Language models write text; world models are designed to understand what happens when a gripper grasps an object or when oil spills on the factory floor. At the Digital Summit, Arthur Mensch (Mistral AI), Tanja Rückert (Bosch), Robin Rombach (Black Forest Labs), and Bernhard Schölkopf (Max Planck Institute for Intelligent Systems, ELLIS Institute Tübingen) discussed this next class of models. Their conclusion was remarkably unanimous: Europe’s opportunity lies in combining language and world models, trained with data that isn’t available on the internet.

What the Industry Demands

Rückert, who is responsible for digital business at Bosch, outlines what the company brings to the table: more than 230 plants and AI-supported optical inspection across over 1,000 production lines. Internet data rarely contains the unusual situations that matter in a factory. A 95 percent accuracy rate is not sufficient for the industry; delays in processing could be dangerous. The next generation of models must be reliable and secure and run locally on machines, robots, and vehicles—as a combination of powerful reasoning models and fast specialized models for on-site decision-making. In this context, sovereignty does not mean isolation: critical applications require full control, while partnerships and open models make sense elsewhere.

The Limits of Reality

Rombach sees two directions: reasoning models that solve complex tasks and coordinate systems, and visual world models that learn intuitive physics and operate on the device without the cloud. Schölkopf adds that reality cannot be recorded indefinitely; simulation and synthetic data are therefore becoming more important, including for preparing models for rare situations. He counts control over security mechanisms as part of digital sovereignty.

The Model Factory

Mensch draws the harshest conclusion: Europe must not give up the competition for foundational technology; reasoning models are the foundation for everything else, and having its own “model factory” for reasoning and world models is “absolutely critical” for Europe. There is no shortage of talent, but rather a lack of capital and computing power. Mistral is therefore building its own data centers. According to Merz, the federal government’s High-Tech Agenda promotes research into world models and physical AI, and the AI Task Force’s report focuses on highly specialized models based on proprietary industrial, research, sensor, or process data.

Berlin is banking on NVIDIA

A memorandum of understanding dated October 19 outlines how this is to be implemented in practice. At the Invest in Germany Summit, the Ministry of Economics and the Ministry of Research signed a Joint Declaration of Intent with NVIDIA regarding an initiative for physical AI models. The federal government cites the open model families Nemotron and Cosmos as the basis for this; the Industry 4.0 platform will be fully geared toward industrial AI and is intended to become the central strategic network, spanning everything from infrastructure to applications. The announcement does not specify funding amounts, a timeline, or specific projects. The Handelsblatt reported in advance that federal funding would be in the range of hundreds of millions, partly for computing time in Telekom’s Industrial AI Cloud, and that there would be about a dozen partners, including Siemens, Audi, and Black Forest Labs. However, this has not been confirmed. The ambivalence is clear: Europe’s global models are to be developed based on the model families of a U.S. corporation whose hardware also powers the Munich-based Industrial AI Cloud.

Interesting Facts about Erlangen

Siemens and NVIDIA are demonstrating what a world model can achieve in manufacturing at the Erlangen electronics plant, which has approximately 1,000 production employees. Agents in Siemens’ Intelligence Center X, built using NVIDIA’s Factory Operations Blueprint, compare materials, line capacity, delivery dates, and changeover costs for rush orders and suggest options. Planners review the change in the digital twin, and a production controller approves it before the agent implements permissible changes. For printed circuit board inspection, the Cosmos world model, together with Teamcenter and Process Preparation X, generates a virtual “golden board” and synthetic images of good and defective boards. The goal is to reduce manual engineering effort for test programs by up to 85 percent. Siemens is investing 500 million euros in the Erlangen Technology Campus. At GTC Berlin on October 21, the company expanded its self-verifying AI agents to include electrothermal analysis of 3D chips; engineers can now evaluate up to 50 times more design variants, says Ankur Gupta, Executive Vice President at Siemens Digital Industries Software.

Locally, under control, using proprietary data

For decision-makers in manufacturing, the question is therefore: What data is used for training, who owns the models, and where do they run? Bosch, Mistral, and Black Forest Labs all give the same answer: locally, under control, using proprietary data. For now, the infrastructure for this comes from California.