Physical AI: Who Owns the Robot Floor

Robots are moving out of fenced industrial cells and into spaces they share with people: operating rooms, distribution centers, mixed factory aisles and, later, homes. The commercial question is who earns money once that happens. Our answer is the owner of the installed base, the party that keeps each machine working, billed and learning after the sale. This report tests that claim against the latest industry statistics and company filings, and finds it holds in one sector, is only half built in another, and is mostly a forecast in a third.

The Short Version

We started with four propositions: robots are shifting from caged cells to open spaces, installed base and recurring revenue per robot are the economics that matter, deployment and service capacity limit growth, and robot cost against the cost of the labor it replaces sets the pace. The evidence backs the second and third strongly and the first only weakly so far.

The International Federation of Robotics (IFR, the trade body that publishes the standard count of industrial robots) reports that more than 600,000 industrial robots were installed in 2025, taking the operating stock to 5 million. Almost all are conventional arms doing fixed tasks in factories. The open-space cases that exist at scale are surgical robots and warehouse systems, and they differ sharply in recurring revenue per machine.

So the robot floor is opening slowly, and the companies that own it combine an installed fleet with a usage-linked revenue stream and a service organization. Intuitive Surgical has that today. Symbotic has the fleet commitment but not yet the recurring layer. Teradyne’s robotics unit has the hardware but a small share of its parent. Tesla has a stated ambition and a factory line, with no disclosed units or revenue.

The Factory Is Still the Market

Start with what is actually installed. The IFR counted 542,076 industrial robots installed worldwide in 2024, against 541,302 in 2023 and a record 552,946 in 2022. The operating stock reached 4,663,698 units, up 9%. In 2025 the IFR reports installations rose 11% to more than 600,000, and the stock rose 9% to a record 5 million.

That outcome beat the IFR’s own forecast. A year earlier it expected 2025 installations to grow 6% to 575,000 units, with 700,000 passed by 2028. The first of those numbers was exceeded within a year. Growth in factory robots is not a story waiting for humanoids to start.

Geography is concentrated. China installed 295,045 robots in 2024, 54% of the world total, and its operating stock of 2,027,190 units was 43% of the global stock. Japan installed 44,453, the United States 34,164, South Korea 30,596 and Germany 26,982. Those five markets took 80% of installations.

The customer mix also says something about where robots work. In 2024 electronics buyers took 128,899 units and automotive buyers 126,088, with metal and machinery third at 88,777. Automotive demand fell 7% as electric vehicle plans slowed. These are structured environments: parts arrive in known positions, and people stay outside the work cell.

Density tells the same story from the labor side. The IFR puts the average at 177 robots per 10,000 manufacturing employees in 2024, with 204 in Asia, 148 in Europe and 131 in the Americas. The United States installed 34,164 units in 2024, down 9%. Even in a country with high wages and loud interest in automation, the factory installation count was falling in the latest full year the IFR has reported in detail.

The IFR also names the binding constraint, and it is not the robot. In its executive summary it says a developed ecosystem of system integrators is key and “quite often the bottleneck”, because vision, process design and peripherals are needed before an arm earns anything. That observation supports the service-capacity half of our thesis directly.

Why the Floor Is Opening

Three pressures push robots toward shared spaces. Labor is the first. The IFR lists robots addressing labor shortage and new customer segments beyond manufacturing among its top market trends for 2025. Its service-robot sample, covering 293 of 944 manufacturers and not projected to the whole industry, counted 199,000 new professional service robots (up 9%) and 16,700 new medical robots (up 91%) in 2024. The sample caveat matters, and the IFR warns against comparing survey waves, so these are direction markers and not market sizes.

The second pressure is software. Mobile fleets can now be coordinated as a system. Amazon says it has deployed its one-millionth robot across more than 300 facilities, and that its fleet-routing model improves fleet travel time by 10%. A million machines moving through shared aisles is the clearest large-scale example of robots working in spaces designed around people.

The third is the labor price itself. The U.S. Bureau of Labor Statistics (BLS) reports median pay for hand laborers and material movers of $18.38 an hour, or $38,220 a year, in May 2025, across about 6.9 million jobs. It projects 4% employment growth for the group from 2025 to 2035. Automation shows up in specific tasks, such as machine feeding, which the BLS expects to keep shrinking. In aggregate, the largest pool of work robots might replace is still growing.

Amazon’s own page shows a shift in the labor mix alongside any reduction. It says its Shreveport facility needs 30% more employees in reliability, maintenance and engineering roles. A fleet that works among people creates demand for technicians. This is the service-capacity constraint again, seen from the customer side.

Installed Base Economics: Surgery Is the Reference Case

Surgical robotics is the sector where robots already work in shared, high-stakes space, and Intuitive Surgical’s filings show what the economics look like when the model works. At September 30, 2026 the company had a da Vinci installed base of 12,094 systems, up 12% from 10,763 a year earlier, plus an Ion lung-biopsy installed base of 1,152 (up 21%). In the third quarter it placed 546 da Vinci systems, 315 of them the newest da Vinci 5.

Those machines were used. The company performed approximately 906,000 da Vinci procedures in the quarter, up 14% from about 797,000. By our arithmetic that is roughly 79 procedures per system on the average installed base over the quarter. Chief executive Dave Rosa said customers performed over 10,000 procedures per day on average across the period.

The pricing structure is what makes the base valuable. Per its 10-Q, a da Vinci system generally sells for between $0.6 million and $3.1 million, while the company earns between $900 and $3,700 of instruments and accessories revenue each time a procedure is performed. Instruments wear out, so each procedure re-orders from Intuitive. Service contracts add a third stream.

The revenue mix reflects it. In the nine months to September 30, 2026, instruments and accessories brought in $5,218.7 million and services $1,388.3 million, against $2,104.8 million from systems, out of total revenue of $8,711.8 million. Instruments plus services were therefore 75.8% of revenue by our calculation.

Intuitive has also moved from selling machines to meter-reading them. Of the 546 systems placed in the third quarter, 291 went out under operating leases, and 160 of those were usage-based, up from 115 usage-based placements a year earlier. In its 10-Q the company says variable lease revenue, primarily usage-based arrangements, was $177.8 million in the quarter compared with $138.3 million. A hospital that pays per procedure carries less capital risk, and Intuitive earns more as utilization rises.

The model is not frictionless. Intuitive recorded a $450 million charge in the quarter to settle antitrust class action litigation, which GAAP (generally accepted accounting principles) net income absorbed. The 10-Q also says the company expects to cut instrument prices in China in response to a national pricing framework. A dominant installed base attracts both legal and pricing pressure. Procedure growth guidance of 14% to 15% for 2026 is the number to watch against that pressure.

Warehouses: A Large Backlog on a Thin Recurring Layer

Warehouse automation is the second large open-space deployment, and Symbotic is the cleanest public example. The company reported third-quarter fiscal 2026 revenue of $721 million, up 22%, with net income of $55 million for the quarter ended June 27, 2026. It had 77 systems in deployment. Its 10-Q reports backlog of approximately $22.5 billion, of which the agreements with Walmart and Exol comprise the vast majority. Exol is the warehouse-as-a-service joint venture formerly called GreenBox.

The backlog is large relative to revenue. Symbotic’s fiscal 2025 revenue was $2,246.9 million, so the backlog is roughly ten times that figure. Concentration is the other side of it: Walmart accounted for approximately 85% of total revenue in fiscal 2025. In January 2025 Symbotic also bought Walmart’s Advanced Systems and Robotics business and signed a master agreement for automated pickup and delivery systems at Walmart stores.

The recurring layer is where Symbotic differs from Intuitive. In the nine months to June 27, 2026, systems revenue was $1,895.7 million, software maintenance and support $36.6 million, and operation services $95.0 million, on total revenue of $2,027.3 million. The two recurring-type lines were 6.5% of revenue by our calculation, against 75.8% at Intuitive. The filing says 56 operational systems are under software maintenance.

The gap is partly timing. Symbotic is early in a build cycle, and each system it completes adds to the base that later generates software and service revenue. Both of those lines grew faster than systems: software and maintenance rose 80% and operation services 33% in the nine months, against 23% for systems. The Exol structure points the same direction, since a warehouse-as-a-service arrangement charges for outcomes. But the ratio today shows a company paid mainly for installation, with one customer making up most of the revenue.

Cash is the related constraint. Symbotic ended the quarter with $1.7 billion in cash, down from $2.0 billion a quarter earlier, while guiding fiscal fourth-quarter revenue of $760 million to $780 million. Building systems ahead of revenue recognition consumes working capital. The evidence therefore supports the deployment-capacity half of the thesis (77 systems in progress is a large installation workload) more than the recurring-revenue half.

Collaborative Robots: The Slow Door

Collaborative robots, or cobots, are designed to work beside people without a cage. They are the clearest bridge between the fenced factory and the shared floor, so their adoption rate is a useful test of how fast the floor is opening. The IFR said in 2024 that collaborative robots were 10% of industrial robots installed in 2023, with cobot installations declining 2% for the first time. Its 2025 presentation shows double-digit cobot installation growth of 12% for 2024, though we read that figure from a chart slide and not a table.

So cobots recovered, and they remain roughly one robot in ten. They have not displaced caged arms. A plausible reason is that a cobot gives up speed and payload to work without a fence, though the IFR data we reviewed does not test that.

Teradyne owns the category leader, Universal Robots, alongside the mobile robot maker MiR, and reports them together as its Robotics segment. In the third quarter of 2026 Robotics revenue was $96 million of $1,306 million total, about 7% by our arithmetic. The company said the segment delivered its highest third-quarter revenue and achieved profitability. The 10-Q attributes the quarter’s $20.7 million (27.6%) increase to higher sales of collaborative arms and autonomous mobile robots, and a $68.1 million (31.1%) increase over nine months.

Teradyne’s filings do not disclose the size of the Universal Robots installed base or any recurring revenue per arm, so we cannot place it on the installed-base spectrum between Intuitive and Symbotic. That gap is itself informative. Without that disclosure we cannot tell whether the vendor owns the service contract and the usage data, or whether the integrator does. The IFR’s remark about integrators suggests the latter is at least common.

Humanoids: Disclosed Facts and Company Targets

Humanoids are the form of robot most closely tied to the open-floor thesis, since a human-shaped machine fits spaces and tools built for people. They are also where the gap between disclosure and ambition is widest. We separate the two.

The IFR’s judgment is cautious. Its 2025 presentation lists “single purpose humanoids” among top trends with few commercial deployments announced, and states that true multipurpose humanoids are far off yet.

Tesla’s disclosed facts are narrow. Its third-quarter 2026 update says Optimus production began in Fremont on high-volume lines, that the output is the first production-quality robots the company has built, and that a Texas Optimus site is under construction. The Optimus passages in that update give no unit count, selling price or cost per unit. Tesla also said it plans to reach 400 MW of training compute by year end to support vehicle and humanoid autonomy software.

The targets are better documented than the results. On the first-quarter 2026 call, as reported by Electrek, Musk called the production rate this year literally impossible to predict, cited more than 10,000 unique parts, and described initial robot skills as simple factory tasks. The same report says Musk’s January 2025 statement that Tesla would build roughly 10,000 Optimus robots that year was missed, and that in January 2026 he acknowledged no Optimus robots were doing useful work in its factories. A second Optimus factory at Giga Texas is expected around summer 2027. These are company statements about intentions, relayed by a news outlet, and we treat them as such.

Chinese makers report actual units. Unitree’s prospectus-linked coverage shows 2025 revenue of 1.70 billion yuan against 392.77 million yuan in 2024, with more than 5,500 humanoid units shipped in 2025 and 5,632 humanoids sold cumulatively from 2023 to 2025, alongside 33,294 quadrupeds. Secondary sources disagree on its net profit figure, so we do not cite one.

UBTech, listed in Hong Kong, reports a business that is further along in revenue and still loss-making. Per KrASIA’s account of its first-half 2026 results, group revenue was RMB 1.27 billion, up 104.2%, full-size humanoid revenue was RMB 590.3 million (46.5% of the total) on 921 units sold, gross margin was 44.7%, and the net loss was RMB 338.8 million. The same article gives a 2026 production target of 10,000 full-size humanoids, which is a target and not a result. Other coverage we saw gave a different target, and we could not reach the company’s exchange filing, so the UBTech figures carry secondary-source risk.

The pattern across these companies is consistent. Thousands of units have shipped, mostly to research, education, exhibition and pilot customers. No humanoid maker has disclosed fleet-level productivity, uptime or a recurring revenue stream per robot, which are the numbers that would show a humanoid earning its place on a floor.

China: The Supply Side

China is the largest buyer of industrial robots and increasingly the largest seller to itself. The IFR reports that Chinese suppliers held 57% of their domestic market in 2024, up from about 28% a decade earlier, and that Chinese manufacturers sold more than foreign suppliers at home for the first time. Its outlook says the domestic share will continue to increase.

That shift has two effects on the argument. First, it pressures price. Chinese demand has rarely declined in the IFR’s history, and domestic suppliers can win by undercutting. Second, the same industrial base that makes arms makes the actuators, sensors and batteries that humanoids need, which is a plausible reason the lowest-priced humanoids we found come from there. The list price of Unitree’s base G1 is US $13.5K, for a robot of about 35 kg with roughly two hours of battery life.

We have not found reliable, primary-source data on component makers such as reducer, motor and sensor suppliers, and we leave them out instead of estimating. The IFR’s domestic-share figure, Unitree’s prospectus-linked numbers and UBTech’s results are the China evidence we can stand behind, and the last two come through secondary reporting.

Unit Cost Versus Labor Cost

The cost comparison that decides adoption is simple to state and hard to complete. A robot buyer compares the all-in cost of a machine doing a task against the all-in cost of a person doing it, over the machine’s life.

Take the cheapest humanoid we can price. At $13.5K, a G1 costs about 35% of one year of median U.S. material mover pay of $38,220. That is a hardware price only. With about two hours of battery life, a single unit does not cover a shift, and the price excludes charging, maintenance, software, supervision and the integration work the IFR identifies as the bottleneck. The ratio looks favorable only before those costs are added, and the sticker price says nothing about productivity, which is the missing term.

Surgery works through a different calculation. A hospital pays $0.6 million to $3.1 million for a system and $900 to $3,700 of instruments per case, and accepts that because the robot changes outcomes, stay length or surgeon capacity, not because it replaces a wage. Warehouse systems sit between: Symbotic customers buy throughput and storage density. In neither case does robot price versus a single wage settle the question.

The honest reading is that unit cost versus labor cost works as a ceiling and not a driver. It tells us a humanoid priced in the tens of thousands of dollars can compete with a low wage on paper. Whether it does depends on uptime and task breadth, and on who bears the cost of failure. That points back to the owner of the fleet, who either absorbs that risk through usage-based pricing, as Intuitive does, or pushes it to the customer.

What Would Change Our View

We would revise the thesis on these signposts.

• IFR installations for 2026, published in September 2027, fall below 2025’s 600,000. That would show factory demand stalling and weaken the premise that the installed base keeps compounding.
• Intuitive’s da Vinci installed base growth falls below 8% year over year, or procedure growth falls below 10%, for two consecutive quarters in 2027. That would mean the reference case is saturating.
• Symbotic’s software and operation services lines stay below 10% of revenue through fiscal 2027 while Walmart remains above 70% of revenue. That would say the recurring layer is not forming.
• Tesla discloses external Optimus customer revenue or unit shipments in any 2027 quarterly update, with a stated cost per unit. Absence through 2027 would keep humanoids in the target category.
• A humanoid maker discloses fleet uptime and recurring revenue per robot at a deployed customer. That would be the first evidence for the installed-base model in humanoids.

The Companies

Tesla (TSLA)

Tesla is the largest public bet on humanoids, though the robot is a small part of what it sells. In the third quarter of 2026 it reported revenue of $28,046 million and operating income of $718 million, with research and development of $2,761 million and capital expenditure of $4,732 million. Free cash flow was $1,531 million. Spending is rising faster than operating income, and the Optimus line sits inside that spending. Our full analysis is in the Tesla memo.

For this thesis the facts are that Optimus production has begun in Fremont and a Texas site is under construction. The targets, including units, skills and customers, are statements of intent. The test for Tesla is whether it discloses a cost per robot and an external customer before the end of 2027. Until then, Tesla fits our framework as a fleet-data and manufacturing story on the vehicle side, and a pre-revenue option on the robot side.

Intuitive Surgical (ISRG)

Intuitive is the reference case for installed-base economics in an open, high-stakes space. It had 12,094 da Vinci systems in service at September 30, 2026, about 906,000 procedures in the quarter, and revenue of $3.05 billion, up 22%. Instruments and services made up 75.8% of nine-month revenue by our calculation, and usage-based leasing is growing. The full analysis is in the Intuitive memo.

Two risks remain, one legal and one about price. The $450 million antitrust settlement is subject to court approval, and the 10-Q flags planned instrument price cuts in China. Procedure growth guidance of 14% to 15% for 2026 is the metric that connects both to the model.

Symbotic (SYM)

Symbotic is the largest public test of whether warehouse automation can become a recurring-revenue business. It reported fiscal third-quarter revenue of $721 million, net income of $55 million and a backlog of about $22.5 billion, mostly Walmart and Exol. Recurring-type lines were 6.5% of nine-month revenue by our calculation. The full analysis is in the Symbotic memo.

Two numbers define the debate: Walmart at approximately 85% of fiscal 2025 revenue, and cash falling from $2.0 billion to $1.7 billion in one quarter. Fiscal fourth-quarter guidance is for $760 million to $780 million of revenue.

Teradyne (TER)

Teradyne is a semiconductor test company that also owns Universal Robots and MiR. Robotics was $96 million of $1,306 million third-quarter revenue, and the segment turned profitable. Semiconductor Test, at $1,072 million, drives the company’s results and its exposure to AI data center demand. The full analysis is in the Teradyne memo.

For this report Teradyne is the cobot read-through. The segment is growing at roughly 28% to 31%, per the 10-Q, but it is about 7% of revenue and its filings show no recurring revenue per installed arm. It tells us cobots are growing, and tells us little about who captures value on the shared floor.