Machine vision got cheap. The silicon underneath it did not get plural.

The economics of industrial vision changed when inference moved off the plant server and onto a module inside the machine. Bin picking, weld inspection and quality sorting stopped being integration projects and started being product features.

The chips that made this possible are embedded accelerators: a few families from a small number of vendors, each with its own toolchain, quantisation behaviour and model conversion path.

Why substitution is hard

A robot builder can usually swap a microcontroller with a board revision. A vision accelerator is different, because the value sits in the software layer around it. Models are trained, quantised and validated against a specific runtime, and safety-relevant inspection functions are validated against measured performance of that exact combination.

Change the accelerator and the validation evidence does not travel. In regulated production, particularly automotive and medical device manufacturing, that is the expensive part, not the board.

What European builders are doing about it

Three approaches recur among firms interviewed across Denmark, Germany and Switzerland. Abstracting the inference layer behind an internal interface so the model pipeline is portable, even at some performance cost. Keeping a validated fallback path on general purpose compute for functions that cannot stop. And, less often, standardising on open runtime formats so model assets are not locked to one vendor's converter.

None of these removes the dependency. They reduce the cost of changing it from a product redesign to a planned engineering project, which is a realistic goal where independence is not.

The policy point

European industrial policy has concentrated on fabrication capacity. For vision, the binding constraint is software ecosystem depth rather than wafer supply. A European accelerator with a weak toolchain does not get designed into products, regardless of where it is manufactured.

That is an unglamorous conclusion and it is the one builders repeat. The chip is not the lock-in. The compiler, the model zoo and the validation evidence are the lock-in.