In seven days, humanoid robotics held three grand masses on two continents — the EMEA Humanoid Robot Summit in Munich on 16 and 17 June, VivaTech in Paris from 17 to 20, and then Automate in Chicago on 22 — and all recited the same new word. In Paris, VivaTech made ‘embodied AI’ (Embodied AI) its central trend, enacting the shift from on-screen AI to an AI that reasons and acts in the physical world. In Munich, the programme aligned sessions titled ‘From AI to Embodied Intelligence’. In Chicago, Automate’s theme no longer asked if humanoids are ready, but at what speed the industry can absorb them. The press saw a simultaneous breakthrough. It is a misinterpretation: ‘embodied AI’ does not name a faculty of the robot, but a training pipeline. And this pipeline has an owner.
The screen word and what it hides
The vocabulary of ‘embodied intelligence’ suggests that the machine has gained a proper competence, housed in its body. The industrial reality exposed this week is the inverse: what circulates under this label is a chain of behaviour production — data generation, simulation training, validation, deployment — whose value concentrates entirely upstream, far from the chassis. Placed side by side, the announcements from the three trade shows read like so many independent press releases. Cross-referenced, they draw a graph that no one has plotted on stage: under almost every flagship demonstration, the same supplier.
The entity graph no one has drawn
Foxconn makes its European debut at VivaTech with an industrial humanoid trained via the simulation-to-factory pipeline designed with NVIDIA, around Isaac GR00T: we simulate, we iterate on site, we deploy in the workshop. ABB Robotics presents its Physical AI Toolchain at Automate, a software stack covering data generation, training, validation, deployment and optimisation to achieve industrial precision — the fruit of a partnership with NVIDIA sealed in March 2026. Richtech Robotics exhibits two DEX robots in the humanoid pavilion of Chicago and dedicates a second stand to its use of Isaac Sim for Sim2Real training. Doosan deepens at the same trade show its partnership with NVIDIA, integrating the chipmaker’s computing platforms into its two-armed robots, its humanoids and its Bobcat machines.
The pattern does not stop at the exhibitors: it structures the venue itself. The first humanoid pavilion of Automate is sponsored by NVIDIA; the trade show’s Startup Challenge, endowed with 10,000 dollars for its ten finalists, is co-sponsored by NVIDIA and Microsoft. And where a champion escapes the chipmaker, it is not that it frees itself from the rented model — it simply changes landlords: Boston Dynamics’ Atlas runs on Google DeepMind’s Gemini Robotics foundation models, a vision-language-action intelligence charged with real-time adaptation and the interpretation of complex goals.
The hard metric, drowned in folklore
Meanwhile, the screens sell spectacle. HABS and Unitree offered VivaTech an experience presented as ‘telepathy’: impressive, but multiple choice — the robot does not read the raw content of thoughts, it identifies which of three predefined categories corresponds to the captured cerebral activity. AgiBot had eight robots dance synchronised in front of two thousand people; Unitree’s humanoids boxed and balanced against visitors. This folklore masks the only figure that counts: the global humanoid market weighs 4 to 5 billion dollars in 2026. A trifle, on the scale of the industry rushing in — and whatever the winner of the body, the brain is rented from two or three American addresses.
The fracture is not what we think
The convenient opposition has been repeated often: China does volume, the West does narrative. Chinese manufacturers did indeed control 87% of global installations in 2025, led by Unitree and AgiBot. But this figure says the opposite of what we make it say: it measures dominance over the commoditised layer — the chassis, the actuator, the envelope. Proof by the reverse: Foxconn, yet anchored on the Chinese side, trains on NVIDIA. The layer that captures value — foundation model plus simulation — is not Chinese. It is not European either.
For Europe does not just tell stories: it delivers. Wandercraft presented in Paris its Calvin-40, already in production in the Renault factory in Douai to transport tyres on the electric vehicle line. Enchanted Tools showed Mirokaï, a social robot speaking more than 50 languages, deployed in hospitals and airports, of which more than 60% of components are manufactured in Europe. Real robots, real deployments. But bits of bodywork: the question is not whether Europe knows how to build a body — it does — but to note that the training layer that animates this body remains outside its borders.
A consumable and a licence
Here is what the three trade shows say together without any one admitting it alone. The robot’s body has become a consumable, manufactured 87% in China, assembled all over the place, now also in Douai and in Foxconn’s workshops. The brain, however, is rented — at NVIDIA for Foxconn, ABB, Richtech, Doosan and the Chicago pavilion; at Google DeepMind for Atlas. ‘Embodied AI’ is the word that disguises this dependence, and this week’s triptych of trade shows was its clearest demonstration: on two continents, in seven days, the same value chain was exposed in the open. The sector’s feat is not to have embodied intelligence. It is to have convinced that a proprietary pipeline was a faculty of the robot.