How Europe is catching up with "Modular AI": The price trap of major US language models – Image: Xpert.Digital
The Architecture of Freedom: Why Europe Must Rely on Modular Language Models
Whoever controls the models controls the knowledge – and Europe is still just watching
The global market for large-scale language models resembles an oligopoly with a familiar pattern. A few US technology companies determine which models are available, under what conditions they may be used, and which information architectures they support. In the enterprise segment, three providers shared the lion's share in 2025: Anthropic controlled around 40 percent of enterprise spending on language models, OpenAI accounted for 27 percent, and Google for 21 percent. The entire US enterprise market for generative AI tripled to approximately $37 billion. European providers play no measurable role in these statistics.
This concentration is not just an economic problem; it is a problem for democracy. Monolithic language models operate as black boxes for their users. Their training data, internal weightings, bias structures, and decision-making logics remain opaque. In an open society that relies on diversity of opinion, verifiability, and institutional oversight, this lack of transparency poses a systemic risk. Autocratic regimes can use centralized AI architectures as instruments of surveillance and information control. Democracies need the opposite: transparency, modularity, and the capacity for self-correction.
Related to this:
The fairy tale of open AI from overseas
The common answer to the sovereignty problem is often that Europe can rely on open-weight models from the United States or China. This approach is naive and strategically short-sighted for several reasons.
Open-weight AI models like Meta's Llama family operate under one-sided community licenses that can be modified, restricted, or revoked at any time. The corporations behind these models are not acting out of altruism, but rather out of strategic calculation. In July 2025, Meta demonstrated its disregard for European interests by refusing to sign the voluntary EU AI Code of Practice. Joel Kaplan, Meta's Vice President for Global Affairs, publicly stated that Europe was on the wrong track regarding AI and criticized the code as over-regulating and stifling innovation. This is noteworthy because Meta simultaneously plans to aggressively position its AI models in the European market, for example, by integrating them into Qualcomm smartphones and Ray-Ban glasses.
Chinese models like DeepSeek are technologically impressive. DeepSeek V3 was trained for a mere $5.6 million, while GPT-4 cost between $78 and $191 million. However, for security-relevant, industrial, or public applications in Europe, Chinese models are often unsuitable, whether for regulatory, geopolitical, or data protection reasons.
The real problem lies in the platform economy's playbook: US companies lure customers with low entry prices and transparent weightings. Companies implement these models in their processes, replace human workers with machines, and become dependent. Once this dependency is established and the models are mature, prices rise. Customers have to pass these costs on, without any guarantee that their customers are willing to accept the increased prices. OpenAI can afford aggressive pricing strategies because ChatGPT subscriptions alone generate $3.6 billion annually, thus cross-subsidizing API prices. European companies don't have a comparable bargaining position in this game.
The investment gap: Europe's structural deficit
The figures speak for themselves. In 2023, an estimated $8 billion was invested in AI in the EU. In the United States, it was $68 billion, and in China, $15 billion. European AI startups attract just 6 percent of global AI funding, while US startups receive 61 percent. The European Commission has announced a €200 billion program with its InvestAI initiative, of which €50 billion is to come from public funds and €150 billion from private investors. Whether these sums will actually be mobilized remains to be seen. By comparison, the Trump administration alone pledged $500 billion for comparable AI development programs.
Against this backdrop of declining transatlantic reliability, Europe faces a fundamental strategic decision. So far, it has not been possible to pool data, talent, and financial resources in such a way as to create basic models with several hundred billion parameters in numerous European languages. The institutional hurdles between countries, research institutions, and companies are considerable. Corporate politics, siloed thinking, and regulatory requirements often prevent even the merging of comparatively modest amounts of data.
Modular intelligence: Europe's asymmetric advantage
If Europe cannot win the race for the largest monolithic model, it must change the rules of the game. Modular architectures offer precisely this possibility. They require significantly fewer resources in terms of GPUs, data, and talent, and can be developed decentrally. This is a crucial aspect in times of uncertain markets and often short-term research budgets.
The central building block of modular approaches is the Mixture-of-Experts (MoE) architecture. Large models like ChatGPT, DeepSeek, and Mistral already use MoE mechanisms internally. For each input, only selected specialized experts are activated, thus using computing resources efficiently. The Allen Institute for AI has significantly advanced this approach with FlexOlmo and released it as a commercially available open-source solution. FlexOlmo uses a 7x7B architecture with a total of 33 billion parameters, where each expert is trained independently on local, non-shared datasets. The results are remarkable: a 41 percent relative improvement over purely public models and a 10.1 percent superiority over previous merging methods, confirmed across 31 benchmarks and presented at NeurIPS 2025.
The key to FlexOlmo is its paradigm of data collaboration without data sharing. Each data owner creates their expert locally, based on a shared public base model. A router learns which experts provide the best answers to which queries. Experts can be activated or deactivated at any time, and in a targeted reconstruction attack, a maximum of 0.7 percent of the training data could be recovered. With pseudonymization measures, this figure could be reduced to below 0.1 percent, which would even meet stringent European data protection requirements. This concept is suitable for use both within a corporation across divisions and for distributed learning between multiple companies.
A new dimension of digital transformation with 'Managed AI' (Artificial Intelligence) - Platform & B2B solution | Xpert Consulting
A new dimension of digital transformation with 'Managed AI' (Artificial Intelligence) – Platform & B2B solution | Xpert Consulting - Image: Xpert.Digital
Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.
A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.
The key advantages at a glance:
⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.
🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.
💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.
🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.
📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.
More information here:
Project SOOFI: Germany's AI factory is developing the European answer to ChatGPT
Reasoning models: Logic instead of size
A second crucial component is Large Reasoning Models. Models like ChatGPT-o3, DeepSeek R1, or OLMo 2 are designed to solve complex problems through step-by-step, logical reasoning, creating coherent chains of argumentation. They utilize techniques such as chain-of-thought prompting to break problems down into individual steps and symbolic reasoning to analyze logical relationships. The year 2025 was widely dubbed the Year of Reasoning, a year in which RLVR and GRPO placed the teaching of models for logical reasoning at the heart of their development efforts.
Of particular relevance for Europe is the cost-efficiency of these models. Training DeepSeek R1 based on DeepSeek V3 cost only an additional $294,000. Reasoning models use and extend the knowledge from the base models, which is why they can be built even with limited computing infrastructure. Domain-specific reasoning models already exist for coding, mathematics, and medicine. The SOOFI project explicitly plans to develop a reasoning model alongside the basic LLM.
This opens up concrete business opportunities for companies: customer inquiries, error analyses, legal reviews, and preliminary medical assessments can be processed automatically and transparently. This not only saves time but also reduces the costs associated with errors. Medium-sized businesses and specialist departments can develop customized AI solutions without large investments, initially based on existing open-source models and later migrated to a European base model.
Related to this:
Agents in test-time compute: Intelligence at runtime
The third component of modular systems is agents in test-time compute. In this approach, a language model initially generates potential answers during inference. Highly specialized agents then independently verify these answers. The key advantage: Test-time compute costs have decreased significantly over the years, and model adjustments during training are unnecessary.
The most impressive example of the power of this approach was provided by Microsoft with its AI Diagnostic Orchestrator. MAI-DxO utilizes five specialized AI agents, each fulfilling different medical roles: a hypothesis generator, a test selector, an evidence interpreter, a consensus builder, and a final diagnostician. In a comparison using 304 complex cases from the New England Journal of Medicine, the system achieved a diagnosis rate of 85.5 percent, while experienced physicians, under limited conditions, correctly diagnosed only 20 percent of the cases. Simultaneously, the system reduced the need for laboratory and imaging tests by 28 percent.
This generator-verifier paradigm can be implemented by individual companies, even with their own IT staff. Agents can be developed independently, enabling distributed development. Many companies can now afford this approach because no complex model adjustments are required.
The SOOFI project: Europe's answer is taking shape
The SOOFI project demonstrates that Europe is not only theoretically but also practically capable of taking action. SOOFI stands for Sovereign Open Source Foundation Models and is one of the most ambitious projects for strengthening European AI sovereignty. A consortium of six German research institutions, including Fraunhofer IAIS, Fraunhofer IIS, DFKI, and the Universities of Würzburg, Hannover, and TU Darmstadt, is developing an open language model with approximately 100 billion parameters together with two startups.
The German Federal Ministry for Economic Affairs and Energy is funding the project with €20 million until July 2026. The model is being trained in T-Systems' Industrial AI Cloud, one of Europe's largest AI factories with over 10,000 GPUs, a computing power of 0.5 exaFLOPS, and a storage capacity of around 20 petabytes. SOOFI is intended to replace the existing Teuken-7B model, which Fraunhofer developed in 2024 as a multilingual European model with seven billion parameters. In addition to the basic model, a reasoning model capable of structured thinking and solving multi-stage problems is also being developed.
Funding is provided through the 8ra initiative, established by twelve EU member states. In parallel, Germany and France have launched another initiative, the Franco-German AI Executives' Dialogue, involving leading European companies such as Siemens Energy, Deutsche Telekom, Arte, and Schwarz Digits. The goal is an industry-oriented, implementation-focused AI roadmap for Europe, driven by Fraunhofer, Inria, and the Institute Mines-Telecom as core partners.
The triad of European sovereignty
The technological building blocks result in a concrete three-stage plan that is feasible within the existing European framework.
The first step involves promoting a European baseline model as a mixed-experts initiative, designed as an open-source infrastructure measure. Developing a high-performance, open model is the digital equivalent of the electricity or transport network. SOOFI and Teuken form the starting point. The baseline model can be gradually expanded with high-quality, domain-specific data and as a Model of Enterprise (MoE) architecture.
The second step involves building specialized reasoning models, supported by companies. These projects are significantly less complex than training base models. Reasoning models would initially build upon existing open-source base models from the US or Mistral and later migrate to a European base model. Smaller teams could achieve substantial results with budgets in the six- to seven-figure range.
The third step involves expanding the use of agents in test-time compute, creating modularity, feedback loops, and ecosystems. Companies can extend models with agents in parallel. The resulting feedback data improves the reasoning models, which in turn enrich the base models with additional world knowledge. This creates a circular system that improves itself with each new expert added to the base model. This learning ecosystem would be open to businesses, academia, and open-source communities.
The window is closing: Action instead of hope
The strategic situation is clear. As long as access to open models is maintained, Europe can pursue the path of modular language models. The prerequisites are in place: a high level of vertical integration in industry, a rich talent pool at universities and research institutions, and a regulatory framework that demands transparency and data protection, which, with modular architectures, is not a disadvantage but a competitive advantage.
However, this window of opportunity is not unlimited. While the trend toward regional and specialized language models is increasing worldwide, the dominance of US providers is solidifying with each passing quarter. By 2026, a clear shift from monolithic language models to specialized, autonomous AI agents will be evident. European companies that fail to develop their own expertise now will be entirely dependent on external providers within a few years, similar to the situation with cloud services, where Europe has become a mere user of foreign core technologies.
The necessary technologies exist, the concepts have been tested, and the first projects are underway. What's lacking is not technical feasibility, but the political and entrepreneurial will to scale up these approaches. Europe faces a choice between technological autonomy through smart architecture and perpetual dependence through inaction. The decision must be made now.
Your global marketing and business development partner
☑️ Our business language is English or German
☑️ NEW: Correspondence in your native language!
Konrad Wolfenstein
I and my team are happy to be available to you as your personal advisor.
You can contact me by filling out the contact form here wolfenstein@xpert.digital:or simply call me at +49 7348 4088 965. My email address is
I'm looking forward to our joint project.
