Physical AI's network demands are telco opportunities

• Putting sufficient compute on a robot drives unit cost prohibitively high, making network-based inference the only economically viable path
• Mobile Experts forecasts uplink traffic from physical AI could outstrip mobile network capacity by 2029, with demand rising 20-25% annually through the 2030s
• But will telcos be able to capture business value from new AI demands?

Robots and other physical AI will require network capacity beyond the capabilities of today's networks, creating bottlenecks to development and telco opportunities, said Joe Madden, founder and chief analyst at Mobile Experts.

Today's physical AI runs on the device — for example, in a delivery robot or self-driving car. But that's not sophisticated enough. Practical, physical AI will require intelligence on the network, demanding reliability and performance greater than today's networks can deliver, Madden said.

Diagram showing four AI inference locations — on-device, 5G tower sites, enterprise premises and central data center — with decision speeds from microseconds to seconds.

Physical AI pushes mobile network capacity

"Mobile Experts estimates that uplink traffic demand could outstrip capacity in 2029," the research organization said in a report: "The Impact of AI on Mobile Telecom." Physical AI will drive spectrum requirements in the 2030s, with uplink traffic rising 20-25% annually.

"Today, AI models run in two locations: on smartphones and other devices, and in large centralized data centers. With the rise of Physical AI, things are likely to get more complicated. Robots that deliver food, assemble cars, or fold laundry will require rapid decision-making. New localized AI computing resources will be needed," the report states. 

Chart showing AI-driven traffic load rising sharply above the more gradual capacity boost AI provides, projecting a network capacity gap over time.

Today's robots are too dumb to be useful — and too expensive to get smarter

The limitations of today's physical AI is on display in YouTube videos of these robots falling over and getting stuck, Madden said. "A robot running a marathon tripped on something tiny in the road and broke itself into pieces," Madden said. "You can go on YouTube and watch videos of food delivery robots that do silly things — drive in front of a truck or fall into a hole."

For example, this video from the New York Post demonstrates robot limitations and potential.

"Dozens of humanoid robot runners competed in the Beijing half-marathon to mixed success," according to the Post. "While a Chinese-built robot named Lightning comfortably beat the human half-marathon world record Sunday [April 19] during an annual race in Beijing, several other humanoid participants malfunctioned and fell, including one bot which had to be carried off in pieces at the very start of the race."

And here's a TikTok video of delivery robots failing, including one that seems to be taking its own life by throwing itself down a flight of stairs.

Businesses are experimenting with using robots in the real world for simple work, and the AI models and decision-making are not up to par, Madden said. "They have to get a lot smarter."

On-device AI makes cheap robots expensive – the network is the alternative

Smarter robots require more computing power and a bigger battery. "You need a $10,000 compute engine and sophisticated AI model, with storage and memory. Your $500 food delivery wagon turns into a $50,000 food delivery wagon," he said.

For that kind of investment, that better be a great burrito.

"For machines costing less than $2,000, automation may require either inference or training data to be handled in the network," according to the analyst report. "For safety reasons, decisions must be very quick, but the low-cost platform won’t be able to adequately analyze its environment."

Offloading AI to the network only works if operators can guarantee performance

Safety isn't the only concern. "If the food delivery bot can't tell the difference between oncoming traffic and leaves blowing in the wind, it can stop and wait. But then your food gets cold or it never gets there, so that's not a good solution," Madden said.

To keep cost down, AI will be located on the network. But network operators need to be able to guarantee performance and reliability. "That's a challenge for operators. It's difficult to guarantee radio signals go everywhere," Madden said.

Telcos are already positioning themselves to seize the opportunity. For example, T-Mobile and Nvidia are working with Nokia and a growing developer ecosystem to deploy physical AI applications over distributed edge networks.

Larger business customers can set up private 5G networks and on-premises edge computing racks to meet their needs. This trend is already well underway for industrial enterprises. "Many of these customers choose to buy their own networks and computing racks because they insist on controlling and protecting their data very tightly," according to the report.

Hyperscale, centralized data centers will be the location for the most sophisticated generative AI models with trillions of parameters, but response times can be slower from this location, said the report.

Physical AI could revive edge computing but telcos have been down this road before

All of this has echoes of Mobile Edge Computing (MEC), a technology and business model from the early/mid-2010s that failed to generate significant revenue for telcos. "This time around, the question is whether AI will create a market for edge computing. So far we don't see this happening," Madden said.

It might be 2035 until the market matures sufficiently to justify compute resources at the base of every tower, the analyst said.

And when AI on the RAN matures, will telcos benefit? Madden said it's an open question whether enterprises buy AI services from their telecom provider, from a hyperscaler in a centralized data center, or own and operate the intelligence themselves.