Every industrial revolution eventually gets its absurd-seeming number, and the humanoid industry just received its annual physical in the form of one. On September 15, the IoT market research firm Berg Insight published findings from a new strategic report, The Future of Humanoid Robots, estimating that annual shipments of humanoid robots will grow from 16,000 units in 2026 to 26.0 million units by 2040, a compound annual growth rate of 63.7 percent. Over the same horizon, the firm sees global market value climbing from US$890 million to US$554 billion.
The instinctive reactions are both wrong: uncritical awe at the big number, and lazy dismissal of it. The interesting exercise is the third one, which is to take the forecast seriously enough to ask what the world would have to look like for it to be true, what it implies about supply chains and labor, and which observable checkpoints between now and 2030 would confirm or break it. That is what this piece does.
What Berg Insight Actually Published
Start with the record, because the summary numbers are only part of the release. Berg Insight frames the industry as having “reached an important commercial inflection point,” in the words of market analyst Julia Wahlen: “For the first time, we are seeing several companies simultaneously scaling production, establishing dedicated manufacturing facilities and deploying humanoid robots in real-world commercial environments. The industry is clearly moving beyond prototype development and into the early stages of commercialisation.”
The report counts more than 100 companies developing full-size humanoids, most still in research, prototype or pilot stages, and names Figure AI, Tesla, Agility Robotics, Boston Dynamics, Apptronik, 1X Technologies, NEURA Robotics, Unitree Robotics, UBTECH Robotics, AgiBot, Leju Robotics, EngineAI and RobotEra as the group currently shaping the competitive landscape. It highlights the 2025–2026 investment wave, with NVIDIA, Microsoft, Google, Amazon, Hyundai and OpenAI deepening strategic involvement, and it argues, through analyst Emilia Rickard, that competition is “increasingly shifting from hardware performance alone towards software, embodied AI and scalable manufacturing.”
One more number deserves attention because it reveals the analyst’s frame: Berg Insight forecasts that cellular subscriptions associated with humanoid robots will grow from 5,000 in 2025 to 50.6 million in 2040, an 84.9 percent CAGR. We will return to that, because it quietly says something important about how these machines are expected to operate.
The Arithmetic of Belief
A 63.7 percent CAGR sounds like an abstract finance-department phrase. Translate it into physical reality and it becomes starker: shipments must double roughly every 17 months, sustained for about fifteen consecutive years. Not one doubling, not three. Ten full doublings back to back, from 16,000 units this year to 26 million.
No physical product category has done that at scale. The closest analogs are instructive rather than reassuring. Global electric vehicle sales grew on the order of 28-fold between 2015 and 2023, one of the fastest manufacturing ramps in industrial history, and even that ran at roughly 45 to 50 percent annual growth before decelerating hard in 2024. Solar photovoltaics sustained 30 to 40 percent annual growth for two decades, from a much smaller industrial base, with heavy state coordination on the demand side. Berg Insight’s curve asks the humanoid industry to run hotter than both, for longer, while producing a machine that is categorically more complex than a panel or a drivetrain.
There is a second, less discussed implication. If annual shipments follow that trajectory to 26 million in 2040, the cumulative installed base does not stop at 26 million. Summing the geometric ramp, the world’s operational humanoid fleet by 2040 lands somewhere around 65 million machines, on my arithmetic from the published growth rate. For scale, that is roughly 2 percent of the global labor force, comparable to the entire population of France, and it sits meaningfully below Morgan Stanley’s widely cited 1 billion humanoids by 2050, which RBC Capital Markets has framed as a $9 trillion market. The 2040 number, in other words, is the conservative entry in a rapidly inflating genre.
Then there is the price tell. Divide Berg Insight’s $554 billion market value by 26 million units and you get an implied average selling price of about $21,300 per robot in 2040. Now do the same with Goldman Sachs’ revised numbers, and the convergence is striking.
Two Forecasts, Two Different Presents
Goldman Sachs’ 80-page Physical AI report raised its 2035 humanoid shipment forecast nearly fivefold, from about 1.4 million to approximately 6.5 million units, lifted its 2030 forecast from 256,000 to 890,000, and marked its 2026 estimate up from 51,000 to 75,000 units. The bank values the 2035 market at roughly $138 billion. Divide $138 billion by 6.5 million units and the implied ASP is about $21,200.
Two independent research houses, working different methods and different horizons, converge on almost exactly the same average robot price a decade or more out: around $21,000. That convergence is doing quiet analytical work. It says both forecasts embed a world where the $16,000 Unitree G1 and the $20,000-or-$499-per-month 1X Neo are not loss-leading curiosities but representative price points, where industrial units with heavier payloads and consumer units with shorter duty cycles average out to something close to a well-equipped car, and where the cost curve falls far enough to enable volume but not far enough to collapse revenue. If you believe one thing about 2040, believe that number: it is the load-bearing assumption.
But look at what the two houses say about the present, and the agreement evaporates. Berg Insight puts 2026 shipments at 16,000 units. Goldman’s updated 2026 estimate is 75,000. That is a 4.7-times disagreement about what is happening this year, not in some distant future. Carry the published growth rates forward and the gap persists: Berg’s 63.7 percent CAGR implies roughly 115,000 units by 2030 on my arithmetic, versus Goldman’s published 890,000, an eightfold spread at the first serious checkpoint.
A gap that large about the present is not sloppiness. It is definitional. One house is likely counting commercial, revenue-generating deployments; the other is likely counting everything that ships, including developer units, research units and pre-production hardware flowing to universities and corporate labs. The industry still has no shared, auditable definition of a “shipped humanoid,” and until it does, every forecast in this genre carries an asterisk. The first company to publish audited unit economics, real utilization hours per robot per day, will do more for forecast hygiene than any analyst model.
The Bill of Materials Test
Could the supply chain even physically support 26 million humanoids a year? Run the bill of materials.
Take actuation. A modern full-size humanoid carries somewhere between 28 and 50 actuators depending on design philosophy; Tesla’s Optimus Gen 3 is reported at around 50. Use 40 as a working number and 26 million robots implies more than one billion precision actuators manufactured annually, each with gearboxes, encoders, motor windings and force sensors. For comparison, the entire global automotive industry produces on the order of 90 million vehicles per year, and each of those is a solved mechatronic problem with a century of process maturity behind it. A billion precision actuators a year means the planetary roller screw, harmonic drive and rare-earth magnet supply chains each need to scale by orders of magnitude beyond today’s capacity, which is sized for tens of thousands of humanoids plus industrial robotics demand. This is not a reason the forecast fails. It is a map of where the margin will sit in five years: whoever industrializes cheap, reliable actuation captures the picks-and-shovels rents of the entire wave, which is precisely the logic Goldman applies to semiconductors.
On chips, Goldman estimates each humanoid carries $3,000 to more than $6,000 of semiconductor content, spanning compute, analog and mixed-signal, and edge storage. At the bank’s 6.5 million 2035 units, that is $19.5 billion to $39 billion of annual silicon demand, a new leg of AI infrastructure spending that exists even if the robot market consolidates to a handful of winners. At Berg’s 2040 volumes, that same content math implies a semiconductor market approaching $80 to $160 billion a year, comparable to today’s entire global automotive chip market. The forecast is not really a robotics claim. It is a claim that physical AI becomes a first-order consumer of the world’s fab capacity.
And the $21,000 ASP, held against that bill of materials, imposes its own discipline. A machine with 40 actuators, several kilograms of battery, a sensor suite and an onboard compute cluster, assembled and warranted, must fit a price that today barely covers the component list for well-specified industrial units. That means the 2040 robot is not today’s robot at scale. It is a redesign around cost the way the Model T was a redesign around cost: fewer parts, radical actuator integration, and software margins subsidizing hardware that sells near build cost.
The Connected-Robot Tell
Berg Insight’s cellular forecast, 50.6 million robot subscriptions by 2040, is the most revealing line in the release, because Berg is an IoT analyst and it instinctively modeled humanoids as fleet-managed connected devices rather than autonomous appliances. Each of those subscriptions is a machine phoning home: telemetry, diagnostics, over-the-air updates, fleet coordination, and in the near term, teleoperation.
That framing matches the deployment reality of 2026 better than any autonomy narrative. The 1X Neo shipping to US homes at $20,000, or $499 per month, is designed around remote operators who can take over, a design choice that has already generated a privacy debate precisely because a company employee can, under defined conditions, see inside a customer’s home. If 2040’s fleet of 65 million machines still requires any meaningful human-in-the-loop ratio, the labor economics of the entire forecast bend around teleoperation centers, and the robot becomes less a replacement for labor and more a redistribution of it, from on-site work to remote operation at some operators-per-hundred-robots ratio. The subscription forecast is Berg telling you, without quite saying it, that the robots will not be alone.
Demand: The Ledger Runs Ahead of the Hardware
The labor market, meanwhile, is not waiting for the volume curve. Challenger, Gray and Christmas counts AI as the leading stated reason for announced US job cuts in 2026, cited in 116,175 reductions through August, about 22 percent of all cuts this year, while global humanoid shipments are still counted in the tens of thousands. The anticipation effect, hiring freezes, automation-justified restructuring, insurance and retraining flows, is an order of magnitude ahead of the installed base.
On the demand side, Goldman expects the traction to come from logistics and warehousing first: structured environments with enough variation to defeat fixed automation, which is exactly the niche Agility Robotics and Figure have been selling into, with automotive manufacturing as the second wave. The bank’s own case study is Amazon, which already operates more than 1 million robots across 300-plus facilities and could see roughly $72 billion in cumulative service-cost savings from 2026 through 2030, worth up to 240 basis points of operating margin in an upside scenario. Set against that, Berg’s 2040 installed base of roughly 65 million machines starts to look less like science fiction and more like the arithmetic of what warehouse economics alone could absorb, before consumer deployment matters at all.
Checkpoints That Will Settle It
The forecast makes testable claims. Four checkpoints matter most.
First, 2026 actuals. Berg says 16,000 units, Goldman says 75,000. When year-end shipment figures land, the definitional gap becomes visible, and whichever way audited numbers break, they will reset every model in the genre.
Second, 2030 volume. The published spread runs from roughly 115,000 units implied by Berg’s curve to Goldman’s 890,000. Capacity announcements are the leading indicator: Tesla’s Fremont Optimus line, Figure’s BotQ, Agility’s RoboFab, and the production line XPeng activated in Guangzhou this month, with mass production targeted by end of 2026, are all stated targets. Announced capacity that converts to shipped, revenue-bearing units on schedule would mark the first time this industry has ever hit its own dates.
Third, the ASP print. If average selling prices in commercial contracts hold near $21,000 while utilization rates justify the spend, the Berg and Goldman convergence holds. If prices collapse toward $10,000 without utilization improving, the market value forecasts break even if unit volumes arrive.
Fourth, the actuator supply chain. Watch capital expenditure into precision gearbox and roller screw capacity, and rare-earth magnet supply agreements. A billion actuators a year does not happen silently; the suppliers will telegraph it years in advance, and today they are not yet telegraphing anything close.
The Verdict
Berg Insight’s 26 million is best read not as a prediction but as a specification: here is the world in which the current investment thesis, roughly $270 million into London-based Humanoid alone, 34,000 pre-orders and a $2.4 billion pipeline behind it, actually pays off. The forecast requires the fastest sustained manufacturing ramp in industrial history, an actuator supply chain that does not yet exist at that scale, a stable $21,000 price point that two rival research houses independently converged on, and a demand side whose appetite is currently running years ahead of its hardware.
The honest position is that the direction is no longer in serious dispute: warehouses first, factories second, homes last, with software and manufacturing, exactly as Berg’s analysts argue, deciding the winners. The magnitude is. Between 115,000 and 890,000 units by 2030 lies the entire difference between a promising niche industry and a second automotive-scale industrial complex. The industry’s job for the next four years is to make that spread embarrassing. The forecasts, at their best, are how we will know which way it broke.
Sources: Berg Insight via Robotics and Automation News · Berg Insight release via I-Connect007 · Goldman Sachs Physical AI report coverage via 24/7 Wall St. · Challenger job-cut data via Yahoo Finance · XPeng IRON production via Electrek · Humanoid unicorn profile via Forbes