In 19 centers spread across Europe, 1.5 billion euros worth of supercomputers await their users. The European Union has built its AI Factories infrastructure to democratize access to intensive computing, with a simple promise: 50,000 free GPU hours for every SME that requests them. But the first months of operation reveal a more complex reality.
The real battle is not being fought over the quantity of available GPUs, but over who controls access, how queues are managed, and what standards define services. The HORIZON-JU-EUROHPC-2026-COAIF-03 call for proposals, endowed with 25 million euros to standardize AI services, confirms that Europe has understood the stakes: governing critical infrastructure rather than being subjected to it.
The Essentials
- 19 operational AI Factories in Europe with total investment of 1.5 billion euros
- Free access capped at 50,000 GPU hours per SME, upon justified request
- 77 AI Gigafactories proposals received in 16 Member States for the next generation
- 25 million euro call for proposals launched to standardize service access protocols
- Tension between promised democratization and risk of capture by large technology integrators
Factories Operating, Queues Growing
Attendance figures exceed initial projections. The 19 European AI Factories currently process more than 2,000 access requests per month, primarily from SMEs and research laboratories. The Barcelona center records an 87% occupancy rate on its NVIDIA H100 GPUs, Munich reaches 93%. Available time slots are now reserved three weeks in advance.
This rapid saturation masks an unbalanced distribution of usage. According to EuroHPC data, 40% of computing time is monopolized by around fifty organizations, primarily public research institutes and digital service companies that know how to navigate allocation procedures. SMEs represent 65% of applicants but obtain only 35% of actual time slots.
Allocation through equal quotas, designed to guarantee equitable access, produces the opposite effect. The 50,000 free hours per company creates an artificial ceiling that pushes the largest users toward workaround strategies: multiple subsidiaries, technical partnerships with public laboratories, disguised outsourcing to specialized integrators.
The governance of queues becomes a political issue. Who decides that an AI medical diagnosis project takes priority over the optimization of a supply chain? Attribution committees, composed of national representatives and technical experts, struggle to establish objective criteria. Result: the best-documented projects prevail, mechanically favoring organizations with substantial administrative resources.
The 25 Million Euro Call Reveals True Priorities
The European Union has just published the HORIZON-JU-EUROHPC-2026-COAIF-03 call, endowed with 25 million euros over three years, to “standardize and optimize AI services in the European ecosystem.” This envelope, modest compared to the 1.5 billion invested in hardware, targets a strategic challenge: defining access protocols, user interfaces, and security standards that will structure the use of AI Factories.
The specifications reveal internal tensions in the European project. The Union demands open and interoperable standards between the 19 centers, while requesting digital sovereignty guarantees that limit dependency on American or Chinese solutions. Applicants must propose queue management tools that are “transparent and equitable,” without specifying how to reconcile technical efficiency with social equity.
This standardization conceals a power issue. Whoever controls the standards controls the ecosystem. Large European technology companies – SAP, Atos, OVHcloud – are positioning themselves to win this market and impose their proprietary solutions as a mandatory intermediation layer. SMEs risk finding themselves dependent on these new intermediaries to access resources they were supposed to use directly.
The call for proposals also requires developing usage monitoring and auditing tools. Every GPU hour consumed will be traced, every trained model will be catalogued, every produced result will be potentially auditable. This transparency aims to justify public investment but creates conditions for unprecedented industrial surveillance in Europe.
Gigafactories Map Europe’s 2030
The 77 AI Gigafactories proposals submitted in 16 Member States map Europe’s technological geography for the coming decade. These next-generation facilities, equipped with NVIDIA H200 GPUs and future B200s, will target capacities 10 times greater than current AI Factories.
Geographic distribution reveals national ambitions. Germany proposes 18 sites, France 12, Italy 9. Nordic countries are betting on their decarbonized electricity: Norway is candidating 6 installations powered by hydroelectricity, Finland 4 sites cooled by arctic air. Spain and Portugal are leveraging their geographic positioning to serve as a bridge to Africa and Latin America.
This expansion masks a strategic dependency. European AI Gigafactories will depend 95% on non-European components: NVIDIA chips, Samsung or SK Hynix memory, American or Japanese cooling systems. Europe is pooling its computing to catch up with AI giants but remains dependent on global supply chains for critical elements.
Massive public investment – 20 billion euros planned for InvestAI and AI Gigafactories – raises the question of return on investment. Europe is betting that these infrastructures will foster the emergence of European technology champions capable of competing with American and Chinese giants. The risk: financing through European taxes infrastructures that will primarily serve to train models designed and commercialized by non-European companies.
SMEs Against Integrators: The Battle for Real Access
Usage data from the first six months reveals a gap between the stated objective – democratizing AI for SMEs – and observed reality. Companies with fewer than 250 employees represent 71% of access applicants but consume only 28% of actual computing time. The gap is explained by unequal mastery of tools and procedures.
A Toulouse-based technical textile SME specializing in technical clothing obtained its 50,000 GPU hours in March 2024 to develop a fabric defect detection algorithm. Six months later, it had used only 3,200. “The interfaces are designed for IT professionals, not for industrial managers,” explains its CEO. “We ultimately outsourced the development to a Paris consulting firm that masters these tools.”
This phenomenon of disguised outsourcing is becoming widespread. Around fifty European integrators are specializing in accompanying SMEs for AI Factories access. They offer a turnkey service: file submission, model development, industrial deployment. SMEs retain intellectual property but lose technical control.
This intermediation recreates the inequalities that public infrastructure wanted to correct. SMEs that can afford these support services effectively access resources. Others accumulate theoretical GPU hours they cannot use. The objective of European technological sovereignty transforms into a captive market for a new category of intermediaries.
The issue goes beyond simple user training. It touches the very philosophy of the European project: building genuinely accessible public infrastructure or financing through taxes a system that primarily benefits already technically equipped actors.
Open Standards Against Proprietary Solutions
The call for proposals to standardize AI services crystallizes a fundamental tension in European digital strategy. On one side, the requirement for open and interoperable standards to avoid technological lock-in. On the other, pressure from European companies to capitalize on their proprietary solutions in a protected market.
Technical issues mask political choices. Standardizing GPU access interfaces on open protocols would facilitate innovation and competition, but would reduce barriers to entry for non-European technology giants. Favoring proprietary European solutions would protect the market but risks reproducing the lock-in logic Europe denounces in its competitors.
The question of foundational AI models illustrates this dilemma. European AI Factories enable training customized models, but most users start from American pre-trained models (GPT, Claude, Llama) or Chinese ones. European public infrastructure thus serves as a fine-tuning platform for artificial intelligences designed elsewhere.
Several Member States are pushing for AI Gigafactories to prioritize European foundational models. France supports Aleph Alpha and Mistral projects, Germany is betting on its public research ecosystem. This dirigiste approach clashes with European competition rules and resistance from other Member States preferring to maintain technological openness.
The choice of standards will condition the evolution of Europe’s entire AI ecosystem. Europe is betting on quantum standards to not lose the industrial stakes: the lesson could apply to artificial intelligence. Mastering protocols and standards often matters more than owning the machines.
Europe Faces the Test of Technological Governance
European AI Factories work technically. GPUs run, models train, algorithms deploy. But their success is now measured by their capacity to serve the political objectives that justified their creation: democratizing access to AI, reducing strategic dependencies, fostering European innovation.
Early feedback shows that governing critical infrastructure cannot be improvised. Allocating computing resources according to equitable criteria proves more complex than expected. Maintaining open access without creating displacement effects demands constant political arbitration. Reconciling technical efficiency with European sovereignty imposes costly compromises.
The 25 million euro investment in service standardization reveals a growing awareness: Europe will not catch up in AI solely through accumulating computing power. The United States dominates through its ecosystems, China through its integrated industrial strategy. Europe is still seeking its model.
The 77 AI Gigafactories proposals testify to European ambition. Their implementation will depend on the continent’s capacity to resolve contradictions revealed by AI Factories: between openness and protection, between equity and efficiency, between sovereignty and interoperability. The governance of European artificial intelligence is being decided in these technical trade-offs that will determine who controls, uses, and benefits from tomorrow’s tools.