Physical AI in Europe: the map, and the missing layer

Ask a European robotics founder whether the continent can win in physical AI, and you often get a tired look. The same question was asked about search, about social networks, about cloud, about large language models. Each time Europe funded the research and then watched the products ship from California and Shenzhen.

This wave may break differently, and the reason is money. Across 2025 and into 2026, European humanoid and physical AI companies closed rounds that would have been unthinkable for a hardware startup a few years earlier. Old industrial names started running humanoid pilots on the factory floor. The capital and the customers are, for once, in the same time zone.

So the honest question is not whether Europe is in the race. It is. The sharper question is what, specifically, Europe could own for a decade. Most coverage maps the field by robot or by model. This piece maps it by one axis that quietly decides the rest: where does each player get its training data.

The European map is more crowded than the headlines suggest

Start with the loud number. The Robot Report covered NEURA Robotics, the German humanoid maker, announcing a funding round reported on the order of $1.4B, among the largest ever for a European robotics company. NEURA also describes a training center built with the Technical University of Munich, its RoboGym, as the largest Physical AI training center in Europe. Both figures come from NEURA and its coverage rather than an independent audit, but the ambition is real and it is well capitalized.

Underneath the headline sit the industrial backers. German and wider European manufacturers have run humanoid pilots on real production lines, and component giants have taken stakes in humanoid startups; the bearings maker Schaeffler, for one, has backed the sector. The International Federation of Robotics has long documented that Europe installs industrial robots at scale and hosts a deep automation supply chain. The pattern is familiar: strong research, strong manufacturing, a persistent gap at the point where a lab result becomes a shipped product. The continent is not starting from zero. It is starting from a factory.

The moat is not the robot, and it is not the GPU

Here is the uncomfortable part for anyone betting on hardware alone. The robots are converging. Look across Figure, 1X, Agility, Boston Dynamics, and NEURA and you see the same rough recipe: a bipedal or wheeled humanoid, dexterous hands, a few dozen actuators, a sensor head. The mechanical gap between the leaders is shrinking, not widening.

The models are converging too. Most serious embodied-control efforts now train some flavor of vision-language-action (VLA) model, and the recipes are increasingly published in the open. NVIDIA's Isaac GR00T ships an open humanoid foundation model with the tooling around it. Reporting in IEEE Spectrum tracks how fast these architectures spread between labs. A good idea in a robot-learning paper is running in a dozen places within months.

Compute tells the same story. GPUs are a purchase, not a birthright; you can rent the same accelerators in Frankfurt that a lab rents in Fremont. None of these three inputs, hardware, model, or compute, stays put. Each of them crosses a border on a purchase order.

Hardware converges, models are shared, and compute is a rental. The one input that will not move across a border on demand is a lawful record of real human hands doing real work.

Where each player actually gets its data

That last input is the interesting one, because it does not scale the way text did. Nobody left a reservoir of synchronized head-camera, force, and hand-pose recordings lying around to be scraped. Every serious team has to source demonstrations somehow, and the method it picks shapes the cost, the speed, and the legal footing of everything downstream.

Read down the right-hand column. For most rows the supply sits with a robot company in the US or China, a global simulator vendor, or a loose research consortium. The pooled datasets are genuinely valuable, Open X-Embodiment and DROID among them, but they are small next to what a foundation model wants, and their licensing is a patchwork. The bottom row is the only one where jurisdiction is a design choice rather than an accident of where the data happened to be recorded.

Europe's edge is jurisdiction, not just engineering

This is where the map tilts toward Europe. If the scarce input is a record of real people performing real tasks, then that record is personal data, captured from human bodies, faces, homes, and workplaces. How you may collect it, document it, and sell it is not a technical question. It is a legal one, and Europe has already written the law.

The EU AI Act phases in obligations around data governance, documentation, and provenance, while the GDPR already governs consent for personal data. For a data operation in the US or China, these rules read as friction. For a European one, they read as a specification. Capture demonstration data inside the EU, with informed consent and a documented chain of provenance, and you produce something a buyer in a regulated market can actually defend: what the category is starting to call sovereign robot data, or EU-jurisdiction demonstration data.

The corporate structure matters as much as the camera. A dataset held by an EU-domiciled entity, under EU jurisdiction by corporate structure, is insulated from foreign disclosure demands in a way that the same data on a US-owned cloud is not. For an OEM that has to explain its training data to a regulator, an auditor, or a works council, provenance stops being paperwork and becomes a reason to buy.

So the map of physical AI in Europe is not really a map of robots or of models. Those are converging toward a shared global baseline, and no single flag will own them. The part that can carry a jurisdiction is the data: who captured it, under what consent, and which court governs the entity that holds it. That is the missing layer on most maps of the field, and the rest of this cluster follows it down, into how egocentric data gets captured in Europe, what the CLOUD Act means for robot data, and why provenance is becoming a feature you can sell.