European Manufacturers Are Running Qwen and DeepSeek on Private Racks — GDPR and China Revenue Are Doing the Selling
Siemens has publicly described running Qwen and DeepSeek on a self-contained LLM platform powered by vLLM, isolated from external API calls. Other German manufacturers are following the same pattern. The motivations are structural, not experimental.
The GDPR Calculation
European firms calling US AI APIs push customer data across the Atlantic. That transfer requires a valid legal basis under the EU-US Data Privacy Framework — the third adequacy arrangement after judges voided the first two (Safe Harbour in 2015, Privacy Shield in 2020). The current framework is holding but carries known legal risk; another court challenge could void it again without notice.
German firms running DeepSeek on their own racks send nothing back to China. No data transfer to a non-adequate country occurs when inference stays entirely local. That eliminates a class of GDPR compliance risk that US API calls cannot.
The logic is straightforward: if a model runs on hardware in Frankfurt, the data never leaves Frankfurt.
The China Market Constraint
Siemens does significant business in China, which ranks among the company’s largest markets. Beijing has pressed buyers toward domestic technology across multiple procurement categories. Chinese models in the AI stack help maintain eligibility and goodwill in that commercial environment — independent of any regulatory mandate.
For large European industrials with China operations, adopting Qwen or DeepSeek is also a positioning decision. It signals alignment without requiring any formal commitment.
The Economics Don’t Always Work
Self-hosting Chinese open-weight models solves GDPR and market access. It does not necessarily solve cost.
Cloud providers pool a single GPU cluster across thousands of customers. A Siemens division running inference alone has to provision for its own peak demand — which means paying for idle compute at the same rate as working compute. The unit economics only work when utilisation stays high. For large, consistent workloads this is viable. For sporadic inference across business units, it typically costs more than API calls.
That constraint limits self-hosting adoption to larger enterprises with predictable, high-volume workloads. Smaller European firms face the same GDPR pressure but lack the scale to make private racks economic.
What the Pattern Signals
Qwen downloads passed 3 billion in a six-month window. Chinese open-weight models now account for over 30% of US enterprise token use and have hit 46% at peak. In Europe, adoption is taking a different form: not API calls through intermediaries, but direct deployment on controlled infrastructure.
The open-weights strategy from Alibaba and DeepSeek is producing a secondary effect its designers may not have planned for. By releasing capable models under permissive licences, they have created a legal path for European firms to adopt Chinese AI without the data transfer question ever arising. GDPR — a regulation designed to protect European users from US surveillance — is functioning as an adoption accelerant for Chinese models.