The Anthropomorphic Premium
The humanoid robot is a capital-allocation error dressed as a moonshot.
By industry counts, more than seven billion dollars moved into humanoid robots last year, and the round sizes have only grown since. Figure is reportedly valued at around thirty-nine billion dollars, Tesla has floated a target price for its coming Optimus near thirty thousand dollars, and every few weeks another video shows a metal figure folding laundry or sorting parts with unsettling grace. The press files all of it under one headline, the arrival of physical artificial intelligence, the machine that finally works the way we do. Read the unit economics instead of the headline, and a different story appears. The market is not paying for capability. It is paying a premium for a resemblance, and that resemblance is the most expensive design decision in the industry.
What the money is actually buying
Strip away the wonder and the humanoid thesis makes a narrow claim, that a machine shaped like a person, able in principle to do many things, is worth more than a machine built to do one thing perfectly. Versatility is the pitch, and the human silhouette is offered as its proof. This is where the reasoning quietly breaks. A general-purpose humanoid does many tasks poorly at a price premium, while a purpose-built system does one task with a clear return, and in an industrial setting the second machine wins on every line a chief financial officer actually reads. Peter Lasinger, a European venture investor, calls the fixation the anthropomorphic fallacy, the urge to build machines in our own image even when the image adds cost and subtracts nothing. The fallacy is real. The more useful word for a board is premium, because a fallacy is a thinking error, and a premium is a number you overpay every month the machine stays upright.
The equation one layer down is measured in watts
The software side of this market already learned the lesson the hard way, and it is the same lesson. Model capability became cheap and abundant faster than almost anyone expected, and the cost that ended up deciding whether a digital colleague earned its place was never the intelligence, it was the running cost, the price of every inference at scale. In hardware the same equation returns wearing different units. What decides a physical machine is not what it can do in a demo, it is how much energy it burns per task, the ratio of power to payload. A human being walks, lifts, reasons and repairs itself on roughly the power of a light bulb. A humanoid draws many times that, and much of the budget is spent not on the work but on the mere act of staying upright, on solving a balance problem every millisecond that nature only ever solved because it could not grow wheels. Lasinger reads the whole field through intelligence per watt, and the axis he points to is the right one. An anthropomorphic machine spends most of its energy budget on looking like us.
The teleoperation tell
There is a quieter fact under the demos that explains more than any valuation. A large share of the viral humanoid footage is teleoperated, driven in real time by a human in a motion-capture rig or a haptic suit. Tesla’s robots at the October 2024 event were remotely assisted for their crowd interactions, the polished hand demo a month later was teleoperated, and a fall in a late 2025 showcase reopened the same debate in public. This is not a scandal, it is a signal, and it is the physical twin of a pattern anyone who has shipped enterprise artificial intelligence knows by heart. The demo dazzles, the production case stays unpriced. A machine that needs a dedicated operator behind the curtain to cross a dynamic factory floor is not an automation solution, it is an expensive avatar. Real industrial scale runs the ratio the other way, one operator overseeing a fleet of twenty autonomous purpose-built machines, rather than one operator married to a single humanoid for the length of a shift.
If 2026 really is the ChatGPT moment, the thesis holds
The strongest objection deserves the floor. At CES in January, Jensen Huang declared that the ChatGPT moment for physical AI has arrived, and unlike previous years the announcements around him carried shipping numbers rather than concept videos. Boston Dynamics is wiring Google DeepMind’s Gemini models into Atlas with a stated target of thirty thousand units a year by 2028, and the first deployments are committed for 2026. Around the same time, an angel investor who had been shown the next Optimus in Tesla’s lab told his audience that nobody will remember Tesla ever made a car. Musk’s public reply was two words, probably true.
Take the analogy seriously, because it cuts the other way. The ChatGPT moment of software AI did not belong to a machine that resembled a person. It belonged to the least human interface imaginable, a text box, because the revolution was the model and never the body. If physical AI now has its equivalent moment, the same logic applies one level down, the intelligence becomes abundant and portable, and it will flow into whatever body delivers the most work per watt and per dollar in each environment. Nothing about that favors legs. Huang, it is worth remembering, wins either way, since the chips are the same whether they sit in a humanoid or in a wheeled arm. And the Tesla line is not evidence about robots at all, it is evidence about narrative, a company valued in the trillions needs a story larger than cars, and the confirmation from the top was a confirmation of the story’s necessity, not of the machine’s economics.
Where the capital is mispriced
This is the part a board or a fund should sit with, because the error is not only in engineering, it is in how the asset is valued. The market is pricing the humanoid install base as optionality, a versatile platform that will pay off across many future uses, when this cycle the resemblance is closer to a liability than an asset. Every degree of human likeness carries a cost that never appears in the launch video, in maintenance, in the safety margin around real workers, in downtime, in the sheer mechanical fragility of a tall two-legged frame. The wheeled, purpose-built alternative gives up the magic and keeps the margin. The judgment that separates a good underwriter from a late one is exactly this, that versatility is being counted as value when in an industrial setting it is mostly cost, and that the environment can almost always be adapted to fit a simpler machine more cheaply than the machine can be made human enough to fit the environment. Whoever keeps paying for the silhouette is buying a story. Whoever reshapes the shelf, the floor and the bin to suit a wheeled arm is buying a return.
The size of the premium can be read straight off the public numbers. Goldman Sachs projects the entire humanoid robot market at around thirty-eight billion dollars in 2035, which is less than the reported valuation of Figure alone today; other houses reach trillions on a 2050 horizon, but the nearer the date, the smaller the market and the wider the gap to the prices being paid. And the premium has now arrived in Europe. Neura Robotics of Metzingen closed a Series C of up to 1.4 billion dollars in June, led by Tether with Amazon, Nvidia and Qualcomm alongside, the largest robotics financing Europe has ever seen, framed as Physical AI made in Europe. Neura is the instructive case, because the company earns its money today with purpose-built cognitive machines for industry while the humanoid flagship carries the story that raised the round. Europe’s biggest robotics bet confirms both halves of the argument inside a single company, the capital follows the silhouette, the revenue follows the architecture.
And here the debate about steel rejoins the one that runs through all of this work. The decision was never the machine, in software or in hardware. It is the capability of the organization to match the right architecture to the right task, one domain at a time, and to tell the difference between a tool that earns its keep and a tool that merely looks the part. The humanoid is only the most visible case of a mistake that is everywhere in this transition, the mistake of taking the resemblance of intelligence for the result of it.
The way through
None of this means the robots are not coming. They are, and the useful question is not whether but which. Buy the resemblance and you carry the premium for as long as the machine runs. Buy the result and you ask a colder question first, what is the output per watt, and how much human is still left in the loop once the cameras are off. The winning architecture, in the data center and on the factory floor alike, is the one with the best answer to those two questions, and it will almost never be the one that looks back at you.
Whoever keeps paying the anthropomorphic premium will find, a few years from now, that the returns went to the builders who never cared what the robot looked like, and that they never saw it coming.
Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.