Dieser Artikel ist auch auf Deutsch verfügbar
AI should be viewed as a new foundational technology similar to electricity, the internet, and the steam engine. Foundation Models – the technological basis for ChatGPT, Claude, and other AI systems – are becoming essential infrastructure for business processes and innovation. These models are often still called Large Language Models, but their versatility and countless use cases make them the general-purpose technology of our time. Yet Europe faces a stark choice: either use Chinese models that refuse to answer questions about Tiananmen Square or rely on US models that could become geopolitical leverage tools at any time.
Foundation Models as a New Foundational Technology
Foundation Models are pre-trained AI models that serve as the foundation for specialized applications. They process not only text but also images, audio, and video. The principle on which they are based (the Transformer architecture) also works well with molecular structures, demonstrating their remarkable range.
The dimensions of European dependency are sobering. While the US “Stargate” project mobilizes $500 billion, Europe’s efforts appear modest: The EU has increased its investments to €200 billion – with €20 billion for “AI gigafactories” – but these sums pale in comparison to actual needs. Meta trains Llama 4 on over 100,000 H100 GPUs, computing power simply unavailable in Europe. The reason: Europe has yet to produce its own hyperscaler. These past failures are now exacting their bitter price.
The Chinese Alternative: Performance with Political Filters
DeepSeek demonstrated with its R1 model that frontier models don’t necessarily require training on gigantic data volumes. With approximately $6 million, its predecessor model V3 achieved top performance – a fraction of Western investments. The innovation combines Reinforcement Learning with Test-Time-Compute. In Reinforcement Learning, the AI learns through rewards and penalties – receiving automated feedback for correct and incorrect answers without human intervention at each step. Specifically, the model was taught to follow a “thinking process” before giving a final answer and to output this process. Test-Time-Compute means: instead of embedding all intelligence upfront during training, the model “thinks” longer for complex questions – like a human who allocates more time for difficult problems. This architecture efficiently shifts computational work from expensive pre-training to the more cost-effective inference phase.
Alibaba’s Qwen family pushes boundaries further: While Qwen3 itself isn’t multimodal, Alibaba offers sophisticated multimodal models including Qwen2.5-VL (Vision-Language), Qwen2-Audio, and Qwen2.5-Omni. Qwen3 stands out as the first Chinese hybrid reasoning model, seamlessly switching between quick responses and deeper analysis. In July, they launched Qwen3-Coder, a coding-focused model family that’s highly competitive. Other Chinese labs have released equally impressive models with exceptional agentic performance (Z.ai’s GLM 4.5, for example). But there’s a catch: Ask about June 4, 1989, in Beijing, and you’ll hit a wall. This built-in censorship isn’t a bug—it’s a feature.
The American Dilemma: Apparent Openness
Meta’s Llama 4 perfectly illustrates the problem. Advertised as “open,” the license explicitly excludes EU users – allegedly due to regulatory hurdles from GDPR and the EU AI Act. This justification masks Meta’s real message: Europe’s data protection standards make Europeans second-class digital citizens.
To compound matters, there’s the benchmark scandal: Meta used a specially optimized version of Llama 4 for well-known leaderboards that isn’t publicly available. This practice further undermines trust in American “openness.” But realistically, Meta’s reputation in this area was already questionable.
Under the Trump administration, the situation is intensifying. AI technology becomes a bargaining chip, with export controls for AI chips serving as political pressure tools. The businessman in the White House understands leverage – and Europe’s dependency provides ample opportunity.
August brought the bombshell: OpenAI released a semi-open reasoning model family in two sizes (20B and 120B), matching the performance of their closed o4-mini. The larger variant runs on a single Nvidia H100 GPU at impressive inference speeds. And it ships with an Apache 2.0 license. This could be the holy grail for European enterprises. And frankly, there’s little to dispute that. Semi-open model development is even explicitly encouraged in the US Administration’s “America’s AI Action Plan.” Still, you have to wonder—where’s the catch?
Europe’s Options: Between Pragmatism and Vision
The small Gallic village Mistral demonstrates what’s possible – and where the limitations lie. The French company offers specialized solutions like Document AI for document analysis (an important use case in a continent with heterogeneous digitalization status) and Devstral for software development. Their general-purpose models Small 3 and Medium 3 reach GPT-4 levels for some tasks. However, they lack parity in multimodal frontier models and reasoning capabilities. While Mistral introduced the Magistral model in two sizes as a reasoning model in early June 2025, it still lacks multimodality – unlike OpenAI, Anthropic, and Google. Additionally, Magistral doesn’t appear to be state-of-the-art. Another concern: Mistral Cloud’s data processing agreements don’t completely exclude routing through the US.
However, DeepSeek’s efficiency revolution shows that success isn’t solely about training on enormous data sets with partially unclear usage rights. DeepSeek’s R-Zero experiment – a model trained exclusively through Reinforcement Learning – wasn’t practical as a laboratory precursor but indicates more efficient development paths. The final R1 combines innovative architecture with advanced training methods.