There is a phrase spreading through boardrooms faster than any product can ship: "physical AI." Last week Hyundai's executive chair Euisun Chung stood at a San Francisco summit and declared that his company is no longer an automaker but a physical AI company, one that will start by making cars and robots more intelligent. Days later in Tokyo, Mitsubishi Motors announced a partnership with a University of Tokyo startup to mass produce humanoid robots, aiming for 1,000 units a month by the end of 2027. Two carmakers, on two continents, reaching for the same identity in the same week.
The enthusiasm is not confined to Detroit's rivals. In China, the pattern looks less like a pivot and more like a stampede. According to the German analyst Konrad Wolfenstein, more than 320 companies are now chasing embodied AI, with over 46 billion yuan flooding into the sector in the first months of 2026 alone, already exceeding the whole of last year. Individual rounds have grown enormous: one newcomer, TARS, raised 455 million dollars in a single funding round. Wolfenstein's warning is blunt. This looks like China's electric-vehicle boom replayed at higher speed, a wave of state-fuelled startups that will end, as the car industry did, with hundreds of failures and a handful of survivors.
What almost nobody in the funding announcements dwells on is why physical AI remains so stubbornly hard, even as the underlying models grow dazzling. That is the argument made this week by Andrew Glenn, writing in the Building Our Future newsletter, and it is the most useful frame in the pile. Intelligence, he notes, has raced ahead in coding and mathematics because the relevant knowledge already exists in digital form. The physical world does not cooperate. A model cannot infer that a particular machine vibrates differently just before it fails, or that a warehouse floor turns slippery in humid weather, unless something measured those facts first.
Glenn calls the missing ingredient the "reality loop": the cycle in which a robot observes, attempts a task, recognises failure, receives correction and tries again. Much of what makes a factory productive is tacit, living in thousands of small adjustments that operators never write down. A humanoid dropped onto that floor needs more than a clever brain; it needs years of encounters with awkward objects, worn equipment and human improvisation. The interesting bet, on this view, is not the flashiest demo but the infrastructure that captures experience: Agility Robotics building a 60,000-square-foot training facility, or NEURA Robotics assembling a "physical AI gym."
The gap between promise and floor shows up in the demos too. At AMD's developer conference, a startup called Generative Bionics unveiled a humanoid named Gene.01 wrapped in pressure-sensitive "smart skin" that lights up where it is touched. It is a genuine advance in sensing. But as the company itself concedes, gathering the data is the easy half; turning it into real-time decisions is the work still ahead. Even China's own investors admit the valuations have detached from capability. As one told the business paper Caixin, it is now almost impossible to infer the real technological gap between firms from their price tags alone.
The likeliest outcome is the one the EV story already taught. Capital rushes in on the strength of a category rather than a product, a brutal shakeout follows, and the winners turn out to be whoever quietly accumulated the unglamorous asset: not the smartest model or the most lifelike robot, but the deepest record of what happens when a machine meets the messy, uninferrable real world. Everyone can buy the model. The reality loop is the part you have to earn.