The embodied-AI revolution is announced in press releases and measured in units shipped. The units say the future is further away than the videos. The world installed 542,076 industrial robots in all of 2024, according to the International Federation of Robotics. The same world built 96.4 million motor vehicles in 2025, according to OICA's production tally. One number is 178 times the other. This article does the arithmetic the announcements skip. Every figure carries its denominator, its date, and its evidence label.
The method is subtraction, not enthusiasm. Count delivered units, not demonstrated tasks. Separate the production rate from the installed fleet, the fleet from paid work, and paid work from renewed contracts. A video is evidence about one camera session. An installation statistic is evidence about an industry. The two currently stand several orders of magnitude apart, and closing that gap is a manufacturing problem before it is an intelligence problem.
a staged demo becomes a press releasethe press release becomes a valuationthe valuation becomes a shipment forecastthe forecast meets a factory
The kill condition comes first, because the thesis is empirical. This article claims embodied AI will scale on factory clocks, not model-release clocks. If the dated thresholds in the final section arrive early, installation rates, intervention rates, real prices, renewed contracts, the thesis is wrong on schedule. The article itself names the observations that would kill it.
The census the headlines skip
Verdict: the robot industry is large, successful, and nowhere near the scale its press coverage implies. Both halves are measured, and dropping either half is the error this section corrects.
ESTABLISHED: the IFR's World Robotics 2025 report, published 25 September 2025, counted 542,076 industrial robots installed worldwide in 2024. That was the fourth consecutive year above half a million, two percent below the 2022 peak. The operational stock reached 4,664,000 units, up nine percent year on year. Preliminary IFR data, published June 2026, put 2025 installations near 621,000, up fifteen percent. This is a real industry growing at a real rate. The arithmetic below is not an argument that it fails. It is an argument about how long scale takes.
ESTABLISHED: OICA counted approximately 96.4 million motor vehicles produced in 2025. The ratio is the argument. Cars outnumber newly installed industrial robots 178 to one. Cars are also the comparison the humanoid industry itself invites. Tesla, XPeng, Hyundai, and BMW all frame humanoids as the next automotive-scale product. log₂(178) ≈ 7.5. The robot industry must double its annual installation rate seven and a half times to install in one year what the car industry builds in one year.
Now run the doublings at honest growth rates. At the IFR's own forecast pace of roughly seven percent a year, installations double every 10.2 years. Matching one year of car output takes about 77 years, landing near 2101. At fifteen percent a year, faster than industrial robotics has ever sustained, doubling takes 5.0 years and car-scale arrives near 2061. At twenty-five percent a year, held for two decades with no recession, the crossing lands near 2047. Only at fifty percent annual growth, a rate no heavy-hardware industry has held for ten consecutive years, does the crossing arrive near 2037. The growth-rate assumption is not commentary on the forecast. The growth-rate assumption is the forecast.
Humanoids sit far below even this base. Omdia estimates roughly 13,000 humanoid robots shipped worldwide in 2025. IDC's alternate count runs near 18,000. The definitional gap between the two is itself a finding about measurement quality in the category. Against 542,076 industrial installations, humanoids were about 2.4 percent of units on Omdia's count. Most of those units went to research laboratories, data-collection fleets, and demonstrations rather than paid production work. The category being headlined is a rounding error inside the category being ignored.
Falsifier for the section: the IFR's September 2026 report shows 2025 installations at or above 700,000. That would be its own 2028 forecast arriving three years early. The second trigger: humanoid shipments independently audited above 100,000 for 2026. Either observation bends every curve in this section upward.
What a billion robots would take
Verdict: the billion-robot world is a manufacturing program measured in generations, not a product cycle measured in quarters. The arithmetic is short, and it is merciless to the keynote version of the future.
Production first. At today's total industrial-robot output of 542,076 units a year, one billion robots take 1,845 years to build. Scale the entire industry twentyfold, to ten million units a year. No robotics sector has ever approached that output. A billion robots still take a century at that rate. At one hundred million units a year, roughly the whole global car industry converted to robot production, the job takes ten years. The claim that robots will be everywhere soon is secretly a claim about building a car-industry-sized manufacturing sector that currently does not exist, staffed, supplied, and financed.
Replacement second, because a fleet is not a stockpile. It is a flow with a service life. At a ten-year lifespan, a one-billion-robot fleet consumes one hundred million replacements a year merely to hold its size. That maintenance flow alone is 184 times the world's entire 2024 robot output. Depreciation, not production, is the steady-state constraint, and depreciation never appears in the keynote.
Actuators third, and this is where the arithmetic bites hardest. An advanced humanoid carries roughly thirty major actuated joints, each a motor, gearbox, encoder, bearings, and control electronics, thirty small precision assemblies per body. Producing one hundred million robots a year means producing three billion precision actuators a year. For calibration, take the automotive industry. It is the largest precision-electromechanical supply chain on earth. It turns out on the order of one to two billion electric motors annually across 96.4 million vehicles (INFERRED from vehicle counts and typical motor content per car). A billion-robot economy needs actuators at several times the scale of the car industry's entire motor supply. That industry does not presently exist, and no funded, dated plan to build it exists in the public record as of August 2026.
The surprising exception is batteries. One hundred million robots carrying the Figure 03's 2.3 kWh pack each would need roughly 230 GWh of cells a year. That is large, but well inside the EV industry's trajectory. Battery manufacturing already operates at terawatt-hour scale annually (ESTABLISHED as an industry; the comparison is INFERRED). Precision mechanics, wiring that survives millions of bends, thermal management, field repair, and deployment engineering bind harder than energy storage. The bottleneck map matters because capital flows to the wrong layer when the map is wrong.
Falsifier for the section: a disclosed actuator program from any government or company. It must name capacity above three hundred million precision actuators a year, with a completion date before 2035. One such program, verifiably funded and under construction, breaks the constraint this section prices.
The Nokia N95 problem
Verdict: the cost curve that made computers cheap does not run through precision hardware, and nineteen years of smartphone history prove it. The high-end consumer device did not get cheaper in real terms. It got better at the same price.
The dated anchor. In 2007 the Nokia N95, the ultimate pre-iPhone flagship, carried GPS, Wi-Fi, a five-megapixel camera, and arguably more individual features than the first iPhone. Wired's contemporary review priced it at $749 unlocked in the United States. Apply roughly 61 percent cumulative US consumer inflation and that price becomes about $1,206 in 2026 dollars. Now look at the 2025-cycle flagship: Apple's iPhone 17 Pro Max starts at $1,199. The two numbers are, to within half a percent, the same number. Nineteen years of the most aggressive supply-chain optimization in industrial history left the flagship price exactly where it started in real terms. What fell was the price of the mediocre phone. What held was the price of precision.
The mechanism generalizes, and it is the load-bearing claim of this section. Compute obeys Moore's law because transistors shrink. Precision mechatronics obeys machining tolerances, materials, assembly labor, and production yield, and none of those shrinks on a schedule.
compute gets transistors shrink → → gears do not shrink cheaper precision motion the flagship price does not get holds cheaper
Name what must get cheaper for the humanoid price to move: actuators, dexterous hands, tactile skins, safety certification, and field repair. The current evidence runs in both directions and must be read in both. On one side, industry cost analyses put today's deployed humanoid pricing near $90,000 to $100,000 per unit. Bank of America projects unit costs below $17,000 by 2030. These are REPORTED projections, not reproduced prices. On the other side, Unitree already lists the G1 from around $13,500. That proves aggressive cost compression is possible. It also proves what cheapness omits: durability ratings, tactile coverage, safety certification, and a repair network. The N95 lesson is not that prices never fall. It is that the price of the complete, integrated, reliable product held flat for two decades while components inside it got cheap. The robot must fall in price as a complete product, and the complete product is the part that has never followed the chip curve.
One more inheritance from the phone wars. Nokia did not lose on hardware; it lost on integration, body, operating system, interface, developer platform, distribution, payments, service. Today's humanoid field reproduces that structure exactly. Every maker ships a different body. Skills do not transfer cleanly between bodies. There is no robot-app ecosystem, no financing layer, and almost no repair layer. The decisive contest is integration economics, and it will not be settled by a better backflip.
Falsifier for the section: a humanoid with certified safety, published durability, and a working repair network, sold below $30,000 at positive margin by end-2028. The bill-of-materials economics must be verified by a party that does not sell the machine. That price at that completeness would break the N95 pattern and force the cost sections of this article to be rewritten.
The scissors test
Verdict: the honest measure of embodied intelligence is not a backflip and not a folded shirt. It is a twelve-step encounter with an unfamiliar pair of scissors, and no general robot passes it. The gap is not intelligence in the abstract. It is contact-rich manipulation: the physics of touching a world that pushes back.
The test, stated fully, because its length is the point. Enter an unseen room. Find several unfamiliar scissors. Determine whether each is right-handed, left-handed, or ambidextrous. Choose one suited to the material. Pick it up from an arbitrary orientation. Insert the correct fingers without visual markers. Hold loose paper with the other hand. Cut an unseen curved line accurately. Detect slipping, folding, tearing, and jamming as they happen. Switch hands on command. Avoid a moving human hand nearby. Put the scissors away safely. A six-year-old passes this test. No general robot does, ESTABLISHED as of August 2026, by the absence of any published, independently witnessed complete pass.
The standing reproach has a precise name. ScissorBot already cut curves through paper, cardboard, fabric, plastic, and foil at roughly one-handed-human performance. But the scissors were attached to the arm, the paper was suspended, and the action space was specially structured. That is a trained skill in a shaped world, and it is real engineering. The full test measures something else. Right-handed scissors mechanically fail in the left hand, because blade pressure reverses. A passing machine notices the incompatibility, then reorients or selects accordingly. Detecting incompatibility is intelligence. Forcing the wrong scissors through the paper is not.
The measured frontier confirms the shape of the gap. Google reported Gemini Robotics 2 task success ranging from 32 percent to 92 percent across five multi-finger hand tasks. The simpler gripper categories ran 74.2 to 89.6 percent. Reliability is jagged even inside one model release. NVIDIA's GR00T N1.5 reached 83 percent on a real-robot fruit-selection task and 84.2 percent after training on novel-object placement. These are genuine advances, and read carefully they are also single-task rates in prepared settings. An 83 percent success rate is a triumph on a benchmark and a liability in a kitchen, because of what rates do when tasks chain.
What the scissors test does not measure must be named too. It does not measure language, planning, or visual reasoning, where machines now exceed most humans. It does not prove robots will never manipulate well. Fixed industrial cells already outperform people at shaped contact tasks. A commercial UR5e arm repeats positions to ±0.03 mm. It measures the unsolved layer, integrating sensation into immediate corrective action, in matter, on first encounter, without a prepared setup. Sensor sensitivity is sometimes superhuman; research tactile sensors resolve micron-scale surface features, and a DIGIT fingertip sensor sells for about $355. Whole-hand tactile intelligence remains substantially subhuman. A robot can carry a better fingertip than yours and still crush the cup, because the loop from feeling to correction is the missing organ.
Falsifier for the section: a general robot, independently witnessed, passing the full twelve-step test at 95 percent success across 1,000 trials in rooms it has never seen. One verified pass at that reliability rewrites this section overnight.
Reliability multiplies against matter
Verdict: embodiment converts small per-step failure rates into large system failure rates. The software industry's favorite repair, just retry, transfers badly into a world made of atoms. This is the deepest quantitative reason the videos mislead.
The multiplication is elementary and brutal. A humanoid carries roughly thirty powered joints. If each joint is 99 percent reliable over a working period, whole-body reliability is ≈ ≈ 0.99^30 74 percent. At 99.9 percent per joint, the body still reaches only 0.999^30 97 percent. For the body to hold 99 percent, every joint needs about 99.97 percent reliability. That is before counting cameras, hands, wiring, batteries, and software. Each carries its own failure rate. Reliability in embodiment is not a property you add to a machine. It is a property that multiplies across a machine, and multiplication is unforgiving.
Task chains multiply the same way. A thirty-step manipulation sequence at 99 percent per step completes end to end 74 percent of the time. At 95 percent per step, already excellent against the measured rates quoted above, it completes 21 percent of the time. A five-minute demo video shows one draw from that distribution. The metric that prices a deployment is productive hours between human interventions, and that is precisely the metric on which the public record goes silent.
each step succeeds almost alwaysthirty steps multiply the almostthe chain fails one time in foura human walks overthe margin walks over with them
Now the retry asymmetry, because software agents have already shown how much recovery is worth, in software. The arithmetic: fifty dependent steps at 98 percent per step give 0.98^50 ≈ 36 percent end-to-end success. Add a retry that recovers 60 percent of ≈ failed steps and per-step success rises to 0.98 + 0.02 × 0.6 = 0.992, and 0.992^50 67 percent. Recovery halves the failure rate and nearly doubles the yield, which is why software agents look increasingly dependable. But the trick carries a hidden premise: the world must be unchanged after a failed attempt. A failed API call leaves no debris. A failed grasp scatters parts, creases the sheet metal, spills the liquid, or jams the tool. In matter, the retry begins from a worse state than the first attempt. The reset costs minutes and sometimes a human's hands. The failure itself can cost more than the task was worth. A retry is free in RAM and expensive in atoms.
The intervention arithmetic makes the economics concrete. Suppose a fleet of one hundred robots each needs one human intervention per operating hour, an optimistic figure against today's measured systems. Suppose each intervention takes five minutes of a technician's time. That is five hundred technician-minutes per operating hour: 8.3 full-time humans continuously supervising one hundred robots. The fleet has not removed labor. It has converted line labor into exception-handling labor and added a payroll on top of a capital budget. Only when interventions fall below roughly one per shift does the arithmetic flip. That is why Figure's reported ambition at BMW was zero interventions per shift. The word "ambition" carries the weight of the entire industry.
Falsifier for the section: a named operator publishes audited intervention rates below one per eight-hour shift. The fleet must be at least five hundred humanoids performing the same paid task, sustained for six months. That single statistic would break the multiplication argument at fleet scale, and it is publicly observable if it ever becomes true.
The deployment funnel, dated
Verdict: the humanoid industry in 2025-2026 sits between pilot and paid pilot. Exactly one adjacent category, warehouse logistics, touches scaled paid work. There is also one instructive withdrawal. The funnel runs demonstration, pilot, paid pilot, scaled deployment, renewal. Renewal is the only stage that proves value, and almost nobody has reached it.
demopilotpaid pilotscaled deploymentrenewal after the invoice
The strongest record belongs to Figure at BMW, and it deserves a precise reading rather than a headline. Over an eleven-month deployment at Plant Spartanburg, Figure 02 robots logged more than 1,250 runtime hours. They loaded more than 90,000 sheet-metal parts and contributed to the production of more than 30,000 BMW X3s. These figures are REPORTED by Figure and BMW, not independently audited. Run the denominators. 1,250 hours over eleven months averages about 26 hours a week: roughly two-thirds of one full-time worker on one shift. 90,000 parts over 1,250 hours is 72 parts an hour, one every fifty seconds. This is real production-adjacent work. It is the most credible humanoid
deployment on record. It is also one repeated task, in one engineered cell, at one factory, with the robot count unpublished. Both sentences are true. Calibration is holding them at the same time.
The rest of the funnel, by stage. Agility's Digit has moved more than 100,000 totes at GXO's Flowery Branch facility. A cumulative physical-work count is stronger than any demonstration. Robot count, hours, and intervention rate remain undisclosed. Boston Dynamics reports more than 2,000 Spot and Stretch units deployed across customer operations over five years. That is a real commercial base, but for non-humanoid machines doing inspection and palletizing, not general work. AgiBot announced its 10,000th robot in late March 2026 and its 15,000th by late June 2026. That is an acceleration from eleven months per five thousand units to one quarter, the fastest manufacturing ramp in the category. The split between deployed units and warehoused units remains undisclosed. Tesla began Optimus Gen 3 mass production at Fremont on 21 January 2026. It has disclosed only several hundred units, all in learning and data-collection mode, after missing its 2025 target of one thousand units doing useful work. The one-million-per-year Fremont figure is a capacity target. The stated 2026 plan of 50,000 to 100,000 units runs against an analyst consensus near 10,000 to 30,000. And 1X opened preorders for its NEO home robot at $20,000 or $499 a month. It claimed more than 10,000 preorders in five days. It had zero independently verified customer deliveries as of mid-July 2026. And it discloses that difficult chores run through scheduled human teleoperation. A robot with a remote human inside it is partly a labor service wearing a robot body. Its autonomy claims must be discounted by exactly the size of that human.
The funnel also runs backwards. Amazon announced its Blue Jay warehouse robot with fanfare and updated the announcement on 25 February 2026 to say the system was no longer used in operations. A withdrawn system is stronger evidence about the demo-to-scale pipeline than any keynote: the middle stages of the funnel kill products quietly, after the coverage ends.
The calibration anchor for the whole funnel is Waymo, the one embodied-AI category that has verifiably reached scaled paid autonomy. It reported more than half a million paid trips a week across ten US cities as of March 2026. Its June 2026 analysis covered more than 220 million fully autonomous miles, with 94 percent fewer serious-or-fatal-injury crashes than matched human drivers. Waymo proves the funnel can be completed. It also proves the timeline: the program began in 2009. Seventeen years. One narrow operational domain. Geofenced cities and standing remote-assistance infrastructure. And this for the category where the physics is easiest, the task most repetitive, the funding deepest. Humanoids face harder physics, broader task claims, and thinner margins. Expect longer, not shorter.
Falsifier for the section: three or more named customers independently confirm scaled humanoid deployments by end-2027. Scaled means over one hundred units each, doing paid work, renewed after the initial contract. As of August 2026, the count of such customers is zero.
The strongest case for speed
Verdict against myself, stated first: the acceleration case is the strongest it has ever been, and it deserves its best form rather than a caricature. Then the arithmetic answers it.
The case has three legs, and each is real. First, falling hardware cost. Unitree lists humanoids in the low five figures. Bank of America projects sub-$17,000 unit costs by 2030. Chinese supply chains compress hardware prices faster than Western ones. The drone and electric-vehicle industries are the proof of pattern. Second, foundation-model transfer. Gemini Robotics, Physical Intelligence's models, and NVIDIA's GR00T line show single models controlling multiple bodies. They learn new tasks from fewer demonstrations. This attacks the integration cost that fragmented the pre-iPhone market. Third, manufacturing depth: China installed 295,000 industrial robots in 2024, fifty-four percent of the world total, and runs an operational stock above two million. Domestic makers took 57 percent of their home market. AgiBot's ramp from five thousand to ten thousand units in one quarter shows acceleration already present in the record. Assembled: cheap bodies, transferable brains, and the deepest factory base on earth, compounding.
cheaper bodiesmore units soldmore real-world databetter modelsmore useful robotsmore units sold
The flywheel is real. Now price it. Grant the most aggressive leg in full: let humanoid production double every single year from Omdia's 13,000 units in 2025. That is a one-hundred-percent annual growth rate sustained for a decade. It is faster than the car industry ever grew, and faster than smartphones grew in unit terms. The sequence runs 26,000 in 2026, 416,000 in 2030, 6.7 million in 2034, 13.3 million a year in 2035. Even that fantasy curve reaches one-seventh of annual car output by 2035. It does not build a one-billion fleet until the 2040s. And it required every doubling to arrive on time. No hardware curve has ever done that for ten consecutive years. Against it stands the only long series the industry actually possesses: the IFR's record shows global industrial-robot installations growing 2.4-fold across the entire past decade. That is a nine percent compound rate. The steelman needs a growth rate eleven times the reproduced rate, sustained four times as long as any hardware industry has sustained anything like it.
The other two legs meet the same wall at different heights. Transferable brains raise per-step success. The reliability section showed what chains do with that. Lifting per-step reliability from 95 to 99 percent moves a thirty-step task from 21 to 74 percent. That is transformative, and still short of the sub-one-intervention-per-shift regime that fleet economics require. Better models are necessary. They are not sufficient, because the residual failures are physical: dropped loads, worn gearboxes, broken fingers, and the long tail of rooms nobody trained on. Manufacturing depth, meanwhile, builds the bodies without building the demand. Ten thousand refundable $200 deposits are options on a product, not proof of a market. And China's industrial-robot dominance was built over twenty years, inside factories that engineered the environment around the arm. General robots are sold precisely to environments that will not be engineered.
CUT: the claim that humanoids are one product cycle from general household work. The funnel record contradicts it at every stage above pilot, and no evidence in the corpus or the public record supports it. What survives the cut is narrower and still large: a real industry, compounding at factory speed, arriving in waves of structured work over decades.
Falsifier for this rebuttal: humanoid shipments independently audited above 400,000 units for 2027. That is the fantasy curve's own milestone, arrived on schedule. At least half must be paid external deployments, not internal data fleets. If that number prints, the arithmetic in this section is dead and the doubling clock governs everything above it.
What would change the verdict
The verdict, restated once: embodied AI is real, improving, and priced for a future the unit counts do not yet contain. The binding constraints are factories, actuators, reliability compounding, and renewal economics. They run on decade clocks. Every claim in this article is dated and falsifiable, and the kills are specific. Any two of the following, independently verified, overturn the thesis.
1. Installation. Global industrial-robot installations above 700,000 in the IFR's final 2025 figures, three years ahead of its own forecast, and audited humanoid shipments above 100,000 for 2026. This bends the doubling clock hard enough to matter. 2. Reliability. Audited intervention rates below one per eight-hour shift across a named fleet of five hundred or more humanoids on one paid task, sustained for six months. This kills the compounding argument where it lives. 3. Cost. A certified, repairable humanoid with published durability sold below $30,000 at positive margin by end-2028, with bill-of-materials economics verified externally. This breaks the N95 pattern at the flagship tier. 4. Capability. An independently witnessed pass of the full twelve-step scissors test at 95 percent over 1,000 trials in unseen rooms. The passer must be a robot that also does other jobs. This dissolves the manipulation gap the entire cost case waits on. 5. Demand. Three named customers, each running over one hundred humanoids on paid work, renewed after the initial contract, by end-2027. Renewal is the only stage of the funnel that cannot be staged.
None of these is reproduced as of August 2026. What is reproduced is a real industry growing at nine to fifteen percent a year. Its celebrated category ships one unit for every forty industrial robots, and one for every seven thousand cars. The videos are genuine. The progress is genuine. The distance is also genuine, and it is measured in units, doublings, and interventions per shift, quantities that do not respond to keynotes. Intelligence may improve exponentially. Atoms do not, and the future of AI is made of atoms.
