We covered the brain. Now comes the body.
For two years this publication has been obsessed with the brain of the AI economy. The accelerators that do the math. The data centers that house them. The grid and the reactors that feed the whole hungry machine. That coverage was right, and it still is. But it describes only half of an economy. A brain with no hands is a very expensive way to produce text and pictures.
The category I want to introduce today is the other half. Robotics and autonomy is the layer where a trained model stops being a chat window and starts moving pallets, driving down a sidewalk, picking an order, or pouring a drink. It is the part of the buildout that touches things you can drop on your foot.
Here is the thesis in one sentence, and the rest of this piece is me defending it: the same breakthroughs that made software models suddenly useful are now flowing downhill into physical machines, and the market has not finished repricing what that means.
A brain with no hands is a very expensive way to produce text and pictures.
I want to be careful and honest about this, because robotics has burned investors before. The 2015 to 2020 era was littered with autonomy companies that promised a self-driving world and delivered press releases instead. So my job in this feature is not to sell you a dream. It is to separate the part of this category that is becoming real and ownable from the part that is still a story.
Why now, and why it ties to the AI buildout
The reason robotics keeps failing to live up to its hype is that the hard part was never the motors. We have had good actuators and decent batteries for a while. The hard part was perception and decision-making — letting a machine look at a messy, unstructured world and figure out what to do. That is exactly the problem that modern AI models got dramatically better at.
Three things changed at roughly the same time. Vision models got good enough to identify and locate objects reliably in cluttered scenes. The cost of the perception stack — cameras, radar, and especially lidar — fell hard as volume rose. And the data-center capacity to train control policies on enormous amounts of real and simulated motion finally existed, because we had already built it for everything else.
This is the honest tie to the AI buildout, and I will not stretch it further than it goes. Robotics is not riding the buildout because of a slogan. It is riding it because the literal inputs are shared: similar model architectures, the same kind of compute used to train policies in simulation, and the same collapse in the cost of running a model that lets a robot carry a capable brain on board without a server rack strapped to its back.
The hard part was never the motors. It was teaching a machine to understand a messy world.
There is also a labor reason that does not depend on any forecast. Warehouses, last-mile delivery, and food service all face structural shortages of people willing to do repetitive physical work at the wage on offer. When the cost of a machine that runs for a known amount per year crosses below the cost of labor that is scarce and rising, adoption stops being about enthusiasm and becomes about a spreadsheet. We are at that crossover in a few narrow verticals right now. Not everywhere. A few.
How I am drawing the map
I split this category into four layers, because lumping them together is how people lose money. Each has a completely different risk profile.
The first is structured-environment automation. These are robots that work inside four walls where the company controls the lighting, the layout, and the rules. This is the most real layer. The problem is bounded, the customer is a business with a budget, and the payback is measurable. Warehouse automation lives here.
The second is the perception and sensor layer. This is the picks-and-shovels play. Every robot, in every other layer, needs eyes. A company that sells the eyes does not have to bet on which robot wins. It is the most leveraged-to-the-whole-category exposure you can get, and historically the shovel-seller is often a steadier business than the prospector.
The third is unstructured-environment autonomy — sidewalk delivery, public-road driving. This is the hardest layer and the one with the longest history of broken promises. The world outside is endlessly messy. I treat everything here as speculative until the unit economics are proven at scale, not in a pilot.
The fourth is the humanoid layer. This is the one generating the loudest headlines. I think it is real eventually and wildly early today. The honest position is that the largest humanoid programs sit inside giant companies where the robot is a rounding error on the valuation, so there is no clean small or mid-cap way to own it yet. I would rather say that plainly than invent a pure-play that does not exist.
The picks, and the discipline behind them
My favorite single idea in this category is the perception layer, and the reason is structural. If I am right that robotics is a real growth wave, the sensor maker can win whether the winning robot is a warehouse arm, a delivery cart, or something not yet invented. That is the position I want when I am confident about a trend but humble about which player captures it. It is the cleanest expression of the thesis without betting the farm on one form factor.
On structured automation, I have enormous respect for the leading warehouse-automation name. It is the rare robotics company with real revenue at scale, a serious backlog, and brand-name customers. My only hesitation is price. When a stock has already made a large move and the valuation prices in years of flawless execution, the business can be excellent and the stock can still disappoint. That is also why I watch its customer concentration closely. So it sits as a Watch rather than a Buy for me. Quality is not the question. Entry is.
The sensor maker can win whether the winning robot is a warehouse arm, a delivery cart, or something not yet invented.
The delivery and service-robotics names are where I get most cautious, and I want to be direct about why. These are real products. There are real robots on real sidewalks and real machines pouring real drinks. But a real product is not the same as a real business. Revenue can grow quickly and still be tiny against the cash being spent, the unit economics still have to prove out, and several of these stocks move on partnership headlines that run well ahead of the income statement. I keep them on the radar because the optionality is real, but I size them like the lottery tickets they currently are, and one of them I would simply pass on at today's setup.
How to hold this category without getting hurt
Robotics rewards patience and punishes the impatient with unusual cruelty, because the timelines are long and the narratives are intoxicating. So here is how I actually intend to hold it.
I anchor the position in the layers with bounded problems and real customers — perception and structured automation — and I let the speculative layers be small, deliberate bets that I can be wrong on without it mattering to the whole portfolio. I treat any company whose stock moves more on letters of intent than on revenue as a story until the story shows up in the numbers. And I keep reminding myself that being right about the category and wrong about the timing is the most common way to lose money in a real trend.
The label I keep coming back to is anti-fragile. The companies I want most are the ones that get stronger as the field gets messier — the sensor maker that can sell to every camp, the automation provider whose value rises every time labor gets scarcer. Those are the businesses that tend to hold up when the hype cycle inevitably turns, and the hype cycle in robotics always turns at least once before the real thing arrives.
This is the body of the AI economy. We spent two years on the brain because that is where the money went first. The money is starting to move to the hands. I would rather be early and disciplined here than late and excited.
This article is opinion and content only. It is not financial advice, and the author and publication accept no liability for decisions made based on it. Do your own research.