Physical AI: the industrial stack, and where Europe stands

The sober institutional frame: robots that perceive, learn and adapt — reinforcement learning, imitation learning and multimodal foundation models make variability tractable. The most important line sits in there almost as an aside: robotics belongs in the long-term strategy, not on the quick-win list. And the workforce changes with it — the machine operator becomes a robot technician, maintenance goes predictive. Exactly the maturity question my clients actually face.

The US wake-up call with hard numbers: China’s robot density passed the US in 2021, then Japan and Germany in 2024; Unitree sells humanoids for $5,900. Casado and Neuberger call for permissionless innovation and allied alignment — otherwise the gap becomes irreversible. Read from Europe it is doubly uncomfortable: in the America-versus-China frame we do not even appear. A German industrial reader should argue less with the thesis and more with our own pace.

The European answer to the a16z alarm, and it is uncomfortable: in general-purpose robotics Europe has hardly any manufacturers that can match China and the US — against forecasts of ten million AI robots shipping per year within a decade, a dangerous dependency. The recommendations are right: prioritise strategic sectors, improve access to capital, protect hardware strengths from acquisitions. But papers do not replace factories — the classic European gap between analysis and execution remains.

The practitioner’s point of the week, from a robotics CEO: more than half the technology in Europe’s critical infrastructure comes from outside the continent — and sovereignty is not created by decree, taxonomy or investment plan. It is created by companies doing the patient work of engineering, production and deployment, Mittelstand included. Exactly my line since the sovereignty week: posture is not a strategy. Building is.

For once, listening instead of reading: the a16z vision of the electro-industrial stack — simulation, autonomous design and AI-driven operations as a new industrial base, not a modernisation of the old one. The ambition is American-sized, but the bar is right: whoever builds AI-native starts with simulation and software-first, not with a retrofit. The right closing question for European industrials: are we building what comes next — or digitising yesterday?