People that follow me say that I’m a skeptic when it comes to new applications and technologies. Recently I have been pretty negative about integrated sensing and communication (ISAC) and GPUs. I was the guy that predicted that femtocells would “flop” and that 6G would taste like 5G, just re-heated in the microwave.
But it’s no fun to be negative all the time. I’ve just finished a report, in partnership with Sebastian Barros, on the topic of The Economics of Kinetic Tokens. In this investigation, we were able to identify multiple physical AI applications that work outdoors over a mobile network, and that will require AI inference in the network.
Here’s the basic assumption that needs to be re-examined: Most robots are designed with on-board compute horsepower and AI models to work autonomously. The plan is for every robot to make its own decisions, 100% of the time. This is an extension of the self-driving car model, where a huge amount of compute is placed on board the vehicle, so that all important decisions can be made without relying on a network. For a $90,000 car, this architecture makes sense.
However, a giant computer is not the right answer for smaller robots. Over the past year, we’ve looked at automated consumer products like lawn mowers and snow shovels. Humanoid robots for the household are one step removed from these simple machines. We’ve also looked at food-delivery robots, security robots and drones for various enterprise applications. These platforms can’t support 2,500 trillions of operations per second (TOPS) of compute horsepower like a Tesla car, due to constraints on power, weight, size and cost.
These small robots will make most of the necessary decisions for operation on-board. A food delivery robot can drive down the sidewalk and make good decisions about avoiding pedestrians and potted plants. But periodically, these little robots will come across a situation that requires a bigger brain. When the food-delivery robot needs to cross a major four-lane street, its on-board AI model may be inadequate. Imagine a case where the robot pauses at the street corner, sends 10 frames of video somewhere for a “second opinion” and then crosses the street?
This is how we teach our children. When my daughter was three years old, the rule was that she must always hold my hand in crossing the street. Her brain simply was not capable of making the decision, so she submitted to parental control. Little food-delivery robots can do the same thing, relying on their on-board AI model 99% of the time and occasionally outsourcing the decision for safety. There’s no shame in it.
I believe that this distributed model of AI inferencing will be necessary for the explosion of physical AI.
AI inferencing for sale
The next question is: Who will sell this AI inferencing service to the millions of little robots?
Many developers are looking at this like cloud computing, where the data can be sent over the network and the big AI players will dominate the inference. That has been the case for GenAI, and for AI agents of various kinds. But in physical AI, the telcos have an opportunity to capture the AI inference market.
The telcos are in a much better position to identify the robot as an authorized user, to control its service based on its subscription terms and to guarantee a response in a reasonable timeframe.
Keep in mind that robots can be patient where people are not. A self-driving car could never stop at every intersection in order to wait 800 milliseconds for a decision. But a food-delivery robot can do that, a security robot can do that, and a construction robot can do that. Low latency is not the key driver here…the robot simply needs to get an answer every time within a reasonable 1-2 second timeframe.
We’ve used these examples as a straw man to consider what a kinetic token (K-token) will include, and to calculate the cost of delivering K-tokens as well as the potential pricing for K-tokens.
I won’t list the details here, but I will point out that a K-token will include authentication and security features, policy governance features, deterministic timing and an AI inference based on data from the robot (typically a few frames of video). There’s a range of pricing based on video inference factors and the level of latency required.
The revenue opportunity
The answers may surprise you. Telcos can add a rack of GPUs for every 20 million robots, in a centralized cabinet somewhere, and add hundreds of billions of dollars in new revenue.
Mobile Experts is estimating a potential market opportunity in the range of $250 billion or more. This could be a very profitable business for telcos, with strong differentiation over hyperscalers and AI players that will try to compete with them.
The alternative is for mobile operators to treat these little robots like IoT devices, allowing them to send their data up to a cloud AI model over the top. Yes, the operator can make a little money selling an API call to authenticate the robot and prevent fraud. And it can sell the uplink data connectivity. But that is a dumb business model.
The telecom operators have a lemon tree growing in their backyard. In fact, it’s the only lemon tree in town. Now, the cloud AI players are setting up a lemonade stand in the front yard. If the telcos sell network slices and API calls, it’s the equivalent of selling lemons for pennies so that somebody else can sell lemonade for dollars.
Telcos should get serious about physical AI and start playing hardball to keep the GenAI players from taking over.
Joe Madden is principal analyst at Mobile Experts, a network of market and technology experts that analyzes wireless markets. Disclaimer: Nokia is a client of Mobile Experts.
Opinions from industry experts, analysts or our editorial staff do not represent the opinions of Fierce Network.