Physical AI and Telecom Operators: Most Are Selling the Road, Not the Robot

Physical AI is the point where artificial intelligence stops producing words and starts moving things. A robot dog walks into a tunnel before a work crew does. A drone follows an emergency vehicle without a pilot. A humanoid picks up a box and puts it where a warehouse system told it to. In each case a machine senses its surroundings, an AI model decides what to do, and the machine acts in the physical world.

Telecom operators have good reasons to care. These machines need reliable connections, they generate heavy uplink video, and many of them will run part of their AI somewhere other than on the robot itself. That sounds like a natural fit for networks that already reach factories, ports, hospitals and homes.

TeckNexus has reviewed public evidence on physical AI activity at 40 telecom operators across the Americas, Europe, the Middle East, Asia-Pacific and Greater China. One pattern stands out: most operators are entering physical AI by selling the road the robot travels on, not the robot or the work it does. About a dozen of the 40 show a hands-on role in the robot application itself. For the rest, the role is connectivity, compute, investment or a stated ambition. See which six are building beyond it in our free physical AI brief →

That is not a criticism. It is a reasonable starting point. But it raises a strategic question every operator will need to answer: how much of the physical AI value chain can a connectivity provider actually capture?

What counts as physical AI for telecom operators

The term is used loosely, so it helps to be precise. In this analysis, physical AI means a machine that senses, decides with AI and then acts physically: it moves, picks, welds, flies, patrols or stops. The network may connect the machine, guarantee its performance or run part of its AI, but the machine has to do something in the real world.

That definition leaves out several things that often appear in the same announcements:

• AI that runs the network itself, such as AI-RAN features that improve spectral efficiency. These matter for operators, but no machine acts.

• AI data centres and GPU capacity on their own. Compute becomes part of physical AI only when it is used to train, simulate or run robot AI.

• Video analytics that only raise alerts. Cameras that spot a hazard and notify a person are valuable, but the action is still human.

• Vision statements. Saying robots will need 6G is a view of the future, not an activity.

Applying this test strictly is what makes the pattern visible. Once network-only AI and general announcements are set aside, the remaining operator activity is smaller, and much of it sits in the network layer rather than in the machine.

Physical AI roles telecom operators are taking today

The roles operators play fall into six broad groups. They are not mutually exclusive, and the most active operators already combine several.

The first two groups account for most of the activity. In more than six in ten of the operator records we reviewed, the operator’s stated role is to provide infrastructure. The middle groups, where the operator gets closer to the machine, are concentrated in a small number of companies, with China Mobile, China Unicom and HKT the most visible. The investment route is spread more widely and gives operators a stake in the robot layer without running robots themselves.

Connectivity: the most common physical AI role for operators

Connectivity is where most operators start, and it covers more than a standard mobile subscription. The examples in the evidence show three distinct products.

The first is the dedicated or sliced network. At Singapore’s Tuas Port, 5G network slicing supports automated guided vehicles, cranes and other port systems. In Beijing, China Unicom ran a separate 5G-Advanced network for the World Humanoid Robot Games, with a platform that showed each robot’s location and status in real time. In both cases the operator provides a network shaped around the machines rather than around people with phones.

The second is positioning. In Australia, Optus, Ericsson and FrontierSI paired an emergency vehicle with an autonomous tracking drone on a live commercial network, using 5G positioning and a dedicated slice. Knowing exactly where a machine is, without relying only on satellite signals, is a capability robots and drones increasingly need.

The third is global connectivity for machines that move or ship across borders. T-Mobile provides eSIM connectivity for Figure AI’s humanoid robots. That relationship began with a connection and has since extended to work on running parts of the robot’s AI closer to the network edge.

Compute: the second physical AI step for operators

Compute is the natural next step. Physical AI models need large amounts of processing for training and simulation, and some of the processing a robot does in real time could move off the machine if the network is fast and reliable enough.

Deutsche Telekom’s Industrial AI Cloud in Munich, built with NVIDIA, gives industrial companies sovereign capacity to train and simulate AI for robots and factories, and has been available to customers since February 2026. SoftBank’s AITRAS platform runs AI workloads inside its radio network and at the edge, and SoftBank has shown robots offloading AI tasks to that environment with Ericsson. T-Mobile is piloting edge AI servers for physical AI applications with partners.

The technical case is real. Robots that stream video for real-time control put unusual pressure on the uplink, which networks were not originally built for. Research from Nokia Bell Labs suggests that keeping end-to-end delay at around 20 milliseconds for almost all packets can require provisioning three to four times the average uplink video rate (Nokia Bell Labs, April 2026). That is a network design problem, and operators are the ones equipped to solve it.

Much of this compute work is still at laboratory or early trial stage. That is normal for a new architecture, and it means the commercial shape of robot AI at the network edge is still being decided.

Why the network is the natural entry point for physical AI operators

Starting with the network plays to what operators already do well. They own spectrum, sites and fibre. They run edge facilities close to customers. They have enterprise sales teams that already sell private 5G to manufacturers, ports and logistics groups. Physical AI adds new demand on top of those assets without asking the operator to become a robot company.

There are also good reasons not to rush further up the stack. Robot hardware is capital intensive and changes quickly. Robot AI models are improving at a pace set by specialist developers and chip companies. Integrating robots into a specific factory or warehouse takes domain knowledge that operators do not always have. Selling the road keeps risk manageable while the market settles.

Where the margin sits in the physical AI value chain

The difficulty is that connectivity is the part of the physical AI stack where pricing is most familiar to buyers. Enterprises know what a SIM, a slice or a private network costs. The value of a robot, by contrast, is measured in the work it does: inspections completed, totes moved, injuries avoided, hours of labour released.

That value tends to flow to whoever owns the outcome. Today that is usually the robot maker, the AI model provider or the systems integrator that designs the workflow. An operator that only supplies the road risks seeing robots multiply across its network while its share of the value stays flat.

There is a middle ground. Differentiated connectivity, where an operator guarantees latency, reliability or positioning accuracy for a specific machine workload, can be priced on performance rather than volume. Network slicing for port equipment and dedicated networks for robot events are early examples. It is still the road, but a toll road with a service level attached.

The layers of the physical AI value chain, and how close each sits to the operator’s current business, look like this:

Three ways physical AI operators are moving closer to the robot

The operators getting closest to the machine are following three distinct routes. None requires an operator to manufacture robots.

• Become the platform that manages many robots. China Unicom’s Gewu platform connects, controls and schedules more than 30 robot types and is offered as a service; the company reports more than 70 customers. A platform position sits between the robot maker and the enterprise, which is where scheduling, data and service revenue accumulate.

• Package a complete solution for one vertical. HKT, working with YuShu Technology, deployed a quadruped patrol robot with vision AI at a Hong Kong concrete and aggregates provider and offers the package with installation, training, support and maintenance. Telefónica is validating drones as a service for mast and critical-site inspection, with human supervision throughout. The operator sells an outcome, such as a safer site or a completed inspection round, not just a connection.

• Use robots in the operator’s own business first. Vodafone is trialling an AI-steered robotic arm in Tirana that re-aims a live mast antenna in 20 to 30 minutes, work that otherwise needs a site visit. Optus and Nokia used a robot dog to measure indoor coverage automatically. China Mobile runs a coffee-making robot on its own robot software at its Hangzhou research park. Internal use builds skills, data and reference cases before an operator takes an offer to customers.

The three routes can be combined. China Mobile, for example, uses robots in its own service centres, has launched a consumer robot product and has released robot software as open source. Telstra offers a different combination: with construction group John Holland it trialled Astro, a 5G-connected robot dog that walked into confined underground spaces on Melbourne’s Metro Tunnel Project and sent gas and air-quality readings before workers went in.

What physical AI operators need to build next

Moving up the stack is less about technology than about capabilities. The evidence points to four that separate the operators getting closer to the machine:

• Robot integration skills. Connecting robots from several makers to one management layer requires engineering teams who understand robot software, not only network software.

• Vertical partners. The clearest operator offers pair the network with a robot maker and a customer in one industry, such as construction safety, port logistics or site inspection.

• Performance-based network products. Slices and dedicated networks priced on guaranteed latency, uplink or positioning turn network quality into a sellable feature for machines.

• A data position. Robots generate operating data that improves models and services. Operators that manage robots, rather than only carry their traffic, can build on that data.

What to watch as telecom operators build physical AI businesses

Three signals will show whether operators can move beyond the road over the next 12 to 18 months:

• Robot workloads on live networks. Several AI-RAN and edge trials are moving from the lab towards field validation. The step that matters is robot AI running on operator infrastructure at a customer site, not only network performance figures.

• Priced service levels for machines. Offers that charge for guaranteed latency, uplink or positioning for robots and drones would show operators turning network quality into revenue.

• Named customers and measured results. Most physical AI announcements across the market today describe capabilities and pilots. Published outcomes from named enterprise customers will mark the move from experimentation to business.

Physical AI and telecom operators: frequently asked questions

What is physical AI in telecoms?

Physical AI refers to AI-driven machines such as robots, drones and autonomous vehicles that sense their environment, decide what to do and act physically. In telecoms, the term covers the networks, edge computing and services that connect and support these machines.

Why do robots need telecom networks?

Robots that move across large sites, stream video or offload AI processing need reliable wireless connections, strong uplink capacity, low delay and, increasingly, precise positioning. Private 5G, network slicing and edge computing are designed to provide these.

How can telecom operators make money from physical AI?

Operators can sell connectivity and positioning, edge and cloud compute, robot management platforms, complete vertical solutions or performance-based network services. The further an operator moves from connectivity towards outcomes, the more of the value it can capture, and the more partners and skills it needs.

Which telecom operators are most active in physical AI?

Public evidence shows the most visible activity at a small group of operators in China, Hong Kong, Japan, South Korea, Europe and the United States, including China Mobile, China Unicom, HKT, Deutsche Telekom, SoftBank, T-Mobile and Vodafone. Most other operators are at the connectivity, investment or planning stage.

Physical AI will need networks, and operators are right to start where their strengths are. The open question is whether they are content to carry the traffic or want a share of the work the machines do. The answer will shape which operators are still talking about physical AI as a connectivity opportunity in three years, and which are talking about it as a services business.

The road, not the robot: Most operators still sell the road. See which six are building beyond it in our free physical AI brief →

Related reads: Physical AI and telecom operators

Can the Network Be the Robot’s Brain? Edge AI, AI-RAN and the Case for Offloading Robot AI
What it takes to run part of a robot’s AI in the network, and what operators are testing.

Physical AI Business Models for Telecom Operators: Who Pays, for What, and How
The offers operators can sell, from machine connectivity to packaged robot services.

Physical AI in China: Why Chinese Telecom Operators Are Furthest Ahead in Putting Robots to Work
How China’s operators put robots to work, and what other operators can learn from them.

NVIDIA and Telecom Physical AI: Why One Platform Sits at the Centre, and How Operators Can Build on It
NVIDIA’s role across AI-RAN, operator AI clouds and robot AI, and how operators can build on it.