Europe's AI Advantage Is Your Factory Data. Siemens Just Described the Deal
Europe is losing the AI investment race by a wide margin. Stanford figures put US private investment in AI at $285.9 billion in 2025, against $20.9 billion in Europe. In an interview published by Fortune on 5 October, Peter Koerte, who runs Siemens’ Smart Infrastructure division, made the case for where Europe can still win: AI for industrial processes, built on data that only European manufacturers have.
The interesting part of the interview is not the optimism. It is that Koerte described, quite plainly, the transaction that makes it work. And every manufacturer in Europe will eventually be offered some version of it.
What Siemens is building
Siemens is developing an Industrial Foundation Model which, according to Koerte, could shorten engineering cycles by up to 40%, with use cases in automotive and aerospace. The company plans to invest more than €1 billion in industrial AI over the next three years.
The reasoning is that industrial models are trained on manufacturing and engineering data rather than language. Koerte argued that language models simply do not work for engineering and production, because words are too imprecise, and that an AI which makes up a calculation in engineering creates a real problem.
Europe’s advantage, in his view, is longevity. More than half of the companies on the Fortune 500 Europe are over a century old, and decades of operations in pharma, automotive and chemicals have produced the largest installed base and the data to match.
The deal on the table
That data has to come from somewhere, and Koerte was candid about the difficulty. Companies like BMW or AstraZeneca will not put their data on the internet, so they have to trust someone else to train models on it.
The terms Siemens offers so far: companies share their data in return for access to the AI model and the ability to request new use cases. The data is used only for training, and ownership stays with the company that created it.
Those are reasonable terms, offered by a company with genuine decades-long customer relationships. But they deserve a closer reading, because “ownership retained” and “control retained” are not the same thing.
What ownership does not cover
Once data has been used to train a model, what the model learned from it does not go back to its owner. You keep the files. The knowledge extracted from them now sits in a model that others use too, possibly including your competitors. Removing what a model has learned from a particular dataset is technically very difficult, which is why the withdrawal question has to be settled before training, not after.
None of this makes the trade a bad one. A model trained on pooled industry data may well serve you better than anything you could build alone, and refusing means watching others benefit while you do not. The point is that it is a negotiation, not a formality, and the value of your side of it is your process knowledge.
A mid-size manufacturer is unlikely to get a phone call from Siemens headquarters. You will meet the same trade in smaller form: in the terms of service of industrial software with AI features, in an “allow us to use your data to improve our models” setting in a cloud MES, in a platform that offers better results if you contribute your drawings. The question is identical each time.
Questions to ask before your data trains anyone’s model
Ask exactly what is shared: raw records, anonymised extracts or aggregated statistics. Ask who else can use the model trained on your data, and whether that includes companies competing for your customers. Ask what happens if you stop the relationship: does the model keep what it learned? Ask what you get in return in concrete terms, whether priority use cases, better pricing or a model tuned for your processes.
And ask the question that comes before all the others: is the data actually yours to share? Drawings and specifications from automotive and aerospace customers usually come under confidentiality terms, and those terms rarely anticipated model training.
Koerte is right that Europe’s industrial data is a strategic asset. The open question is whether it stays an asset for the manufacturers who generated it, or becomes one mainly for the platforms that train on it. That depends less on Brussels than on the terms each company agrees to.
“Content generated using AI”