Now Is The Time for The Middle Market to Start Paying Attention to Physical AI
Last week I attended AUTONOMOUS 2026 in San Francisco, a gathering of about 700 founders, operators, and investors building the physical AI economy. Leaders from companies in autonomous trucking, humanoid robots, agricultural autonomy, drones, prosthetic hands training robot dexterity, and mobility devices for the blind all shared the same stage.
Let me get the obvious out of the way first. Almost nothing at this conference is a middle-market PE target today. These are venture-backed companies burning massive capital to solve hard technical problems, most years away from material EBITDA. Walden Robotics announced a $300 million seed round the day before the event. That is not a middle-market profile.
So why should middle-market investors care? Because the physical AI wave will reshape the industrial and operational technology companies that middle-market PE already owns and buys. The winners of this transition will not only be the robot makers. They will be the component suppliers, the safety and trust infrastructure providers, the integrators, the incumbents who partner well, and the industrial businesses that adopt early to solve labor problems their competitors cannot. Those are middle-market stories, and they are forming now.
Here is what I took away.
The bottleneck is data, not hardware
The most consistent technical theme of the conference: hardware is largely solved, and data is not. Compute is sufficient. Sensors are good enough. Simulation platforms are maturing. What every company is fighting for is training data that reflects the messy, variable physical world.
SkilldAI’s founder walked through the industry’s fifteen-year search for scalable data, from curiosity-based exploration to teleoperation to simulation to human video. His conclusion was that there is no golden path, so you need all of them. Industrial Next’s founder went further, arguing that the AI model itself is only 10 to 15 percent of what you need to run an AI-powered production line. The rest is manufacturing know-how.
Hardware is solved, data is not. The process data and field records inside current industrial portfolio companies become more valuable as physical AI matures.
The investor takeaway: proprietary operational data and deep domain expertise are the scarce assets. This favors incumbents and operators over pure technology plays, which is exactly where middle-market PE lives. A portfolio company sitting on years of process data, field service records, or machine telemetry may be holding an asset that becomes far more valuable as physical AI matures. One speaker made the point sharply in the semiconductor context: all the IP is in the process data, and that data will need to live in sovereign, customer-specific models. Companies that own and control valuable operational data will have leverage. Companies that gave it away will not.
The economics live in removing the driver, not building the robot
Torc, the autonomous trucking company Daimler acquired in 2019, was refreshingly direct about where the profit pool sits: in the autonomy itself, in removing the driver from Class A trucking. A truck runs 200,000 miles per year and lasts a million miles. The autonomous driver can be sold as a subscription on an autonomy-ready truck, or wrapped into transport-as-a-service.
This is the pattern middle-market investors should watch across every physical AI vertical. The hardware is a delivery mechanism. The recurring revenue attaches to the autonomy, the software, and the service layer. When these companies eventually mature into buyable businesses, the ones with subscription and service models attached to installed fleets will command the multiples. The same logic applies to portfolio companies today: any industrial business that can convert a hardware sale into a recurring autonomy or intelligence subscription is repositioning itself up the value curve.
Trust is becoming an industry of its own
If I counted the number of times the word “trust” was spoken from the stage, it would beat any technical term. Regulators need to trust the safety case. Customers need to trust the brand. The public needs to trust the robot on the sidewalk. Kubota’s team was explicit that a 135-year-old brand cannot afford to betray farmer trust with an autonomous tractor, because the farmer’s livelihood depends on it.
Trust is not just a talking point. It is spawning real businesses. Fort Robotics is building safety-certified control components and already has 650 customers and 20,000 units in the field. Koop is writing insurance for robots. Simulation-based safety validation platforms are expensive but improving fast, because you cannot road-test your way to safety on edge cases.
The enabling layer extends beyond safety. Point One Navigation, a San Francisco company that raised a $35 million Series C late last year, sells centimeter-level positioning that autonomous trucks, tractors, and field robots depend on, and its business scales across verticals, regardless of which platform companies win.
For middle-market investors, this is the picks-and-shovels layer worth tracking closely. Safety certification, functional safety components, testing and validation, insurance, and compliance infrastructure will scale with the entire industry regardless of which robot makers win. These businesses look a lot more like traditional middle-market industrial technology companies than the humanoid startups do, and they will get to profitability sooner.
Labor scarcity is the demand engine, and it makes robots a painkiller
The deployment panel crystallized something important about demand. Agriculture has fields going unharvested for lack of labor. Construction faces persistent shortages worsened by immigration constraints. Trucking has been short drivers for a decade. The framing across the board was not replacement but fulfillment: robots filling work that is going undone, and moving people out of dirty, dull, and dangerous jobs.
Trust infrastructure is likely the first commercially viable layer. Safety components, validation, and robot insurance will hit real EBITDA before most robot platforms do.
One panelist put it in classic terms: make a painkiller, not a vitamin. Physical AI companies selling into acute labor scarcity have a fundamentally different sales motion than those selling productivity gains. Carbon Robotics (laser weeding), Built Robotics (autonomous retrofits for solar construction), and Dusty Robotics (printing construction plans on site) are all painkiller businesses. When evaluating any company in this space, the first diligence question should be whether the customer is buying relief from a problem they feel today or an improvement they can defer.
Partnership beats vertical integration, and that opens doors for incumbents
Robotics has historically been vertically integrated, with each company building hardware, software, and AI in-house. The panelists agreed that a more horizontal industry structure is coming. The Kubota and Agtonomy partnership is the template: Kubota brings 135 years of manufacturing, distribution, and brand trust, while Agtonomy brings the autonomy stack. The autonomous tractor supports over 200 implements and works day and night, something neither party could have delivered alone.
This is directly relevant to middle-market portfolios. Established industrial OEMs and equipment makers do not need to build autonomy from scratch. They need to be the trusted manufacturing and distribution partner that an autonomy company plugs into. PE owners of industrial equipment businesses should be asking now which autonomy partnerships would extend their products’ relevance, because the startups need exactly what these incumbents have: manufacturing scale, service networks, and customer trust.
Middle market deals in physical AI don’t need to build autonomy. They need to be the manufacturing, distribution, and service partner the autonomy startups can’t live without.
Retrofit is the related theme. Torc is exploring retrofitting existing trucks. Built Robotics retrofits heavy equipment. Retrofit models raise real technical and legal complications, but they also mean the installed base of industrial equipment, much of it owned or serviced by middle-market companies, becomes the deployment surface for autonomy rather than being displaced by it.
Timelines are long, which is actually good news for the middle market
A dose of realism ran under everything. Waymo took 20 years from the original X challenge to commercial deployment. Torc was acquired in 2019 and is targeting driver-out testing next year with commercial launch around 2030. When a panel was asked to predict when a robot would be doing chores in your house, one answer was 20 years.
For venture investors, long timelines are a financing risk. One VC on the capital panel made the point that being ahead of the market can kill you: raise too much too early and you burn through it before the market gives you the traction to raise the next round.
For middle-market investors, the long timeline is an advantage. You do not need to underwrite technology risk at seed stage. You need to watch the space now, build theses on which layers will consolidate, and be ready when these companies, or more likely the enablers and adopters around them, grow into real EBITDA. The safety component makers, the integrators, the data-rich incumbents, and the early industrial adopters will hit middle-market profiles years before the humanoid platforms do. The buyers who understand this ecosystem before the banks start running processes will have the edge.
The bottom line
Physical AI is not yet a middle-market asset class, but it is already a middle-market force. It will change what industrial data is worth, which business models command premiums, where recurring revenue attaches to hardware, and which incumbents stay relevant. The companies on stage in San Francisco are building the future. The companies in your portfolio are going to live in it. The work of understanding this industry starts well before the first deal does.
