Every year, the mobile industry has a new buzzword. This year, it’s “AI.” That’s artificial intelligence (AI), not an abbreviation for “Albert” or A1 steak sauce. Companies are enamored with this term because Nvidia has soared past a $2 trillion valuation.  

But let’s look more closely at what AI will really do in the mobile telecom market:

First off, let’s address the use of generative AI (GenAI) models to “chat” with customers or network techs and provide a human-language interface to the network. This might be useful as a way to replace human beings in the Customer Service call center, or in field operations, but my point of view is that this is a peripheral change, unlikely to work better than humans, and probably won’t realize as much business impact as people hope. 

The human-language aspect of AI is a nearly infinite problem: It requires huge databases and truly massive computing power. For that reason, I wonder whether the business benefit will be worth the investment.

Second, I’m more interested in the use of AI to enhance capacity in the network. In contrast with the boundless problem of human interaction, this is a finite problem with clear boundary conditions and clear metrics. Applying AI to the RAN, for example, can improve the way that users are grouped, the way that beams are steered and coordinated and the way that modulation and coding settings are used. This directed use of technology is aimed at the heart of the business case for a mobile operator: They need more capacity from each dollar of investment.

AI for capacity enhancements

Over the past year, I’ve had access to some field trials that are very promising for AI-directed capacity enhancements. The results are very encouraging, with a 20% – 40% increase in RAN capacity through the application of AI.

Each network vendor has a different way to view AI implementation and the best approach to capacity enhancement, but in general they choose 20-30 variables or “settings” in the network and remove the human algorithms previously used to choose the settings. In this case, “AI” means that the optimization engine can create its own algorithm to determine the best settings and achieve the highest possible capacity.

Recently, I saw a GPU workshop advertising a discussion about how to use AI to stretch 6G beyond Shannon’s limit. Hogwash! AI is not magic. It simply helps to bring the communication channel back to its ideal state when the network setup is not ideal.

In 5G, we have brought OFDMA to a level where we are approaching Shannon’s limit, meaning that we’re reaching a point where we are transmitting the maximum possible amount of information within a given bandwidth.

Well-optimized 5G networks are near this limit today in test networks under ideal lab conditions. In commercial networks in the field, nobody achieves the same results as ideal lab conditions because there are obstacles and interference and misalignments in the network. The real-world spectral and spatial efficiency is often half of the results achieved in a lab.

Half! Some of that loss cannot be recovered, because path losses (like the loss in penetrating a building) force the 5G link to a less efficient mode. But a great deal of inefficiency is built into today’s networks based on the constraints of human algorithms used to set up the network.   

Here’s an example: Setting up a 5G network, the technicians point the antennas in a given direction with some overlap between sectors. They set various frequency channels and turn on the network. Users will pop up with assignments to resource blocks and beams that are relatively random based on what’s available at the time. To improve the performance for an individual user, the network can reassign the user to a new band or massive MIMO beam, but there is not much coordination with other sites to ensure low cross-talk between beams.

With AI optimization, the beam selection, frequency channel selection, carrier aggregation settings and other technical parameters can be set holistically, instead of sequentially. That is how the AI engine can find improvements: by breaking users out of conditions that are sub-optimal.

In fact, what AI will do is to bring the average performance of the RAN closer to Shannon’s limit, not to make any user exceed Shannon’s limit.

The mobile industry is starting a new era, where capacity gain will not come from large blocks of spectrum or from a new “G” with better peak efficiency. Instead, the industry needs to learn how to squeeze more capacity out of existing spectrum. This will become one of the biggest differentiators in the RAN market three years from now.

Original article can be seen at: