate an AI feature, especially at the gNodeB, we are adding a layer of complexity and power consumption that must be justified by a meaningful performance gain over the existing rule-based baseline. We look for specific high-friction problems—such as interference optimization or complex beamforming—where traditional logic hits a ceiling and only a machine learning model can find the extra efficiency. If a proposed AI solution doesn’t demonstrably improve the user experience or lower operational costs through automation, it doesn’t pass the gate. This rigorous vetting process allows us to focus our resources on high-impact areas like Layer 1 and Layer 2 functions where real-time processing can truly revolutionize how traffic is steered and managed.
In what ways is the move toward real-time processing at the gNodeB affecting your strategy for energy efficiency and operational cost reduction?
Moving intelligence deeper into the radio access network, specifically into real-time Layer 1 and Layer 2 functions, is a double-edged sword that we are managing very carefully. On one hand, these AI-native capabilities allow for incredible precision in channel estimation and fault detection, which can trigger closed-loop remediation before a human technician even knows there is an issue. This speed reduces manual operations and accelerates issue resolution, which are massive wins for our bottom line. On the other hand, we have to balance these performance gains against the energy footprint of the additional compute power required at the edge. Our focus is on finding that “sweet spot” where the AI helps us save energy—perhaps through more intelligent traffic steering or power-down cycles—while still maintaining the peak capacity and stability our customers expect.
As you look at the lifecycle management of these models, how do you manage the frequency of updates and the accuracy requirements for AI operating at the cell edge?
Lifecycle management at the cell edge is one of the most demanding aspects of our current operations because the cost of an inaccurate model is so high in a live environment. We have to be incredibly disciplined about how many models are deployed and how frequently they are retrained to ensure they don’t drift from their intended performance targets. Our operations teams are now tasked with understanding the “provenance” of every model, knowing exactly how it affects troubleshooting and how it interacts with other automated systems in the network. We are moving toward a more fluid, programmable platform that allows us to work with various vendors and third parties, which prevents us from being locked into a single source and gives us the flexibility to update models as quickly as the local conditions demand. This requires a robust data foundation where the feedback loop from the network edge to the management layer is virtually instantaneous, allowing for a level of precision we have never had before.
What is your forecast for the evolution of AI-native RAN over the next 18 months?
Over the next 12 to 18 months, I expect we will see a major shift where the momentum built during our long-term planning turns into a widespread, tangible reality for the average customer. We are going to see AI move beyond optimization of the status quo and start enabling entirely new use cases at the gNodeB that were previously technically impossible. My forecast is that the “AI-worthiness” test will become the industry standard, leading to a much leaner, more efficient deployment of technology that focuses strictly on areas where rule-based logic fails, such as complex interference environments. Ultimately, we are building an architecture that isn’t just “smart” for today, but is open and programmable enough to support the radical AI capabilities that will emerge over the next several years, ensuring the network remains a platform for constant innovation.