Germany's AI asset is its knowledge. "Open your data" won't unlock it.
· Christoph Heuwieser
On Monday evening Jensen Huang shared a stage with Chancellor Friedrich Merz at the first "Invest in Germany" summit in Berlin. The government used the occasion to announce a declaration of intent with Nvidia on "physical AI" and on bringing industrial AI "systematically into broad use". It is being presented as a big win. I don't think it is one, but the evening did produce one point worth keeping, and it came from Huang.
Huang is right about the data
Huang argued that Germany should treat its data as a national resource: not export it, but use it at home to build its own "national intelligence". He warned against handing over knowledge and values only to buy them back later as imported intelligence, and said AI should extend and scale what German engineers already know rather than replace them.
I agree with the core of that, and I would put it more bluntly. What Germany still has in abundance isn't compute, cheap energy or a growing industrial base. It is knowledge: decades of process data, engineering know-how and hard-won judgement about how plants are built and why they run the way they do. With the industry itself under pressure and likely to shrink further, that knowledge may well be the most valuable thing the country owns.
The deal itself
Per the Tagesschau report, the federal government gave no figures for its own contribution. According to a media report it cites, the project is to be supported with a three-digit million sum, and part of that money is meant to rent compute in Deutsche Telekom's AI cloud.
If that is how it lands, it is industrial policy picking a winner: taxpayers' money paying for compute in one national incumbent's data centre, with a US chip vendor's models on top. That looks more like a planned economy than like a market. And nobody should pretend that a few more GPUs at Telekom will change anything for the mid-sized plant that is supposed to benefit. Renting GPUs was never the hard part.
The same money could do something. It could fund ideas that have a real chance to scale: young companies building industrial AI products that can be sold beyond Germany's borders, or work that removes the regulatory hurdles described below. Either would enable people to build something. A compute contract mostly enables a press release.
To me this looks like a familiar pattern. Politicians spend taxpayers' money, pat themselves on the back, and leave everyone a little poorer, instead of creating the conditions in which people and companies can create value themselves. I suspect part of the reason is uncomfortable: really enabling people means fewer rules, lower barriers and, in the end, a smaller share of the economy passing through the state. That is not a direction Berlin seems to want.
"Open up your data" won't work
The part that worries me most is the Chancellor's appeal to companies to open their valuable industrial data to AI. It sounds right on a summit stage. In practice it fails on several levels at once.
German companies struggle to sell services globally. The value of data is realised in software and services, not in the raw files. A manufacturer that opens its data is, in most cases, handing that value to whoever can build the service, and today that is rarely a German company.
The data isn't theirs alone to open. Industrial data is wrapped in contracts. Customers bind their suppliers to strict information-security rules: ISO 27001 processes, NDAs, data-processing agreements and, in automotive, TISAX assessments. A supplier's production data is often also its customer's design data. Even a company that wanted to open up would first have to unwind a web of obligations that it doesn't control.
The GDPR is in almost everything. Shift logs, operator IDs, maintenance tickets with names in them, emails attached to orders: very little real plant data is free of personal data. Each use has to be justified, minimised and documented.
And that is before the data that really matters. Anything about people, and above all health data, sits under special rules. That is where some of the biggest gains from AI are, and it is exactly where "just open it" is least realistic.
Without real regulatory change, not another summit, the appeal will mostly produce pilots that never leave the NDA.
Move the knowledge, not the data
So how do you use a nation's knowledge without opening its data? I think the honest answer is that much of what makes a German plant work isn't in any one customer's records. It is in the structure: how the systems are built, where orders get re-keyed by hand, why the warehouse and the planning system disagree, which small fixes move on-time delivery and which don't.
That structure can be modelled without a single real customer record. This is what we have been building with World-Sim: simulated companies that run around the clock, with siloed systems (ERP, MES, warehouse, planning, mail, logistics), staff who key orders late and sometimes wrongly, and inefficiencies built in that nobody points out. Agents start with no map. They have to find the systems, ask for access and work out where the work gets stuck. An evaluation oracle forks the live company in two, runs a proposed change on one copy, and returns the difference in each KPI with a signed receipt.
None of it involves anyone's real data. It holds the kind of knowledge Huang is talking about, written down as models, and agents can practise on it. A simulated company is not your plant, and a change that wins in the sim wins against that model's assumptions. But it is a way to put German process knowledge to work for AI today, without waiting for a regulatory change that may not come, and without asking anyone to hand over what their contracts forbid them to share.
If Germany wants "national intelligence", I'd start there: make the knowledge usable, and leave the data where its owners are allowed to keep it.
a4sx is an exchange for AI coding-agent sessions: publish, share and continue real work, dead ends included. World-Sim is its simulated-company test bed for agents.