AI Factories: Europe's new infrastructure for AI innovation

Democratizing AI computing power

On October 10, 2025, a significant shift occurred in Europe's AI strategy. The European Union announced the expansion of its AI Factories network with six new locations, including the Netherlands. This is more than an infrastructural announcement - it marks the moment when Europe's ambition to become an 'AI Continent' takes tangible form.

The timing is not coincidental. Since August 2025, the General-Purpose AI obligations under the EU AI Act have been in force. Foundation models must comply with strict transparency and documentation requirements. Meanwhile, European AI startups watch their American competitors grow with access to virtually unlimited cloud computing capacity. This combination - stricter regulation and technological dependency - threatened to undermine Europe's innovation capacity.

AI Factories are Europe's answer to this paradox. They provide infrastructure that complies with EU norms around privacy, transparency and data sovereignty, while simultaneously delivering the computing power needed for cutting-edge AI development. It is infrastructure policy, industrial strategy and geopolitical positioning in one.

What makes an AI Factory different from commercial cloud?

To understand the unique proposition of AI Factories, we must examine what fundamentally distinguishes them from AWS, Google Cloud or Azure. The difference lies not primarily in hardware - although EuroHPC supercomputers rank among the world's most powerful - but in mission and accessibility.

Priority for SMEs and startups

Commercial cloud services operate on a simple economic model: whoever pays more gets more capacity. This creates a natural threshold for startups and SMEs. A training run for a medium-sized language model can quickly cost tens of thousands of euros. For a scale-up with limited runway, this is often unaffordable, effectively limiting innovation to well-funded players.

Built-in AI Act compliance

An often overlooked advantage is that AI Factories are designed to be AI Act-compliant from the ground up. For companies struggling with the regulation's complexity - documentation requirements, risk assessments, transparency logs - this offers a practical shortcut.

Training data is automatically logged in compliance with Article 11 AI Act (technical documentation). Audit trails of model decisions are generated by default. Data governance follows GDPR principles and sector-specific regulations like the Medical Device Regulation for healthcare applications. This compliance is not an add-on implemented afterwards, but baked into the infrastructure.

For a fintech startup developing a credit scoring model - a high-risk AI system under the AI Act - this means required documentation is automatically generated during the training process. For a medtech company building diagnostic AI, training data remains within healthcare-certified environments without extra configuration.

Expertise ecosystem

AI Factories are more than just computing power. They combine infrastructure with hands-on support from AI specialists who help companies with model optimization, architecture choices and troubleshooting. This is crucial because having access to supercomputing doesn't mean you know how to use it efficiently.

An SME producer of industrial sensors wanting to implement predictive maintenance may have excellent domain knowledge but limited experience with training large-scale neural networks. Factory experts help translate the business case into technical architecture, advise on model selection and assist with hyperparameter tuning. This knowledge transfer accelerates time-to-market and dramatically increases the success rate of AI projects.

The Netherlands gets its first AI Factory

The announcement that the Netherlands will have an AI Factory is strategically significant. The country has strong positions in logistics, agrifood, water management and healthtech - sectors where AI has transformative potential but where traditional software companies are less dominant.

Expected sector focus

While exact specifications are still being finalized, based on existing factories and Dutch strengths we can expect the factory to focus on specific use cases:

This sector focus means the factory offers not just generic computing capacity, but also specialized tooling, datasets and expertise relevant to these domains. A Brabant agrifood startup can count on pre-trained computer vision models for plant pathology, while an Amsterdam healthtech scale-up gains access to anonymized medical imaging datasets for model training.

Local access and economic impact

The physical presence of a factory in the Netherlands has concrete advantages over remote access to foreign facilities. Latency-sensitive workloads - for example real-time inferencing for autonomous systems - perform significantly better with local infrastructure. Additionally, the factory creates a gravity center for AI talent, strengthening the ecosystem.

The economic impact extends beyond direct infrastructure access. Experiences from Finland and Germany show that factories function as catalysts for local AI ecosystems. They attract venture capital by increasing the feasibility of AI startups. They stimulate partnerships between universities and industry. They generate spin-offs from researchers who see commercial applications for their technology.

The broader European network: from fragmentation to federation

The Dutch factory is not a standalone facility, but a node in a growing pan-European network that fundamentally differs from the centralized megadatacenters of American cloud giants.

The three waves of expansion

December 2024 marked the first wave with seven factories distributed across Finland (focus on sustainable AI), Germany (automotive AI), Greece (maritime AI), Italy (manufacturing AI), Luxembourg (fintech AI), Spain (agriculture AI) and Sweden (forestry AI). This initial selection reflected Europe's industrial strengths and deliberately chose sector specialization over generic capacity.

March 2025 brought the second wave with six new factories in Austria, Bulgaria, France, Germany, Poland and Slovenia. The addition of Eastern European locations addressed a deliberate strategy to spread AI capacity and prevent innovation hubs from concentrating only in Western Europe. This helps talent retention in regions that traditionally experience brain drain to Silicon Valley or London.

Interoperability and data sovereignty

The federative model has deliberate design choices that distinguish Europe from centralized American or Chinese AI infrastructure. Training data can be shared between factories within strict data governance rules, but remains subject to local sovereignty requirements. A Polish healthtech company can use Dutch medical imaging data for model training, but that data never physically leaves the Dutch factory - only the trained model is transferred.

This architecture solves a fundamental problem for cross-border AI collaboration within Europe. GDPR prohibits in many cases the transfer of sensitive personal data between member states without adequate safeguards. Federated learning via factories makes it possible to train models on data from multiple countries without physically centralizing that data. This opens new possibilities for pan-European AI applications in healthcare, finance and public services.

The Apply AI Strategy: how factories fit the bigger picture

AI Factories are not a standalone initiative but a crucial component of the Apply AI Strategy that the European Commission launched on October 8, 2025. This strategy has one overarching goal: bridging the gap between AI capacity and AI adoption.

The ecosystem of supporting structures

The Apply AI Strategy introduces an integrated ecosystem where different elements reinforce each other:

European Digital Innovation Hubs (EDIHs) function as the local "front door" of the system. An SME in Groningen seeing AI potential for their production process contacts the local EDIH. There they get an assessment: is AI even the right solution? What type of model fits the use case? Is there sufficient data? The EDIH can refer to a factory if heavy compute is needed, or to a Testing and Experimentation Facility if the focus is on compliance validation.

Testing and Experimentation Facilities (TEFs) offer sandbox environments where companies can test AI systems in realistic scenarios without direct compliance risks. For a fintech startup wanting to launch a credit-scoring model - a high-risk AI system under the AI Act - a TEF is the place to validate that the model meets non-discrimination requirements before going into production. The TEF simulates real-world conditions and generates the documentation needed for AI Act compliance.

AI-first policy in practice

A striking element of the Apply AI Strategy is the "AI-first policy" that organizations - especially in the public sector - are encouraged to apply. This means that with every strategic or policy decision, the question is asked: could AI offer a solution here? This is not technology determinism but a systematic way to avoid blind spots.

A municipality stuck processing permit applications would traditionally hire more civil servants. An AI-first approach first asks: can we automate this process with natural language processing that classifies applications and automatically handles standard cases? This frees civil servants to focus on complex cases requiring human judgment.

Crucially, the policy explicitly states that AI should not be applied blindly - benefits and risks must be carefully weighed. This is where EDIHs and TEFs show their value: they help organizations make this assessment without first building deep AI expertise themselves.

Gaining access to AI Factories: practical routes and realities

The most urgent question for companies: how do you actually get the computing power? The system has three primary access routes, each with its own pros and cons.

Route 1: Via European Digital Innovation Hubs

For most SMEs, the EDIH is the logical starting point. The Netherlands has multiple EDIHs distributed across different sectors and regions. The process typically starts with an intake conversation discussing the use case. Is this a good fit for a factory? Are there alternatives that fit better? What is the expected compute need?

The advantage of this route is guidance. The EDIH helps formulate a project proposal, advises on data preparation and can provide support with technical implementation. The disadvantage is that the process can take weeks to months, which can be a barrier for startups with urgent timelines.

Route 2: Direct application at EuroHPC JU

Companies with more technical maturity can directly submit an application to the EuroHPC Joint Undertaking. This requires a detailed project proposal: which models are being trained, what is the compute footprint, what are the expected outputs, how do these align with EU priorities?

This route is faster if you know exactly what you need, but does require being able to write a technically solid proposal. For companies without dedicated AI engineers, this can be a threshold. The approval rate for direct submissions is lower than for EDIH-referred projects, suggesting that EDIH guidance adds value in formulating successful applications.

Route 3: Via research partnerships

Universities and research institutions often have preferential access to factory capacity for academic projects. Companies collaborating with these institutions - for example via an SBIR grant or consortium project - can piggyback on this access. This combines computing power with research expertise and also opens doors to talent acquisition.

Costs and subsidies

The pricing structure varies per factory and project type, but follows consistent principles. Startups and SMEs typically pay 20-40% of what the same compute would cost on AWS or Google Cloud. For projects with strong societal impact - for example climate AI or healthcare innovations - full subsidies are possible via programs like Horizon Europe or the Digital Europe Programme.

Concrete pricing examples from existing factories: a mid-scale language model training that would cost €25,000 on AWS comes down to approximately €7,000 for a startup in a factory. A large-scale computer vision project that normally costs €100,000+ can be run for €30,000-40,000. This subsidy has direct impact on runway - a startup can do 2-3x more experiments with the same capital.

What factories enable: concrete transformations

The theoretical advantages are clear, but what actually changes when a company gains access to factory capacity? Examples from the first factories illustrate transformative impact.

Case: Precision agriculture in Spain

A cooperative of olive growers in southern Spain struggled with diseases that sometimes destroyed 30% of the harvest. Early detection was crucial but required expert inspection of every tree - practically infeasible with tens of thousands of hectares. The cooperative developed a computer vision system that could identify diseased trees in early stages via drone imagery.

The training process required analyzing millions of images to teach the model to recognize subtle visual indicators of different pathologies. On commercial cloud, this training budget would have cost €80,000+ - a prohibitive investment for a cooperative with tight margins. Via the Spanish factory, the model was developed in three months for €20,000. The ROI was back within one season: early intervention reduced crop loss by 60%.

The secondary effect is equally important: the model is now available to other agrifood cooperatives in the factory network. A Portuguese vineyard cooperative uses an adapted version for grape disease detection. This knowledge transfer between sectors and countries is where the federative model shows its value.

Case: Predictive maintenance in Italian manufacturing

An SME producer of industrial pumps with 200 employees wanted to implement predictive maintenance. Pumps run in critical industrial processes - an unexpected failure can shut down production lines with costs of tens of thousands of euros per hour. The company had decades of sensor data from pumps in operation, but no expertise to train machine learning models on it.

Via the Italian factory they gained access to both computing power and AI specialists who helped with feature engineering and model architecture. The resulting predictive maintenance system predicts failures 3-7 days in advance with 85% accuracy. This reduces downtime by 40% and transforms the business model: the company now sells "uptime-as-a-service" instead of just hardware. The competitive position versus Chinese low-cost manufacturers improved dramatically - not by producing cheaper, but by being smarter.

The pattern: from unaffordable to feasible

These cases share a fundamental pattern: tasks that were previously economically or technically infeasible become feasible. This is not because the technology is new - computer vision, predictive analytics and climate models have existed for years. The breakthrough is in accessibility. Capabilities previously only available to well-funded tech companies or governments are now within reach of a Spanish agri-cooperative or an Italian SME producer.

This is where the democratization narrative becomes concrete. Innovation is no longer limited to organizations with deep pockets for AWS spend. A good use case, domain expertise and some data are sufficient to begin. This level playing field has fundamental implications for where Europe's next AI innovations come from.

Realistic expectations: limits and challenges

With all the positive potential, nuance is warranted. AI Factories are not a silver bullet and the ecosystem has teething problems that need time to be resolved.

Capacity limits and waiting times

Even with 15+ factories, total compute capacity is limited compared to the virtually unlimited scale of hyperscalers. This creates allocation challenges. During peak periods - for example when multiple large-scale projects run simultaneously - waiting times can extend to multiple weeks. For startups with tight product launch deadlines, this can be problematic.

Factories apply prioritization algorithms that rank projects based on innovation potential, societal impact and urgency. A healthcare project with breakthrough potential gets priority over a commercial recommendation system. But this introduces subjectivity - who determines what "breakthrough potential" is? Transparency about these allocation decisions remains an issue the system must address.

Technical threshold remains high

Access to infrastructure doesn't eliminate the need for technical expertise. You can get 1000 GPUs, but if you don't know how to configure distributed training or optimize gradient descent, you'll get suboptimal results. For many SMEs, finding and affording AI engineers is a higher barrier than infrastructure costs.

EDIHs offer training and support, but this is necessarily generic. Deep domain-specific expertise - for example medical image analysis or supply chain optimization - must be brought by the company itself or hired. This skills gap is a structural bottleneck not solved by infrastructure alone.

Geographic inequality

Not every country gets a factory. Companies in smaller member states - think Cyprus, Malta, the Baltic states minus Lithuania - must apply for cross-border access. This introduces bureaucratic friction and possibly latency issues for real-time workloads. Remote access works fine for batch training but is problematic for inferencing workloads requiring millisecond latency.

The Commission tries to address this with "Antennas" - smaller satellite facilities connected to large factories - but coverage remains unequal. A Maltese startup has structurally less convenient access than a German competitor. Whether this geographic disadvantage is compensated by other factors (lower labor costs, niche specialization) will become visible in coming years.

Sustainability after subsidy period

Many factories run on temporary EU funding via the Digital Europe Programme and Horizon Europe, with budgets running until approximately 2030. What happens after that? Business models for self-sufficiency are still underdeveloped. Must factories charge commercial rates that price out startups? Will member states take over national co-financing, with risk of budget differences between rich and poor countries?

One possibility is developing hybrid models where commercial use subsidizes non-profit projects. Large corporates pay market-conform prices; those revenues finance subsidized access for startups and public sector. But this requires governance frameworks that don't yet exist and risks mission drift if commercial interests become dominant.

Strategic dimension: Europe's AI sovereignty

Zoom out to the geopolitical level, and AI Factories gain meaning extending far beyond infrastructure policy. They are a central pillar in Europe's strategy to maintain technological independence in a world where AI is becoming increasingly central.

Alternative to American cloud dominance

Currently, the majority of Europe's AI development runs on AWS, Google Cloud or Azure. This creates strategic vulnerability. The US Cloud Act gives American authorities jurisdiction over data stored on servers of American companies, regardless of where those servers physically stand. For European companies working with sensitive data - healthcare, defense, critical infrastructure - this is problematic.

AI Factories offer an alternative that falls entirely within EU jurisdiction. Data governance follows European rules, not American. There's no risk that a geopolitical conflict leads to access restrictions, as recently visible in tech export controls to China. For sectors where digital sovereignty is crucial, this is not a nice-to-have but a necessity.

Industrial strategy: play to your strengths

The sector focus of factories is no coincidence but reflects deliberate industrial strategy. Europe will most likely not produce the next Google or Meta - the consumer tech train has left and Silicon Valley dominates this domain. But Europe has comparative advantages in manufacturing, automotive, chemicals, healthcare and green tech. These are sectors where domain expertise and regulatory compliance create barriers to entry - areas where European companies can excel.

Norm-setting via Brussels Effect

By linking AI development to compliance with EU norms (AI Act, GDPR, sector regulations), Europe exports its values. Models trained in factories follow European principles around privacy, transparency and non-discrimination. If these models become internationally successful - for example a medical diagnostics AI used worldwide - EU norms become the de facto global standard.

This is the Brussels Effect in action: not by imposing regulation on foreign companies, but by building infrastructure that makes compliance easier than non-compliance. Companies wanting to sell globally then choose EU-compliant development because it opens the largest market with least friction.

Talent retention and brain drain

European AI researchers traditionally often leave for the US for access to resources. Factories create incentives to stay: you can do cutting-edge research on world-class infrastructure without needing to go to OpenAI or DeepMind. This is crucial because AI development is fundamentally dependent on talent - visible in how AI companies fight competitive battles over top researchers.

First signs are encouraging. Finnish AI researchers who previously left for Silicon Valley now stay for projects on the Finnish factory. French PhD students more often choose postdoc positions in Europe versus US when they have access to comparable resources. But it's early days - the test comes when companies like OpenAI aggressively start recruiting with multi-million compensation packages.

Practical roadmap for companies

For organizations wanting to move now and explore factory access, a pragmatic roadmap avoiding common mistakes.

Step 1: Use case fit assessment

Not every AI application requires factory capacity. A customer service chatbot or simple classification model can run fine on own servers. Factories make sense for compute-intensive workloads: training foundation models, large-scale computer vision, complex simulations, generative AI applications, reinforcement learning for robotics.

Concrete criteria: if your expected training time on own hardware takes months, or if your dataset is so large that local processing is impractical, then a factory is probably a good fit. If your model trains in a few hours or days, factories are overkill and only introduce administrative overhead.

Step 2: Build internal capabilities

No infrastructure compensates for lack of expertise. Before applying for factory access, ensure your team has at least one engineer with hands-on ML experience. This person doesn't need to be a world expert, but must be familiar with frameworks like PyTorch or TensorFlow, understand how model training works and be able to troubleshoot.

If you don't have this talent in-house, consider partnerships with universities or hiring consultants with specific factory experience. Some EDIHs offer subsidized access to AI consultants for SMEs - make use of this.

Step 3: Contact EDIH and start dialogue

Even if factory access seems far away, contact your local EDIH early. They can advise on alternative EU programs you might benefit from: AI TEFs for compliance testing, Digital Europe grants for AI projects, Horizon Europe subsidies for research partnerships. Building early contact pays off - there are more funding opportunities than awareness suggests.

Step 5: Explore cross-border options

If you can't wait for the Dutch factory, investigate access to existing facilities. The German, Belgian and French factories are accessible to Dutch companies. Cross-border applications are slightly more complex but certainly possible. If your project is sector-specific (for example agrifood), the Spanish factory might fit better than waiting for the Netherlands.

Some factories prioritize pan-European projects where companies from multiple countries collaborate. If you can form an international consortium - for example with partners in Germany and France - this significantly increases your approval chances.

The future: where is this heading?

Looking toward 2026 and beyond, what evolution can we expect in the factory ecosystem and Europe's broader AI infrastructure strategy?

Further scaling and specialization

The Commission has hinted at 30-40 factories by 2030, covering all EU member states plus associated countries like Norway and Switzerland. This creates a truly pan-European grid with regional specializations. Expect more vertical focus: factories dedicated to specific domains like healthcare AI, climate modeling, financial systems, automotive, where deep domain expertise is built.

Parallel to this come edge AI facilities for real-time applications. Autonomous vehicles, smart cities and industrial IoT require inferencing with millisecond latency - something centralized supercomputers cannot deliver. Distributed edge factories closer to end-users solve these latency problems, but require different architecture and governance models.

Integration with Common European Data Spaces

Europe is developing sector-specific data spaces for healthcare, mobility, energy, agriculture and finance - curated datasets accessible for innovation within governance frameworks. The synergy between data spaces and factories is obvious: combine high-quality training data with compute infrastructure to accelerate AI development.

A concrete example: the European Health Data Space makes anonymized patient data available for research. A medtech startup can use this data in combination with factory compute to train diagnostic AI - without having to collect decades of clinical data themselves. This level of data-infrastructure integration is where Europe's ecosystem approach can realize its full potential.

Commercialization paths and venture capital

Expect more formalized startup programs where promising factory projects get fast-tracked to venture funding. Factories can function as "proving grounds" for investors - proof of concept at enterprise scale before capital is committed. This reduces technology risk, a major concern for VCs investing in early-stage deep tech.

Some factories already experiment with demo days where portfolio companies present their results to investment consortia. If this model proves successful, we may see the emergence of factory-linked venture funds that specifically invest in companies incubated via the ecosystem. This could help address Europe's notorious funding gap for scale-ups.

Standardization and API interoperability

As the network grows, interoperability becomes crucial. Uniform APIs across factories make it easy to move workloads depending on availability. Shared development tooling reduces learning curves - an engineer working on the French factory can seamlessly switch to the Dutch facility.

This standardization also helps portability between factories and commercial cloud. Hybrid workflows become possible: rapid prototyping on cloud, heavy training runs on factories, production deployment back to cloud. This pragmatism - not everything must use EU infrastructure - increases attractiveness versus dogmatic approaches that exclude commercial platforms.

Conclusion: Europe's AI ambition becomes reality

The expansion of AI Factories in October 2025, with the Netherlands as newest addition, marks the moment when Europe's AI strategy shifts from policy to practice. Infrastructure takes shape, governance structures become operational, first success stories generate momentum.

For Dutch companies - especially SMEs and startups - an opportunity emerges that was unthinkable five years ago: access to world-class AI capacity without prohibitive investments. The democratization of AI technology that policymakers promise takes concrete form in megawatts and petaflops.

The question is no longer whether Europe can keep up in the global AI race. The infrastructure is here, funding is secured, regulatory frameworks take shape. The question is whether companies and organizations will actually use the resources now being built. Whether promising use cases in sector plans get converted into deployed systems. Whether the Netherlands' factory in two years is seen as underutilized capacity or as the catalyst that elevated the Dutch AI ecosystem to a higher level.

The factories are ready. The door is open. Who will step through first and show what's possible when world-class infrastructure becomes accessible to everyone with a good idea?

Sources and further reading:

• EuroHPC JU: AI Factories Initiative
• Apply AI Strategy (European Commission)
• EU AI Factories network overview

Want to explore whether AI Factories are suitable for your organization? Or need help navigating EU AI programs and subsidies? Contact us for a strategic consultation on how to access Europe's AI infrastructure.