Germany is preparing a wide-ranging cooperation agreement with Nvidia to increase industrial AI adoption, Handelsblatt reported through Reuters on 15 October. The reported plan has two targets: physical-AI tools used in areas such as robot training and autonomous vehicles, and high-performance models built for industrial work. Investment would run into the hundreds of millions of euros.[1]
That agreement was not yet an announced government programme when the report appeared. Reuters said the plan was expected to be presented at an Invest in Germany conference the following week and that the German government had not immediately commented. The most useful interpretation is therefore conditional: if confirmed, the pact would add a deployment layer to infrastructure and model projects that already exist.[1]
Germany has already built two pieces of the stack
Three layers of the industrial AI strategy
| Layer | What already exists or is reported | What it is meant to supply |
|---|---|---|
| Compute | A Germany-based industrial AI cloud announced with 10,000 GPUs | Capacity for simulation, engineering, digital twins and robotics workloadsSource 4 |
| Foundation model | SOOFI and its first model, Soofi S 30B-A3B | An open German–English model, reasoning work and early agent use casesSource 2Source 3 |
| Industrial deployment | A reported Germany–Nvidia pact involving hundreds of millions of euros | Physical-AI tools and high-performance industrial modelsSource 1 |
Nvidia announced the infrastructure layer in June 2025: an industrial AI cloud in Germany containing 10,000 GPUs, including DGX B200 systems and RTX Pro servers. The stated workloads range from engineering simulation to factory digital twins and robotics. That cloud provides scarce compute, but hardware alone does not supply a sovereign model or a useful factory application.[4]
SOOFI supplies the model layer. The German research programme runs from September 2025 to September 2026 and aims to produce a larger open-source language model, a reasoning model and initial agent applications for fields including Industry 4.0 and health. Its first reported model, Soofi S, was trained on the German Industrial AI Cloud.[2][3]
A 31.6-billion-parameter model that activates about a tenth at once
Soofi S is a hybrid Mamba–Transformer mixture-of-experts model with 31.6 billion parameters in total and about 3.2 billion active per token. That means roughly 10.1% of the model’s parameters are engaged for a given token, before embeddings are counted. Only six of its 52 layers use attention with a key–value cache; the remaining architecture is designed to keep per-sequence state close to constant as context grows.[3]
What the SOOFI model contributes
| Measure | Value | Context |
|---|---|---|
| Total parameters | 31.6 billion | Capacity distributed across a mixture-of-experts networkSource 3 |
| Active per token | About 3.2 billion | Approximately 10.1% of total parameters, excluding embeddings from the denominatorSource 3 |
| Training data | 26.68 trillion tokens | German–English curriculum with code and detailed source accountingSource 3 |
| Long-context result | 8–9× aggregate decode throughput | Authors’ protocol at 40K context and batch 32 versus selected dense 14–24B modelsSource 3 |
The paper’s openness claim is broader than releasing model weights. The team says it will publish selected checkpoints, per-source data accounting, hyperparameters, training code and evaluation code, with data-construction artefacts where licences permit. Those materials make the model more auditable, although the benchmark and throughput claims still come from the project’s own evaluation protocol.[3]
The proposed pact is trying to connect the layers
The reported agreement matters because industrial AI often fails between a capable model and a functioning operation. Robot training needs simulated environments, sensor data, controls and safety constraints. A factory model needs access to industrial knowledge, software integration and deployment support. Reuters says the planned pact would finance compute and model training around SOOFI while drawing on Nvidia resources.[1]
The strategic question is not whether Germany can assemble another model announcement. It is whether an open model trained on sovereign infrastructure can become a reusable component in German industrial systems without replacing one dependency with another. The model may be locally governed, while the compute stack remains deeply tied to Nvidia hardware and software. The announced terms, participating companies and ownership of resulting models will determine how much sovereignty the third layer actually adds.[1][3][4]
Sources & further reading
- Germany plans AI cooperation deal with Nvidia, Handelsblatt says Reuters · 15 Oct 2026 · Accessed 16 Oct 2026
- SOOFI — Sovereign Open Source Foundation Models DFKI · Accessed 16 Oct 2026
- A Sovereign, Open-Source Foundation Model for German and English arXiv · 10 Jul 2026 · Accessed 16 Oct 2026
- NVIDIA Builds World’s First Industrial AI Cloud to Advance European Manufacturing NVIDIA · 11 Jun 2025 · Accessed 16 Oct 2026