Practical AI for Telco Networks

The AI hype is real, by and large driven by the GenAI element, fueled by mind-boggling bets on AI supercomputing infrastructure that have captured the attention and the imagination of industries, including the telco industry. Is all "AI" hype?

 Thankfully, no.

To discern fact from fiction, the practical from the aspirational, we need to unpack what AI means and what we mean by it. For telco networks, I was fortunate to have the opportunity to unpack and clarify what AI means with Melike Erol-Kantarci , Strategic Product Manager for AI RAN at Ericsson .

To start, Melike and I agree that AI has become overloaded and has largely come to mean generative AI. This appropriation of the term has served to dilute years of R&D and effort Ericsson, and the industry at large, have invested in making more traditional forms of AI, namely machine learning, workable and valuable for the network and operations.

It was a great conversation that I highly recommend that you watch the brief summary video. The link is provided below.

There are several important insights that I took from my conversation with Melike that you should consider as you evolve your thinking about AI for the network and the network for AI.

• Expert-centered approach to AI for the network
• Fit-for-Timescale AI architecture
• Cost of intelligence
• Right-provisioning
• Edge economics
• AI for the Network first

Expert-centered approach to AI for the network

Networks are complex, heterogeneous, and dynamic. Human domain expertise will be essential for progressing and realizing the benefits of AI that cater to the specific readiness and requirements of an operator.

AI-native is and will be a long journey.

Fit-for-Timescale AI architecture

The telco network is a large, spatially distributed system of systems. The idea that a monolithic LLM will execute all the operations across the network is simply not realistic.

For this reason, the kind of AI that is implemented from the radio to the core must be a right fit for the timescale requirements of a feature, function, or process in the network that range from microseconds to seconds and minutes.

You aren't going to run your network on ChatGPT anytime soon or possibly ever. 

Cost of intelligence

I was glad that the cost of intelligence came up in our conversation, which is something often disregarded in vendor talk tracks about AI-RAN. Intelligence comes at a price and should be balanced against value.

Don't discount lightweight and efficient algorithms and machine learning, especially at the far edge of the network where power and economics become a constraint, making it difficult for any meaningful reasoning and learning to occur.

Moreover, the TCO of intelligence must include the full lifecycle of continuous training/learning and maintenance.

Right provisioning

As many network engineers know, over-provisioning compute across the RAN is a no-no, especially at the very edge of the network close to the radio. This means purpose-designed and built compute that is optimized to run domain-specific workloads for high-value use cases of the network, such as link adaptation, will likely remain the baseline value proposition for operators.

Anything more and beyond the network and RAN intelligence and augmentation is and remains hypothetical, and challenging for most, if not all, operators to practically consider.

Edge economics

The edge is a rough place. The economics are very different than what we have seen and continue to see in the hyperscale cloud data centers. At the telco edge, assets need to be designed beyond performance for durability, reliability, and longevity.

As we look at AI for the network, we shouldn't project neocloud economics, which have yet to prove viability, toward the edge of the network. A build-it-and-they-will-come approach could prove ill-fated.

AI for the Network first

AI-RAN will be an evolution. The idea of a converged RAN and AI-MEC (AI-Multi-access Edge Computing) architecture for AI-RAN looks great on paper and PowerPoints but has yet to be proven in principle and viability. The challenges that operators have faced making MEC a viable business are well known. The value of AI-MEC is hypothetical at this point.

According to Melike, Ericsson sees the early days of the AI-RAN evolutionary journey requiring a focus on the network. This AI-for-the-network first approach makes the infrastructure agnostic and independent of how AI-MEC is implemented, developed, and evolved. This segregation could be important as accelerated computing is on an accelerated cadence of innovation that operators are not used to.

Realizing an “AI-native” network promises to help operators scale and economically operate and deliver advanced 5G features and eventual 6G capabilities that could foster a new frontier of mobile AI applications, A.K.A., the network for AI.

Essential Thought Leadership on AI-RAN

For more insights into Ericsson's views and approach to AI for the network and the future network for AI, check out these sources and associated whitepapers.  

• AI RAN – Smarter connectivity (Click to go to site)
• Future-proof your network with AI-driven innovation (Click to go to site)

Also, check out neXt Curve research at www.next-curve.com and our Substack for more AI-RAN insights and analysis that matter!

This material may not be copied, reproduced, or modified in whole or in part for any purpose except with express written permission or license from an authorized representative of neXt Curve. In addition to such written permission or license to copy, reproduce, or modify this document in whole or part, an acknowledgement of the authors of the document and all applicable portions of the copyright notice must be clearly referenced.

© 2025 neXt Curve. All Rights Reserved