AI is infiltrating the networking space at an unprecedented rate, filling transmission pipes with relentless reams of new data unencumbered by time or fatigue, a rush that is set to flood wireless connectivity pipes ill-prepared to handle the flow.

This AI-fueled surge is seen as a boon to wired networks, with firms in that space salivating at the chance to fill their pipes and coffers with AI-generated traffic and revenues. But recent tests have shown cellular spectrum-based wireless networks could struggle with critical uplink speed and latency needs to reap that advantage.

Network testing firm Ookla recently released a report that noted current cellular network configurations are not prepared to deal with AI-generated traffic. The report specifically found that network spectrum resources are heavily tilted in favor of downlink speeds, which are speeds from a cell site to an end-user device, as opposed to the uplink where originating data is transmitted from the device to the cell site.

Mobile network operators have traditionally allocated a vast amount of the spectrum capacity to downlink connections, a path hewn by spectrum-hungry streaming content consumption.

Ookla’s tests showed that percentage to be roughly 90% of resources allocated to the downlink and 10% to the uplink. This ratio can often be seen in consumer speed test applications where downlink speeds vastly outperform uplink performance.

This pattern was backed by Verizon CTO Yago Tenorio, who recently told SDxCentral that, “every system has been designed with downlink in mind, with radios that could be like nine-to-one or eight-to-two, or downlink dominated everything and uplink was a minority.”

This mobile connectivity template is an important consideration for the AI industry as it looks to distribute workloads around a network. AI traffic is variable by nature, with some systems operating in “bursty” patterns while others are more stable, similar with latency needs that can vary from exceedingly latency sensitive to more modest immediacy needs.

Workloads that are housed in large data centers are likely to have robust wired connections able to deal with these divergent network traffic flows. But, as the AI ecosystem looks to move workloads into locations more apt to have cost-efficient power resources or be closer to end users due to speed and latency needs, wireless connections and the performance of those connections will start to come into play.

Tenorio agreed, noting that the current mobile network transmission paradigm “may change. AI may change that.”

What’s causing traffic patterns to change?

Tenorio specifically pointed to AI systems that are being embedded in everyday devices that will change consumption expectations from “a transactional communication that you have today with AI: so you ask a question, you get an answer; to more of a very intelligent assistant that is aware of everything around you without you needing to fill in the context every time you want an action.”

“That is what requires a strong uplink, because either, if it's like video, audio, or sensing, it means that there is information flowing to a processing point in a continuous way,” Tenorio added. “So that's the problem that the industry needs to solve.”

Ericsson, in its most recent annual Mobility Report, pointed to several drivers of increased uplink traffic, including AI-enabled IoT devices, autonomous vehicles, and drones, that “will transmit a lot of data to the cloud as they collect a lot of training data, require data to be stored for legal reasons, and sometimes require remote interventions.”

“As a result, the uplink traffic will increase significantly over the coming years and, indeed, is becoming telecom’s new ‘currency,’” the report stated. There will also be challenges in how AI agents are deployed. An “always-on” agent will more regularly trigger an uplink connection via the mobile network, while an “on-demand” agent will be more parsed in its network interactions.

Similar to Tenorio, Ericsson’s report ascribed this AI connection architecture to the smart glasses market, which is expected to see a boom in penetration. Ericsson noted that this market was currently just a small percentage of overall glasses sold, but the market has “ambitions to sell millions per year going forward.”

“The success driving these sales is in connecting the user to an AI agent that delivers sentient engagements based on video and audio input from the glasses,” the report noted. “Going forward, some models will use AI capabilities right on the glasses and/or tethered devices; however, advanced AI capabilities will need to run in the cloud and – when inference time of the models is low – the uplink network characteristics become critical.

Technology to the rescue

One avenue toward solving that uplink challenge is by advancing cellular technology standards. The most forward-leaning mobile networks today are using a standalone 5G core and 5G-Advanced transmitting technology, which provide the ability to aggregate spectrum resources, dynamically allocate those resources, and manage antennas to best beam and receive data.

These systems were not specifically designed with AI in mind but have enough adaptability to deal with the impact AI is starting to have on mobile networks.

Former Verizon CEO Hans Vestberg told an investor conference last year that the operator had been working for years toward this opportunity.

“We built the whole metro network and the edge capabilities five years ago, way early maybe, but that's really what we want to see when AI will start to have devices that want to connect with the network in a totally different way,” Vestberg told an audience at a Goldman Sachs Communacopia + Technology Conference.

“They're probably a couple of years out, but that is the next boundary of wireless growth for us. We're both going to have new offerings for these type of devices, and of course, the manufacturer of the device needs to have edge capabilities from us that we can charge as well.”

Toward this goal, Verizon in late 2024 signed a deal with Nvidia to power enterprise AI services and digital transformation efforts running over the carrier’s 5G private network and mobile edge compute (MEC) infrastructure. Adam Koeppe, SVP of technology planning at Verizon, touted the carrier’s architecture and its ability to support AI-derived use cases.

“I think it’s really important to look at what these capabilities exist within the architecture, and how can AI, true AI, augment things that are already being done or create things that are brand new,” Koeppe said. “Where I see our evolution occurring is when you have an advanced cloud platform, as we do, you have an orchestration layer on top that we already have, and you then find ways to incorporate new AI capabilities on top of that. That’s going to allow your engineers and your operators to interface differently.”

These new AI interface opportunities are expected to push further edge investments.

“As the focus of AI shifts from training to inference, edge computing will be required to address the need for reduced latency and enhanced privacy,” Dave McCarthy, research VP for cloud and edge services at IDC, wrote in a report last year. “This trend not only optimizes operation efficiencies but also fosters new business models that were previously not possible with centralized infrastructure. Distributing applications and data to edge locations enables faster decision-making with reduced network congestion.”

Jenny Lindqvist, SVP, head of cloud software and services at Ericsson, noted at a recent vendor event that it was a natural fit to explore AI inferencing at the edge, highlighting reasons such as sovereignty, privacy, and security.

"We're adding a colocation opportunity for AI application inferencing together with the user framework," Lindqvist explained. "So we're working here together with some of our infrastructure partners. We're working with Intel and Dell in order to see how we can cost-effectively colocate the user plane function together with the application inferencing at the edge."

Lindqvist added that this work was being done in a lab environment, where “we are running our user plane function (UPF) together with cloud radio access network (RAN) on a single server, and then in addition colocating with an AI application. Needless to say, that’s a very footprint-efficient way to deliver those types of workloads very far up in the edge.”

As the mobile industry moves toward 6G, opportunities start to open in terms of designing those systems from the start to deal with AI.

“We’ve got five years until potentially we launch [6G], but we are discussing exactly what it is and what it isn't and how we're going to go about it,” Tenorio said.

Ericsson noted that the wireless industry ecosystem is working to support this challenge with technology advances like carrier aggregation and advanced multiple-input, multiple-output (MIMO) antenna technology. These will allow for more flexible management of spectrum and link resources across uplink and downlink connections.

Dell’Oro Group VP Stefan Pongratz noted evidence of that MIMO flexibility in a recent column for SDxCentral, writing that, “Nokia recently shared results from initial pilots suggesting the upper 6 GHz spectrum could realize 75% of the C-Band uplink throughput gain.”

“This potential growth of uplink traffic underlines the importance of network capacity planning, spectrum allocation, and RAN feature developments,” the Ericsson report noted.

Uplink to space

This uplink challenge is exacerbated when looking at satellite connectivity.

SpaceX recently made cosmic waves when it filed plans with the Federal Communications Commission (FCC) to potentially launch up to one million satellites, each of which would operate as part of an “orbital data center system.” Each one of these satellites would operate at an altitude of between 500 kilometers and 2,000 kilometers and be “within orbital shells spanning up to 50 kilometers each.”

As with more traditional data centers, these orbiting platforms would likely require robust connectivity specifically tuned to deal with AI-generated traffic.

A separate Ookla test showed that the median uplink speed for Starlink’s current satellite broadband service was just under 17 Mb/s. Those tests also showed Starlink’s median multiserver latency of around 35 milliseconds.

Both of those results were significantly below performance results gleaned from ground-based cellular systems.

Other satellite-based broadband services tested performed worse. Viasat’s constellation, for instance, showed uplink speeds of less than 5 Mb/s, while EchoStar’s HughesNet constellation showed latency in excess of 685 milliseconds.

AT&T CEO John Stankey laid out that satellite-based connectivity challenge during a recent investor conference, explaining that the carrier and its traditional rivals T-Mobile US and Verizon each have access to approximately 300 megahertz of spectrum that can be transmitted from each cell site, which is more than triple the amount of spectrum the newly spectrum-filled SpaceX can provide from its satellite constellation.

“If you think about it, you get 80 megahertz you get to put on a spot beam that has a radius that is dramatically larger than a cell site – you know, two-and-a-half miles, two miles for a cell site, you're talking about a spot beam right now that's probably close to 20 miles, and you're taking 80 megahertz and you're putting over that period, you have a weaker uplink. That's a hard putt to be a like-to-like replacement,” Stankey said.

Stankey specifically pointed to limitations with “upstream bandwidth,” which “becomes more important in the AI dynamic.”

“It’s inherently going to be a more fragile upstream uplink than what you're going to get in a terrestrial network that gets on fiber really quickly,” Stankey said of a satellite upstream connection. “The faster you can get something on fiber, the more robust that upstream uplink is going to be, and the more low-band spectrum you have to do that, the more robust that upstream link is going to be. Satellite doesn't have those things.”

This article first appeared in the SDxCentral Magazine Issue #1

To read the full issue. Simply register.