AI’s Impact on Wireless Infrastructure: More Uplink Traffic and Opportunity

By Paul Reddick, Board Advisor, Inorsa 

Note: This blog was produced under WIA’s Innovation and Technology Council (ITC). The ITC is the forum for forecasting the future of the wireless industry. Participants explore developments in the wider wireless industry, from 5G network monetization trends and streamlining infrastructure deployment to future spectrum needs and cell site power issues. These views are not a WIA endorsement of a particular company, product, policy or technology.

The conversation around AI and infrastructure usually focuses on data centers, power, and training frontier LLMs. But for the wireless industry, that is largely noise. The real story unfolds in two parts: first, how AI-driven inference and edge computing are transforming mobile traffic and creating new demand for physical infrastructure; and second, how wireless operators and infrastructure providers can harness AI intelligence to radically improve how we build, own, and operate networks, moving step-by-step from task automation to true agentic workflows. 

The AI Infrastructure Hype That Doesn’t Involve Wireless 

To date, most of the focus has been on the infrastructure needed to train AI models. Training the models requires enormous amounts of computing power, GPUs, cooling, large expanses of land, electricity, and data. It also requires vast amounts of fiber to move information to and between the data centers where that training takes place. The work is highly centralized, bringing together enormous quantities of text, video, images, audio, and other information to generate the intelligence embedded in a model. 

None of that depends heavily on wireless networks. When and how do wireless infrastructure owners get to play? Just wait, grasshopper. 

The AI Infrastructure Trends that Do Involve Wireless 

The trillions of dollars now going into model training will need to generate a return. This means AI applications must deliver real-world solutions and products people will pay for. As AI moves into everyday applications, inference will become an increasingly important part of the equation. In fact, over the past few months, we’ve passed the point where more data center capacity is dedicated to inference than training.

Inference is different. It is when the intelligence created during training is actually put to work, increasingly through phones, vehicles, factories, robots, cameras, and other connected devices. As more AI applications move into the real world, the networks connecting those devices to the computing resources behind them will become increasingly important. 

That’s where mobile networks enter the AI infrastructure story. 

AI changes the traffic equation 

Mobile networks historically have been built around a significant imbalance between downlink and uplink traffic. Roughly speaking, about 92% of traffic travels down to devices (the downlink), and only about 8% comes from the device (the uplink). 

Think about how we’ve traditionally used a phone. You make a small request for a video stored in the cloud, and the video (a much larger file) comes back down to you. Over the past decade, video has driven much of the growth in downlink traffic. 

AI can begin to shift that relationship. 

Take a picture and ask AI, “Where am I?” You may send a full photo of the Golden Gate Bridge upstream and get two words back: San Francisco. You might send a video or even a gigabyte of information upstream and get a very small answer in return. 

Ericsson studied traffic from consumer AI applications such as ChatGPT and Copilot, and the ratios can be much more symmetrical. Instead of roughly 12 times as much traffic moving down as up, it can move closer to 3:1 and, in some cases, toward 1:1. 

It’s one thing to transmit from a tower, where you can put significant power behind a downlink signal. But on the uplink, battery-constrained devices transmit at milliwatts – and the tower still has to reliably decode what they’re saying. 

As AI generates more upstream traffic, networks may need larger or different antenna arrays to hear all those devices. That means more load on towers and a greater need for infrastructure. More uplink traffic benefits wireless infrastructure providers. 

Where does all that inference happen? 

Much of the AI processing today goes all the way to large cloud data centers. Some happens at metro data centers, especially when applications need lower latency or when it makes sense to avoid extra transit costs. Some can occur directly on the device or on-premises, such as at an airport, where cameras and sensors continuously generate data. 

Some AI computing could also happen at or near tower sites. The industry has discussed mobile edge computing at tower sites for years, but carriers never moved network functions that far out. The closest we got was Dish’s earliest O-RAN designs. 

Now, the rapidly growing demand for scarce AI inference compute could create new reasons to locate that capacity at a tower site. Capacity is critically undersupplied in the near term. 

The attraction may be practical. Land, permitting, a shelter, and available power may already be in place. Available power may be limited. But if developing a new metro location will take much longer, a tower site may be a faster way to get computing capacity up and running. And every little bit helps. 

Further, some use cases require extremely low latency. Instantaneous motor control on a robot, where latency needs can be as low as 5-10 milliseconds, can’t run from a distant metro datacenter. In those cases, the robot would have to be within 1-5 miles of the compute. If those robots are mobile, say cutting tree branches vs. working on an assembly line, tower locations may best meet the requirement. T-Mobile is experimenting with this today, working with humanoid robotics developer Figure AI.

That doesn’t mean all inferencing is moving to towers. Where inferencing happens will depend on the application and what makes sense technically and economically. 

Putting the intelligence to work 

Of course, building infrastructure is only part of the equation. The real test is how we put all that intelligence to work, not just in consumer and enterprise apps generally, but behind the scenes in the wireless industry itself. 

For example, in the case of Inorsa, AI can bring order to an otherwise disorganized set of information about physical infrastructure, equipment specifications, permits, environmental analyses, leases, construction drawings, RF data sheets, mount analyses, structural analyses, and more. 

But that takes more than handing general-purpose model documents and asking for an answer. This requires deeper training, subject-matter expertise, and AI systems designed around infrastructure engineering workflows. The challenge is knowing when and how to apply that intelligence. When do you leverage LLMs with probabilistic answers, when do you need human review, and when should you rely on deterministic sources (e.g., equipment manuals)? 

Beyond just solving a task and getting an answer, AI systems from companies like Inorsa can automate important workflows, such as validating ingested information. Suppose information pulled from a construction drawing doesn’t initially match another project document. An AI workflow might recognize that the drawing is grainy and re-examine it with a different, more precise tool. What initially looked like a 7 might turn out to be a 2. The documents actually matched after all. A well-constructed AI system can decide to pull in the precise tool only when necessary, minimizing expense. 

Once you’ve validated the underlying information, you can use it with more confidence for deterministic tasks, such as creating a construction drawing or performing a mount or structural analysis. 

The Ladder of Adoption: From Automation to Autonomy 

Think about it as if you were running a landscaping company. When you first hire a new employee, you might give them a task: mow this lawn. If they do it consistently well, you might tell them to mow three lawns and issue the invoices. You’re giving them more of the workflow, but you’re still telling them exactly what to do. 

The jump to agentic AI is like telling an employee who’s earned your trust: We have six employees, two lawnmowers, three weed eaters, and twelve lawns to complete today. Here’s today’s weather and traffic. You figure out how to schedule the work, assign the crews, and get everything done as efficiently as possible. 

Earning that level of trust is a step-by-step process. In infrastructure, where every detail feeds critical construction drawings, mount analyses, and structural calculations, confidence is built on rigorous data validation and seamless human-AI partnership. As the system proves its precision, it unlocks greater autonomy and frees teams to scale. 

Ultimately, AI’s impact on wireless infrastructure comes down to two major shifts: a surge in uplink traffic and edge compute that drives new demand on existing assets, paired with reliable intelligence that transforms how we engineer, build, and deploy them. 

That’s where AI starts earning its keep. 

Paul Reddick is an advisor to Inorsa and other firms, offering pragmatic guidance to founders, leadership teams, boards, and investors. He has a proven track record of deploying emerging technologies and currently focuses on applied AI. Explore his practice at Reddick & Company.