Europe's digital breakthrough? Breaking free from the US trap: How Europe is building a completely new AI infrastructure with the SOOFI project

Forget chatbots: Why Europe's new mega-AI SOOFI relies directly on autonomous agents

Safe from the US Cloud Act: This is Europe's ingenious plan for sovereign enterprise AI

Europe is caught in an AI trap. While American tech giants like OpenAI, Google, and Microsoft almost completely dominate the artificial intelligence market, the old continent risks being relegated to the role of a mere technology consumer. For European companies, this means not only a massive outflow of added value but also an incalculable legal risk – especially when US authorities can access sensitive company data via the Cloud Act. But now, industrial and scientific resistance is forming: With the "SOOFI" (Sovereign Open Source Foundation Models) project, a consortium of leading German research institutions and startups is venturing to build its own sovereign AI infrastructure.

This is explicitly not about programming just another witty chatbot for consumers. SOOFI pursues a far more ambitious goal: a 100-billion-parameter model, trained on European servers and compliant from the ground up with the stringent EU AI Act. It is intended to serve as a legally sound foundation for highly specialized reasoning models and autonomous AI agents that will take on complex tasks in European industry in the future. The following article examines why SOOFI is radically changing the debate surrounding Europe's digital sovereignty, what enormous opportunities the project holds for the economy – and what immense hurdles it still faces.

Related to this:

• AI tools, co-pilots, agents and autopilots

SOOFI – Europe's sovereign AI infrastructure

When Europe no longer asks questions but decides for itself – and why that sounds more dangerous than it actually is

For years, Europe watched as American tech giants laid the foundations of the digital economy. Now, a consortium of leading German research institutions is undertaking one of the most ambitious attempts to structurally break this dependency – not with another chatbot, but with a sovereign, foundational infrastructure for artificial intelligence. The project is called SOOFI, which stands for Sovereign Open Source Foundation Models. And it places the debate about European AI sovereignty on a new, more concrete footing.

The starting point: A continent as a pure consumer of technology

A sober look at economic reality leads to an unsettling conclusion. Europe, which likes to present itself as a regulatory power in digitalization, has been relegated to the role of almost entirely an importer when it comes to the use of artificial intelligence. In the market for generative AI models and platforms, OpenAI and Microsoft together hold around 69 percent of the global market share. ChatGPT alone accounts for over 85 percent of all AI chatbots used in Europe. Furthermore, Amazon, Google, and Microsoft control around 65 percent of the global cloud market. Three out of four computers in Europe run on Windows, while iOS and Android dominate the smartphone market with a combined market share of more than 99 percent.

These figures do not describe a natural phenomenon, but rather the result of strategic investment decisions that Europe has failed to make for more than two decades. The consequences are by no means merely technical. European companies that build their AI infrastructure on American platforms are simultaneously subjecting themselves to a legal framework they did not help shape and that systematically relegates their own interests to a secondary role.

Particularly concerning is the impact of the US CLOUD Act (Clarifying Lawful Overseas Use of Data Act), which came into force in 2018. This federal law authorizes US law enforcement agencies to request data from American cloud providers—regardless of where that data is physically stored. Whether company data resides in a data center in Frankfurt, Dublin, or Amsterdam, if the service provider is a US company, it is subject to potential access by US authorities. This situation fundamentally contradicts the European General Data Protection Regulation (GDPR) and creates a legal gray area that has become a serious operational risk for companies in regulated sectors—from financial services to medical technology.

The dependency is not limited to data privacy issues. US providers can unilaterally change their pricing, terms of service, and data access. What appears to be a reliable infrastructure today may be available tomorrow under different conditions or not at all. European companies that have built their AI-driven core processes on such platforms face a structural dependency risk comparable to the pattern known in the cloud sector: They build on someone else's foundation, pay ongoing rent, and have no control over the stability and conditions of the underlying infrastructure.

The conceptual core: What SOOFI really is and why the question is wrongly posed

The SOOFI project is often described in public communication as "Europe's answer to ChatGPT." This phrase is catchy, but misleading. It encourages people to measure SOOFI against the standards of a consumer product—by language quality, humor, image generation, or the ability to create recipes. That is not the relevant frame of comparison.

SOOFI stands for Sovereign Open Source Foundation Models and is a research project developing an open Large Language Model with approximately 100 billion parameters. The model is intended to serve as a sovereign foundational infrastructure upon which companies, government agencies, and research institutions can build their own industry-specific applications – without having to make legal compromises or submit to a foreign legal framework. The crucial difference lies not in the performance of the foundational model compared to GPT-5, Claude, or Gemini, but in its structural nature: it belongs to no one and therefore belongs to everyone.

Every European company, government agency, and research institution can use the model free of charge and run it on their own servers. AI Act compliance is built into the model from the outset – not as an afterthought, but as a design principle. The model is trained in 24 EU official languages, with a particular focus on German. It succeeds Teuken-7B, the previous European language model with seven billion parameters from the OpenGPT-X project. SOOFI thus represents a leap of more than an order of magnitude – from seven to approximately one hundred billion parameters.

SOOFI's true strategic ambition, however, lies not in the language model itself, but in what is to be built upon it. The project is designed in three stages: first, a basic language model; second, specialized reasoning models built upon this; and finally, autonomous AI agents. Reasoning models are systems that do not simply generate answers, but solve multi-layered problems through structured inference—they analyze complex technical, regulatory, and organizational relationships and can access additional information sources as needed. AI agents go a step further: they act instead of merely responding. They conduct regulatory analyses, optimize production processes, and prepare medical decisions.

The consortium: Scientific excellence as a foundation

SOOFI is not backed by a single company or a venture capital-backed startup, but by a broad consortium of six leading German research institutions and two innovative startups. The consortium is led by the German AI Association, which acts as a strategic interface between research, startups, and industry.

The participating institutions include the Fraunhofer Institute for Intelligent Analysis and Information Systems (IAIS) and the Fraunhofer Institute for Integrated Circuits (IIS), the German Research Center for Artificial Intelligence (DFKI), the L3S Research Center at Leibniz University Hannover, the Technical University of Darmstadt, the University of Bonn, the Julius-Maximilians University of Würzburg, and the Berlin University of Applied Sciences. The scientific foundation is complemented by the startups Ellamind and Merantix Momentum.

Each participating institution contributes specific expertise that makes the overall project possible in its full depth. The L3S at Leibniz University Hannover is responsible for key tasks related to multilingualism, security, and value alignment, develops multilingual datasets for fine-tuning the models, and creates security benchmarks. TU Darmstadt, under the leadership of Professor Kristian Kersting, Co-Director of hessian.AI, is building an innovative data pipeline that uses AI-supported quality checks to collect reliable European training data, is developing the reasoning model, and is researching energy-efficient alternatives to classic transformer architectures to enable more cost-effective AI services in the long term.

The infrastructure: Training on European soil

Training a language model with 100 billion parameters requires a computing infrastructure that simply wouldn't have existed in Europe a few years ago. Now it is available – in the form of Deutsche Telekom's Industrial AI Cloud, operated by T-Systems.

Leibniz University Hannover has commissioned T-Systems to provide the technical infrastructure for SOOFI – a contract worth tens of millions of euros. The Industrial AI Cloud boasts more than 10,000 GPUs with a total computing power of 0.5 exaFLOPS and a storage capacity of approximately 20 petabytes. The data center is connected via four 400-gigabit-per-second fiber optic links and meets the highest standards for data protection, security, and reliability. The infrastructure is located in Germany and is therefore subject exclusively to European law – thus structurally circumventing the CLOUD Act issue.

The partnership between T-Systems and NVIDIA for building the Industrial AI Cloud represents an investment of one billion euros. These figures underscore that this is not a niche academic project, but rather an infrastructure decision with significant industrial implications. The SOOFI model is being trained in one of Europe's largest AI factories – a symbolic and practical marker of Europe's new self-image in the global AI arena.

From March 2026, a network of approximately 1,000 of these GPUs is planned to be activated for training the SOOFI model. The scale of this project underscores Europe's ability to provide computing infrastructure of this relevant size itself – provided the political and economic will exists.

The financing: Public money for public infrastructure

The Federal Ministry for Economic Affairs and Climate Action (BMWK) is funding SOOFI with approximately €20 million until July 2026 as part of the European IPCEI-CIS initiative (Important Projects of Common European Interest – Cloud Infrastructure and Services). This funding is provided through a mechanism explicitly designed to support the development of a European cloud and edge infrastructure.
(Note: The original text referred to BMWE, but the ministry is now called BMWK, or historically BMWi. The current name has been used here.)

Twenty million euros is a modest sum compared to the billions that American tech companies invest in individual training runs. OpenAI is estimated to have spent over 100 million US dollars training GPT-4. However, this comparison is misleading in two respects. First, SOOFI pursues a different goal: not maximum performance in the consumer segment, but a reliable, compliant-by-design basic infrastructure for industrial and governmental applications. Second, a purely cost-based comparison underestimates the leverage of public research infrastructure—especially when the developed models can be used as an open-source foundation for numerous further applications and specializations.

The financing model is conceptually consistent: Public funds finance an infrastructure that is open to all stakeholders. Companies building on SOOFI do not have to pay license fees and are not bound by the terms of service of a private provider. The value arises not from monopolizing the base layer, but from the multitude of industry-specific applications that can build upon it.

The EU AI Act as a competitive advantage: Compliance as a feature, not a burden

One of SOOFI's most remarkable features is its handling of the EU AI Act. While non-European providers predominantly view the European regulatory framework as an obstacle and calculate corresponding compliance measures as subsequent adaptation costs, the AI ​​Act has been embedded in SOOFI's design principle from the outset.

The EU AI Act entered its decisive phase on August 2, 2025: On this date, comprehensive provisions for General Purpose AI (GPAI) models became fully effective. Since then, specific obligations have applied to all models that can be used for many different tasks – such as GPT-5, Claude, or Gemini – including technical documentation, publication of copyright policies, and training data summaries. For models with systemic risk, adversarial testing, incident reporting, and cybersecurity measures are added. The European AI Office has assumed full oversight of GPAI models since August 2025.

Non-European providers wishing to operate in Europe must retroactively adapt to these requirements. SOOFI, on the other hand, develops its model with precisely these requirements in mind from the very first line of code. This is not merely an academic advantage. For companies in regulated sectors—finance, healthcare, critical infrastructure—AI Act compliance is not an optional add-on, but a mandatory prerequisite for deployment. A model that natively meets this compliance significantly lowers the barrier to entry for such companies and eliminates the risk of subsequent regulatory uncertainties.

Our EU and German expertise in business development, sales and marketing - Image: Xpert.Digital

Industry focus areas: B2B, digitalization (from AI to XR), mechanical engineering, logistics, renewable energies and industry

More information here:

• Expert Business Hub

A thematic hub offering insights and expertise:

• Knowledge platform covering global and regional economies, innovation and industry-specific trends
• A collection of analyses, insights, and background information from our key areas of focus
• A place for expertise and information on current developments in business and technology
• A hub for companies seeking information on markets, digitalization, and industry innovations

The three-stage architectural concept: From language to decision

The technical core of SOOFI is its three-stage development concept, which represents a conceptual break with the classic chatbot paradigm.

The first stage is a classic Large Language Model with around 100 billion parameters – a basic language model trained on the 24 EU official languages ​​and serving as the starting point for all further specializations. This foundation differs from its predecessor Teuken-7B not only in its more than fourteen times larger number of parameters, but also in its changed industrial focus and the regulatory requirements embedded from the outset.

The second stage comprises specialized reasoning models. Reasoning refers to the ability of an AI system not only to recognize and reproduce patterns in training data, but also to draw multi-stage logical conclusions, link information from various sources, and argue in a structured manner. For German industry, such capabilities are of immediate practical relevance: They enable the analysis of complex technical, regulatory, and organizational relationships and support well-informed decisions in development, production, and knowledge management. Specific application scenarios range from simplifying bureaucratic processes and supporting craft businesses with cost calculations to guiding startups in technical decision-making.

The third and most far-reaching stage is autonomous AI agents. While a reasoning model performs analysis, an AI agent acts: it executes tasks independently, calls up external systems, processes the results, and makes subsequent decisions. The intended application areas are concrete: conducting regulatory analyses, optimizing production processes, and preparing medical decisions. In medicine, for example, autonomous AI agents offer the potential to fundamentally transform healthcare – as researchers at the Technical University of Dresden have demonstrated in an article published in Nature Medicine. At the same time, the same authors point to the growing discrepancy between the capabilities of such systems and the existing regulatory frameworks. SOOFI addresses precisely this gap by aiming for an agent infrastructure designed from the outset for the European regulatory environment.

Related to this:

• From tool to autopilot: Which ten industries are being reinvented by the AI ​​revolution?

The strategic shift: From ChatGPT competition to infrastructure thinking

Perhaps SOOFI's most significant conceptual achievement lies less in the technology itself than in the reformulation of the question Europe poses to itself. The debate of recent years has revolved around the question: "Do we need a European ChatGPT?" SOOFI shifts this question to: "Do we need European AI agents to prepare decisions for us?"

This is a fundamentally different approach. Demanding a European ChatGPT means competing in the consumer market against providers with a head start of several years and billions of training data points – a structurally hopeless battle. Building a European AI infrastructure that serves as a sovereign base layer for industry-specific agents, on the other hand, means opening up a competitive space where Europe's strengths – industrial depth, regulatory know-how, multilingual competence, and data protection consistency – can truly come to the fore.

The underlying economic policy logic is coherent. Europe boasts highly developed industries with complex value chains: mechanical engineering, automotive, chemicals, pharmaceuticals, logistics, and financial services. For these sectors, industry-specific AI applications are far more valuable than general conversational AI. A model that performs regulatory analyses for the German mechanical engineering sector, is fully compliant with the AI ​​Act, can be run on its own servers, and responds in flawless German, has a significantly clearer benefit than a further optimized English-language chatbot.

The European Commission's report on the state of the Digital Decade 2025 explicitly acknowledged this connection: Persistent strategic dependencies threaten the EU's economic security and technological sovereignty, particularly in the areas of semiconductors, cloud and data infrastructure, and cybersecurity technologies. The Commission calls for renewed action in the areas of digital transformation and technological sovereignty.

Risks and limitations: What SOOFI is not and what remains unclear

A sober economic analysis also requires the honest identification of risks and limitations – and SOOFI has several of those.

First, regarding the timeline: The first version of the model is scheduled for release in the third quarter of 2026. Whether the reasoning model and the AI ​​agent layer will be ready for use by then remains to be seen. Timelines are notoriously unreliable in AI development, and the technical complexity of the project makes delays likely. The three-stage approach—first the language model, then reasoning, then agents—is logically sequential, meaning that delays in early phases will cumulatively impact the overall delivery timeline.

Then there's the question of performance. SOOFI isn't aiming to dethrone GPT-5—and for good reason. With a budget of €20 million and a timeline of a few months, it's impossible to create a model that can compete with systems backed by the entire computing infrastructure of Microsoft Azure or Google Cloud. A blog post from February 2026 put it this way: SOOFI could create a frontier LLM on par with Mistral Large 3—a respectable, but not the most powerful model in the world. This isn't a failure, as long as the benchmark remains accurate. For many industrial use cases, a second-tier model that can be operated with complete sovereignty is more valuable than the world's most powerful model under foreign jurisdiction.

Furthermore, the question of market acceptance needs to be critically examined. Open-source models are not a guaranteed success. Companies that want to run a model on their own servers need the corresponding technical personnel, infrastructure, and maintenance capacity. For many medium-sized enterprises—the central component of the European economic structure—this can represent a significant hurdle. For SOOFI to truly have a broad impact, a complementary ecosystem of service providers, system integrators, and cloud providers will be needed, offering hosted and managed versions of the model—while maintaining sovereignty guarantees.

Finally, the question of further development remains. A model trained only once quickly becomes outdated. The real challenge for SOOFI lies not in the initial release, but in the ability to continuously develop the model, adapt it to new use cases, and keep pace with accelerating global progress. This requires sustainable institutional structures, governance models, and financing mechanisms that extend beyond the current project funding, which runs until July 2026.

The geopolitical environment: SOOFI in the context of European vulnerability

SOOFI is emerging in a geopolitical environment that underscores the project's relevance daily. Under President Trump, Europe's dependence on American technology has transformed from an abstract risk into a tangible competitive disadvantage. What appeared as a reliable partnership under previous US administrations has revealed itself as a structural vulnerability, materializing in concrete price risks, access uncertainties, and political pressure.

Particularly worrying is the measured viability of European companies in the hypothetical case of a complete withdrawal of American technologies: On average, companies indicate they could survive for approximately twelve months without technologies and services from the USA. This figure – even though it describes an extreme scenario – illustrates the extent of the structural dependency and the seriousness of the vulnerability.

The European response to this reality must take place on several levels simultaneously. AI infrastructure is just one of them, but a particularly strategically important one. Artificial intelligence is no longer merely a tool for increasing productivity – it is increasingly becoming the infrastructure itself, upon which other critical systems are built: healthcare, tax administration, production control, and infrastructure management. Those who fail to control the AI ​​foundation will gradually lose control over the systems that run upon it.

Comparative overview: European AI models at a glance

SOOFI is not alone among European AI initiatives, but it occupies a special position. A comparative look at the ecosystem helps to understand the uniqueness of its approach.

Teuken-7B (OpenGPT-X) is an open-source research model with approximately 7 billion parameters, covering 24 EU languages, and was released in 2024. SOOFI is planned as an open-source industrial project with around 100 billion parameters, focusing on EU languages, industrial applications, and agents; its launch is scheduled for the third quarter of 2026. Mistral, from France, takes a mixed commercial and partially open-source approach, is multilingual, designed for efficiency, and is currently actively available. Aleph Alpha, from Germany, is proprietary and has repositioned itself as a commercial, sovereign-oriented provider focusing on enterprise AI and government. APERTUS, from Switzerland, is a smaller open-source project that emphasizes transparency but offers limited scalability.

This overview shows that SOOFI occupies a special position in that it is the only project that explicitly relies on the three-tiered architecture of base model, reasoning, and agents, is publicly funded and open source, and treats AI Act compliance as a central design goal. Mistral, as a commercial European provider, is more advanced in terms of performance but pursues a proprietary business model with corresponding dependency risks. Aleph Alpha has repositioned itself in recent years from an ambitious model developer to a provider of sovereign AI infrastructure. SOOFI fills a gap between the two: It is powerful enough for industrial requirements and sovereign enough for regulated application areas.

Economic implications: What's at stake?

From an economic perspective, the success or failure of a project like SOOFI should not be measured solely by the technical performance of the developed model, but by the long-term consequences for the industrial value creation structure of Europe.

If Europe fails to develop its own AI infrastructure, the result will be an increasing concentration of economic value creation among non-European providers. The pattern is familiar: In the cloud sector, Europe missed the critical moment when its own investments would still have been competitive. Amazon, Google, and Microsoft now jointly dominate around 65 percent of the global cloud market, and European alternatives play only a niche role. With regard to AI infrastructure, Europe is still at this crossroads – but the window of opportunity is closing.

The year 2026 is considered crucial for Europe's AI future: If European companies do not quickly realize significant efficiency gains through AI, the lead held by the US and Asia threatens to become overwhelming. For the German economy, which is grappling with structural challenges in the automotive and energy sectors, AI-driven productivity gains are not an option, but an economic necessity. The question is not whether, but on whose infrastructure these gains will be realized and who will benefit from them.

Another often underestimated aspect is SOOFI's importance for building technological expertise in Europe itself. The project aims to develop expertise along the entire development chain of large AI models – from data and software competence and training to the question of which teams, processes, and infrastructures such projects require. This expertise development has an independent strategic value that extends beyond the specific model: it creates the conditions for Europe to independently conduct research and development in the areas that will shape the next wave of technological innovation.

The real challenge lies after the initial release

When SOOFI releases its first model in the third quarter of 2026, it will be an important step – but not the decisive one. The real challenge begins after that.

First, a community must emerge. Open-source models don't realize their value through the initial release, but through the ecosystem that develops around them: developers who use the model for their own applications; companies that use it for industry-specific fine-tuning; and service providers who offer it as a basis for hosted solutions. Without an active ecosystem, even the most technically advanced model remains an artifact of academic research.

Secondly, a governance structure must be established to ensure the model's continued development beyond the initial funding period. Who decides on future training runs? Who finances ongoing maintenance and updates? Who assumes responsibility for regulatory issues? These institutional questions are at least as complex as the technical challenges of the training.

Thirdly, and crucially: SOOFI must produce applications, not just infrastructure. The most convincing answer to the question of the value of sovereign AI infrastructure is not an academic argument about data sovereignty, but a medium-sized machine manufacturer that automates its regulatory compliance with the help of a SOOFI-based agent, a hospital that prepares diagnostic decisions with a natively AI Act-compliant system, or a government agency that simplifies citizen processes through a system fully compliant with European law. SOOFI's persuasive power will be measured by concrete benefits – and that's exactly how it should be.

The debate about AI sovereignty in Europe has been confined to abstract categories for too long: We need a European ChatGPT. We need regulation. We need investment. SOOFI breaks with this abstraction and focuses on a concrete concept: a sovereign basic infrastructure that doesn't just respond, but acts. This is no guarantee of success. But it is the right starting point for the right question.

☑️ 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 or simply call me at +49 7348 4088 965. My email address is : [email protected]

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business