After working on cross-border deals between Korea and the United States and moving to San Francisco, I have been following physical AI with one commercial question in mind: which parts of this technology can actually be sold into factories, and who gets paid for making them work?
I'm most interested in the companies between the model developer and the manufacturing customer. A model may become able to choose useful robot actions, but someone still has to connect it to the customer's machines and take responsibility for the finished installation. I want to understand what that work involves and how much of it can become a repeatable business.
For these notes, I'm using an imagined factory request: take an undamaged blue part from a shelf and put it on the production line. Following one task lets me see what each supplier contributes, without treating every company called "physical AI" as the same kind of investment.
What the factory needs from AI
Before looking at a new model, I want to know why the shelf task needs one. If the same part always arrives in the same position, a programmed arm can repeat a precise movement. That is already useful automation, and a more general model does not automatically improve it.
The case gets more interesting when the part's shape or position varies. Now the system has to interpret its sensors before choosing a movement, and learned perception and control may help. A vision-language-action model adds an ambition: that an ordinary-language instruction, together with what the camera sees, can guide the robot's motor commands.
For the imagined customer, I would ask what the new capability removes from the installation work. Does it reduce the need to program each variation? Can it identify the requested part and take its condition into account? Those answers tell me more about the product than placing it in the newest generation of robotics.
The suppliers behind an installation
To see who could earn revenue from the shelf task, I'm breaking down the work needed to deliver it. The model is one part of a system the factory must be able to operate and maintain.
| Part of the system | Contribution to the shelf task |
|---|---|
| Computing and simulation | Resources for developing, testing and running the software |
| Models | Interpretation of observations and instructions, or selection of actions |
| Integration software | Connections between models, robot controls and the factory's existing systems |
| Hardware | Cameras, sensors, grippers and the robot that moves the part |
| Finished application | A configured system that performs the production task and can be supported on site |
These categories sort responsibilities, not companies; one company can supply more than one part. The distinction I care about is who has to make the pieces work together when a customer places an order.
That is why I'm paying particular attention to integration software, sometimes grouped under middleware. The model's output has to become a movement the chosen robot can execute. Observations have to reach the right software in time, and the next machine has to be ready to accept the part. Much of the work of getting beyond a demonstration sits in those connections.
For me, the investment question is how often the same software can solve that problem elsewhere. If every sale requires a new engineering project, revenue growth may require a matching increase in the delivery team. Before treating middleware as a scalable product category, I want to know what gets easier on the second installation.
Vention and Tulip
One deployment claim I'm following is Vention's February 2026 announcement of GRIIP, its physical AI pipeline for manufacturing. The company reported consistent picking performance over three months of continuous, lights-out operation, while also saying it was expanding its customer demonstration program.
For my shelf example, that raises follow-up questions. Which parts did the system handle, under what conditions, and how often did someone intervene? I can take the supplier's reported result seriously and still need those details before judging how it would transfer to a customer's production line.
I'm looking at Mitsubishi Electric's investment and alliance with Tulip for a different reason. Tulip supplies software for frontline operations; it is not a robot foundation model. What interests me is the connection to an industrial equipment company, as a route to manufacturers that already organize their work around existing equipment.
So I ask different things of each business. With Vention, I want to understand the scope and repeatability of the reported picking performance. With the Tulip alliance, I'm interested in how the software reaches manufacturing customers. I wouldn't learn much by putting both under one "physical AI funding" total.
Manufacturing customers in Asia
My Korea-U.S. deal background is part of why I keep coming back to where the customers are. According to the International Federation of Robotics, Asia accounted for 74% of the world's 542,000 industrial robot installations in 2024.
I don't read that as a forecast for AI-controlled robots. It describes the existing equipment market, including machines that use conventional automation. It does make access to Asian manufacturers relevant to the commercial question I started with. A supplier has to work with the production environments customers already have, not only demonstrate a model where it was developed.
For the shelf task, I'm ultimately trying to identify the company a factory can hold responsible for a working installation, and then what that company can reuse at the next site. That is the opportunity I'm trying to understand through this value chain: a useful model capability becoming a product that manufacturers can buy repeatedly.