← Back to Articles
Segler Consulting

Europe's AI Gambit: An In-Depth Analysis of the EuroHPC AI Factories and the Quest for Digital Sovereignty

June 22, 2025 • By symtr
Europe's AI Gambit: An In-Depth Analysis of the EuroHPC AI Factories and the Quest for Digital Sovereignty

Executive Summary

The European Union's AI Factories initiative, a cornerstone of its broader "AI Continent Action Plan," represents the bloc's most significant and capital-intensive effort to date to address its widening competitive gap in the global artificial intelligence race. This report provides an exhaustive analysis of the initiative, evaluating its strategic context, computational scale, operational model, and ultimate prospects for revitalizing Europe's lagging AI ecosystem.

Our analysis reveals that the AI Factories are not a ground-up creation but a strategic rebranding and repurposing of the existing EuroHPC supercomputing infrastructure. This pivot aims to make these powerful, publicly-funded assets directly accessible to the commercial AI sector, particularly startups and SMEs, in a state-led attempt to simulate the integrated innovation ecosystem provided by US hyperscalers. The initiative is driven by a stark reality: Europe possesses a mere fraction of the world's AI compute power and investment, creating a profound strategic dependency on foreign technology that is now viewed as a geopolitical risk.

In terms of raw compute, the EU's public infrastructure, while formidable in the scientific domain, is an order of magnitude smaller than that of individual US tech giants. The aggregated fleet of the primary EuroHPC systems amounts to approximately 57,000 high-end accelerators. This is dwarfed by Meta's plan to deploy an infrastructure equivalent to nearly 600,000 NVIDIA H100 GPUs by the end of 2024. The EU's multi-year public investment of roughly €10 billion is a fraction of the annual capital expenditures of companies like Meta ($60-65 billion) or Google ($75 billion). The EU's more ambitious "AI Gigafactory" plan, which aims to build centers with over 100,000 processors each, is a high-risk gamble entirely contingent on attracting tens of billions in private capital, a proposition that remains unproven.

Operationally, the AI Factories are designed as "one-stop shops" intended to be lean and efficient. However, their federated, consortium-based structure and project-based, peer-reviewed access model are fundamentally misaligned with the agile, on-demand needs of commercial startups. While providing "free" compute lowers the barrier to entry, it introduces administrative friction and may fail to prepare companies for the commercial realities of scaling, where compute is a major operational cost.

Strategically, the initiative exists within a contradictory policy dynamic. The AI Factories function as a state subsidy designed to counteract the significant compliance costs and innovation friction imposed by the EU's own AI Act. The "trustworthy AI" branding is an attempt to turn this regulatory burden into a competitive advantage, a market thesis that has yet to be validated. The initiative correctly identifies the compute deficit as a critical problem but addresses the symptom more than the root causes of Europe's AI lag: a less dynamic venture capital market, a fragmented commercial ecosystem, and persistent challenges in talent retention.

In conclusion, the AI Factories are a necessary and ambitious defensive maneuver to prevent Europe from falling further behind in the AI race. They provide a vital lifeline of compute resources to a nascent ecosystem. However, they are insufficient on their own to close the competitive chasm with the United States and China. Achieving true digital sovereignty will require not only state-of-the-art public infrastructure but also deeper structural reforms to foster a pan-European private market for capital, talent, and data that can scale homegrown innovators into global champions. The success of the AI Factories will not be measured by the machines they build, but by the commercial viability of the companies they incubate.

Section 1: The AI Continent Action Plan: Ambition in the Shadow of the Compute Gap

The European Union's AI Factories initiative did not emerge in a vacuum. It is the latest and most tangible manifestation of a long-standing, multi-year policy effort aimed at rectifying a deeply felt strategic vulnerability in the global technology landscape. Framed as a cornerstone of the ambitious "AI Continent Action Plan," the initiative represents a direct and urgent response to a profound and widely acknowledged deficit in artificial intelligence infrastructure, investment, and innovation. This section establishes the strategic context for the AI Factories, tracing their policy origins and deconstructing their foundational blueprint to reveal an effort born from geopolitical necessity as much as economic aspiration.

1.1. A Strategy Born from Necessity

The intellectual groundwork for the AI Factories was laid over several years, reflecting a growing sense of urgency within Brussels. The EU's journey began with its first comprehensive AI strategy in April 2018, which launched the Coordinated Plan on AI.1 This was followed by the February 2020 White Paper on AI, which first articulated the EU's signature dual-pronged vision: creating an "ecosystem of excellence" to foster innovation alongside an "ecosystem of trust" to be established through regulation, which would later become the EU AI Act.1 The 2021 Digital Compass further solidified this vision with concrete targets for the decade.1

Despite these plans, the "excellence" pillar remained underdeveloped, a fact thrown into sharp relief by the explosion of generative AI in the 2020s. This new paradigm made access to massive computational power the primary determinant of competitiveness, exposing Europe's critical weakness. The AI Factories initiative, officially launched through the AI Innovation Package in January 2024 and made a centerpiece of the "AI Continent Action Plan" in April 2025, is the EU's most direct attempt to build this missing pillar of excellence.1

The initiative's urgency is rooted in stark, unflattering data. Europe accounts for a mere 4% of the world's AI computing power, a figure dwarfed by the United States' 70% share.4 This compute gap is a direct result of a massive investment disparity; Europe faces an estimated annual investment shortfall of €270 billion compared to the US.5 This financial deficit has tangible consequences for its innovation ecosystem, with European AI startups attracting only 6% of global funding, while their American counterparts capture 61%.6 This has led to what EU officials and analysts describe as a "risk of strategic dependence" on non-EU technologies, transforming the quest for "digital sovereignty" from an economic goal into a matter of geopolitical security.4 The EU's own policy documents explicitly frame the initiative as a response to these vulnerabilities, citing the use of US export controls on semiconductors as a real-world example of how technology can be wielded as a geopolitical lever.4

This context reveals a critical policy pivot. For years, the EU's primary influence in the global technology sphere has been regulatory—the so-called "Brussels effect," where its stringent rules on data privacy and competition become de facto global standards.7 However, this regulation-first approach has drawn escalating criticism for stifling the very innovation it sought to govern, potentially cementing Europe's technological lag behind the more laissez-faire US and state-supported Chinese ecosystems.7 The AI Factories initiative, with its focus on capital-intensive hardware and infrastructure, is an implicit admission that regulatory power is insufficient without sovereign technological capability. It signals a strategic shift where the EU is attempting to build the "hard power" of infrastructure to complement the "soft power" of its rule-making.

Furthermore, the consistent framing of the initiative around "geopolitical risk" and "strategic autonomy" underscores its dual purpose as both an economic stimulus and a project with national security implications.1 The fear is not just about being outcompeted economically, but about being strategically marginalized in a world where control over critical technologies like advanced semiconductors and cloud computing is a key instrument of state power. The AI Factories are thus conceived as a form of critical infrastructure, akin to energy grids or transport networks, aimed at ensuring the EU can innovate and operate without being subject to the strategic decisions of other global powers.

1.2. Deconstructing the "AI Factory" Blueprint

At its core, an AI Factory is officially defined as a "one-stop shop" and a "dynamic ecosystem" built around the EU's world-class supercomputers, managed by the European High-Performance Computing Joint Undertaking (EuroHPC JU).2 Its primary mission is to develop "trustworthy" cutting-edge generative AI models, with a specific focus on empowering European AI startups, small and medium-sized enterprises (SMEs), and researchers.2

The model is not to build an entirely new infrastructure from the ground up. Instead, it leverages and strategically re-tasks the existing EuroHPC assets. This is being implemented via three distinct tracks:

  1. Track 1: Appointing Existing Systems: Utilizing current EuroHPC supercomputers as AI Factories with minimal changes.13
  2. Track 2: Upgrading Systems: Enhancing existing supercomputers with AI-optimized hardware, such as new GPUs or faster networking, to make them suitable for AI Factory activities.13
  3. Track 3: Building New Systems: Procuring and deploying brand-new, AI-optimized supercomputers specifically designed for the needs of generative AI.13

A crucial element of the blueprint is the plan to integrate these factories into a wider web of EU support programs. They are designed to be networked with European Digital Innovation Hubs (EDIHs), which help companies with digital transformation; Testing and Experimentation Facilities (TEFs), which provide sandboxes for validating AI solutions; and the Common European Data Spaces (CEDS), which aim to facilitate data access for training models.12 This structure is intended to create a comprehensive, full-stack support platform for AI innovation.

This blueprint reveals that the "AI Factory" is fundamentally a rebranding of the pre-existing EuroHPC JU, which was established in 2018 with a fleet of supercomputers that were already in place or planned.15 The core infrastructure, including powerful systems like LUMI in Finland and Leonardo in Italy, predates the AI Factory concept. These machines were traditionally the domain of academic and scientific research, used for complex simulations in fields like climate science and medicine.16 The rise of generative AI created a sudden, massive demand for GPU-based computing from a new and more commercially-oriented user base. The EU recognized that its expensive HPC assets could be repurposed to meet this demand. The "Factory" label is a deliberate policy and marketing choice, designed to make these public assets sound more industrial, accessible, and appealing to the private sector than the more esoteric "High-Performance Computing Center." This rebranding, combined with the creation of new, industry-focused access tracks, represents a clear pivot from serving pure science to actively trying to incubate a commercial AI industry.18

In doing so, the EU is attempting to construct a state-directed, full-stack innovation platform to substitute for the lack of a dominant, private-sector European hyperscaler. In the United States, companies like Amazon Web Services, Google Cloud, and Microsoft Azure provide a vertically integrated ecosystem of compute, data storage, software tools, and expertise as a seamless commercial service. Europe lacks a domestic champion of this scale. The AI Factory initiative, with its plan to connect supercomputers with Data Labs, EDIHs, and TEFs, is an attempt to piece together these essential components using a network of public and public-private institutions.12 This makes the initiative a form of state-led industrial policy aimed at simulating the functions of a market-driven hyperscaler ecosystem, a grand experiment in building a competitive advantage through public coordination rather than private competition.

1.3. Funding the Future: Public Billions and Private Hopes

The financial architecture of the AI Factories initiative is complex, relying on a blend of existing EU budgets, member state contributions, and significant hopes for private investment. The overarching EuroHPC JU, which provides the foundational infrastructure, operates with a budget of approximately €7 billion for the 2021-2027 period.15 The acquisition of new supercomputers is typically managed through a co-funding model, where the EU contributes up to 50% of the cost, and the hosting member state or consortium provides the remainder.12

The ambition scales dramatically with the "AI Gigafactories" plan. This is tied to a separate vehicle, the InvestAI facility, which aims to mobilize €20 billion for infrastructure and an even more staggering €200 billion for the broader AI ecosystem.2 However, these headline figures require careful scrutiny. The €200 billion is a mobilization target, not a direct government expenditure; it assumes that €150 billion will come from private investors, catalyzed by an EU contribution of €50 billion.6 The €20 billion for Gigafactories is similarly a target for a public-private fund, not a fully appropriated budget.2 This reliance on leveraging existing funds and attracting private capital has drawn criticism that the initiative's financial firepower is less "new" than the ambitious announcements suggest.9

This funding structure creates a significant gap between the stated ambitions and the committed new public funds. The success of the most transformative part of the plan—the Gigafactories—is not guaranteed by public will alone. It is entirely contingent on whether the EU's de-risking mechanisms, such as the European Investment Bank providing "first-loss" tranches of capital, are attractive enough to persuade private capital to invest tens of billions of euros in these massive, unproven, state-adjacent projects.21

Furthermore, the co-funding model inherently favors larger, wealthier member states, which could potentially exacerbate rather than solve the EU's internal innovation disparities. Because hosting entities are required to match the EU's substantial financial contributions for top-tier systems, only nations with significant domestic science and technology budgets can afford to host the most powerful exascale machines. This reality is reflected in the current distribution of flagship systems: JUPITER is in Germany, Leonardo is in Italy, and MareNostrum 5 is in Spain.14 This creates a hub-and-spoke model where smaller or less wealthy nations are relegated to hosting less powerful petascale systems or establishing "Antennas"—national gateways that provide access to the larger factories on the continent.23 While this structure ensures broad participation, it also risks concentrating the most critical resources in a few key countries, potentially reinforcing the very economic and technological divides the Union aims to bridge.

Section 2: Gauging the Compute: A Comparative Analysis of Global AI Infrastructure

To assess the true potential of the EU's AI Factories, it is essential to move beyond policy ambitions and conduct a granular, quantitative analysis of their computational power. In the modern AI era, raw compute—measured in the number and quality of accelerators like Graphics Processing Units (GPUs)—is the fundamental currency of innovation. This section provides a technical deep-dive into the EuroHPC fleet, juxtaposing its capabilities against the colossal infrastructures of global technology giants. The analysis reveals that while the EU's public assets are formidable in the context of academic supercomputing, they are dwarfed by the scale of individual US private-sector players, highlighting the immense challenge Europe faces in its quest for digital sovereignty.

2.1. The EuroHPC Fleet: A Technical Deep-Dive

The AI Factory network is built upon the supercomputing assets of the EuroHPC Joint Undertaking. This fleet comprises a mix of world-class "pre-exascale" systems, a new exascale machine, and several powerful "petascale" machines distributed across the continent. The flagship systems forming the backbone of the AI Factory initiative include:

  • JUPITER (Jülich, Germany): As Europe's first exascale system, JUPITER represents the pinnacle of the EuroHPC fleet. Its primary power comes from a "Booster Module" equipped with approximately 24,000 NVIDIA GH200 Grace Hopper Superchips. This configuration is designed to achieve over 1 ExaFLOP/s in traditional double-precision (FP64) scientific computing. Crucially for AI, it is projected to deliver up to 90 ExaFLOPS of performance in lower-precision formats (e.g., FP8) used for training large models, making it one of the most powerful AI machines in the world upon its full deployment.25
  • LUMI (Kajaani, Finland): A leading pre-exascale system, LUMI is currently one of the most powerful supercomputers in Europe. It achieves a sustained performance (Rmax) of 379 PFLOPS, primarily driven by its GPU partition, which contains 10,240 AMD Instinct MI250X GPUs.29 LUMI is also noted for its energy efficiency, running on 100% hydroelectric power.29
  • Leonardo (Bologna, Italy): Another top-tier pre-exascale machine, Leonardo delivers a sustained performance of 238.7 PFLOPS. Its AI and HPC capabilities are centered on a "Booster Module" containing 3,456 compute nodes, each equipped with four NVIDIA A100 GPUs, for a total of 13,824 A100 accelerators.31
  • MareNostrum 5 (Barcelona, Spain): This system features a heterogeneous architecture with a powerful "Accelerated Partition" (ACC) designed for AI workloads. The ACC consists of 1,120 nodes, each with four NVIDIA H100 GPUs, totaling 4,480 H100s. This partition alone provides a peak performance of 260 PFLOPS.31
  • Other Key Systems: Beyond the pre-exascale flagships, the network includes other significant assets. France's Jean Zay supercomputer, central to the AI2F factory, has been progressively upgraded and now includes a mix of approximately 3,500 NVIDIA GPUs, including H100s, A100s, and older V100s.38 Smaller but still powerful petascale systems contribute further capacity, including Meluxina in Luxembourg (800 NVIDIA A100 GPUs)40, Vega in Slovenia (240 NVIDIA A100 GPUs)42, and Discoverer in Bulgaria, which is being upgraded with a new partition of 128 NVIDIA H200 GPUs.44

Aggregating the accelerator counts from these primary systems provides a rough estimate of the EU's public AI compute capacity. The total comes to approximately 57,000 high-end accelerators (GPUs and their equivalents). This is a significant pool of resources, placing Europe's public research infrastructure in the top tier globally.

However, a critical challenge lies beneath the surface of these impressive numbers. The heterogeneity of the EuroHPC fleet introduces a significant hidden cost in the form of software complexity and optimization overhead, which undermines the vision of a seamless, unified ecosystem. The fleet is a diverse mix of accelerator architectures (NVIDIA's CUDA-based GPUs and AMD's ROCm-based GPUs) and multiple generations of hardware (NVIDIA's V100, A100, H100, and GH200; AMD's MI250X). An AI developer building a model on LUMI must use the AMD ROCm software stack.30 To run or scale that same model on Leonardo or MareNostrum 5, they would need to port their code to NVIDIA's CUDA platform, a non-trivial engineering task. Even within the NVIDIA ecosystem, optimizing code for the different tensor core capabilities, memory hierarchies, and interconnects of the A100, H100, and GH200 architectures requires specialized work. This fragmentation creates friction, increases development time, and demands a broader and deeper skillset from user teams than would be required for a single, homogenous cloud platform. Consequently, the "network of factories" is less a single, unified entity and more a federation of distinct, powerful systems, complicating the practical realization of the "one-stop-shop" goal.

Table 1: The EuroHPC AI Factory Fleet - Key Systems and Specifications

AI Factory System Location Key Accelerator(s) Total Accelerator Count Peak AI Performance (Est. FP8/Lower Precision)
JUPITER Jülich, Germany NVIDIA GH200 ~24,000 ~90 ExaFLOPS
LUMI Kajaani, Finland AMD Instinct MI250X 10,240 ~1.3 ExaFLOPS
Leonardo Bologna, Italy NVIDIA A100 13,824 ~10 ExaFLOPS (FP16)
MareNostrum 5 Barcelona, Spain NVIDIA H100 4,480 ~1 ExaFLOPS
Jean Zay Saclay, France NVIDIA H100/A100/V100 ~3,500 Not specified
Discoverer Sofia, Bulgaria NVIDIA H200 128 ~0.128 ExaFLOPS
Meluxina Bissen, Luxembourg NVIDIA A100 800 ~0.5 PetaFLOPS (AI)
Vega Maribor, Slovenia NVIDIA A100 240 Not specified
Total (Aggregated) EU-wide Mixed ~57,000 ~102 ExaFLOPS

Note: Performance figures are based on official announcements and technical specifications. "Peak AI Performance" is often theoretical and depends on workload and precision (e.g., FP8, FP4). The aggregated total is an estimate to illustrate scale. Sources:25

2.2. The Global Hyperscalers: A League of Their Own

While the EuroHPC fleet is impressive, it operates in a different universe of scale compared to the private infrastructure being built by leading US technology companies. These firms are not just building supercomputers; they are building planet-scale AI engines, with capital expenditure plans that eclipse the budgets of many national governments.

  • Meta: Mark Zuckerberg has announced one of the most aggressive AI infrastructure build-outs globally. By the end of 2024, the company aims to have an infrastructure portfolio that includes 350,000 NVIDIA H100 GPUs, with a total compute power equivalent to nearly 600,000 H100s when accounting for other accelerators.47 The company's capital expenditure for 2025 is projected to be between $60 billion and $65 billion.48 This includes plans for a future 2-gigawatt data center designed to house a staggering 1.3 million GPUs.48
  • Google: The company plans a capital expenditure of $75 billion in 2025 to support its AI ambitions.50 While Google does not disclose its total number of accelerators, its strategy is defined by a dual approach. It offers the full range of NVIDIA's latest GPUs (including H100, H200, and B200 systems) on Google Cloud, making it a major NVIDIA partner.51 Simultaneously, it develops and deploys its own custom-designed Tensor Processing Units (TPUs) at massive scale for internal use and cloud customers. A single "pod" of its latest TPUs can scale to thousands of chips, providing tens of ExaFLOPS of performance.53
  • Tesla: To power the development of its Full Self-Driving (FSD) and Optimus robotics programs, Tesla pursues a dual-path strategy. It operates a large cluster of approximately 50,000 NVIDIA H100 GPUs for AI training.55 In parallel, it is building out its own custom Dojo supercomputer. While the first version of Dojo is relatively modest, with performance equivalent to about 8,000 H100s, it represents a strategic investment in bespoke hardware.55
  • OpenAI: Although it does not own its infrastructure outright, OpenAI drives some of the largest compute builds in the world through its partnerships. Its demand is a key factor behind Microsoft's massive Azure investments, with OpenAI projected to spend $13 billion on Azure compute alone in 2025.57 More dramatically, OpenAI is a lead partner in the Stargate project, a planned $500 billion initiative with SoftBank and Oracle to build a network of dedicated AI data centers. One proposed Stargate site alone is slated to house 64,000 next-generation NVIDIA Blackwell GPUs.57

This comparison reveals an undeniable and vast compute chasm. The EU's entire public AI compute capacity of roughly 57,000 accelerators is an order of magnitude smaller than the infrastructure of a single US tech giant. Meta's target of 350,000 H100s alone is more than six times larger than the aggregated EuroHPC fleet.47 The investment gap is even more stark: the EU's total multi-year public investment in EuroHPC of approximately €10 billion is a small fraction of the annual capital expenditure of Meta ($60-65 billion) or Google ($75 billion).12 This disparity in both deployed hardware and the capital available for future procurement means the EU is not merely lagging; it is operating on a completely different plane of scale.

Beyond sheer numbers, the true competitive moat for these US tech giants is their vertically integrated, custom silicon strategy—a capability the EU is years, if not a decade, away from replicating. Companies like Meta with its MTIA chip, Google with its TPUs, and Tesla with its Dojo system are not just buying accelerators; they are designing their own bespoke hardware to be perfectly optimized for their specific software stacks and AI workloads.53 This deep co-optimization of hardware and software yields performance and efficiency advantages that cannot be achieved by simply procuring off-the-shelf components. The EU's AI Factories, by contrast, are primarily procurers of foreign technology, mainly from NVIDIA. While European chip initiatives like SiPearl exist and are part of the long-term vision for systems like JUPITER and the French exascale machine, they are not yet at a competitive scale for AI workloads and are not the primary drivers of the current AI Factory systems.62 This leaves the EU dependent on a non-European supply chain and unable to access the profound strategic advantages of a fully integrated, custom-designed technology stack.

Table 2: Comparative AI Compute Power (2025/2026 Outlook)

Entity Estimated GPU Count (H100-equivalent) Key Chip Types Stated 2025 Capex Custom Silicon Strategy
EuroHPC (Aggregated) ~57,000 NVIDIA (A100, H100, GH200), AMD (MI250X) ~€1.5B (Annualized portion of €7B budget) Nascent (e.g., SiPearl for HPC, not yet scaled for AI)
Meta ~600,000 (by end of 2024) NVIDIA H100, Custom MTIA $60-65 Billion Yes (MTIA for inference/training)
Google Not Public (millions est.) Custom TPU (v5e, Trillium, Ironwood), NVIDIA (H100, H200, B200) $75 Billion Yes (TPU is core to strategy)
Tesla ~58,000+ NVIDIA H100, Custom Dojo D1 ~$10 Billion (AI spend) Yes (Dojo for FSD training)
OpenAI (via partners) 100,000s (cluster-dependent) NVIDIA H100/Blackwell, Microsoft Azure Drives partner capex (e.g., $13B Azure spend) Yes (In-house chip development underway)

Note: GPU counts are estimates based on public statements and reports. Capex figures are for the full fiscal year 2025. H100-equivalent is a metric to standardize compute power. Sources:12

2.3. The "Gigafactory" Aspiration: A Credible Leap or a Distant Dream?

To address this staggering compute gap, the EU's strategy includes a more ambitious, longer-term vision: the creation of "AI Gigafactories." These are conceived as massive, large-scale facilities, each equipped with over 100,000 advanced AI processors, with a plan to establish up to five such centers across the Union.2 The goal is to create infrastructure capable of training the next generation of trillion-parameter foundation models, directly competing with the largest private clusters in the world.

This vision is to be funded by the €20 billion InvestAI facility, which aims to catalyze the necessary public-private partnerships to build these centers.2 The proposed timeline is aggressive, with the first call for proposals for Gigafactories expected in late 2025 and the first facilities becoming operational in the 2027-2028 timeframe.22 A consortium of German companies, including Deutsche Telekom and SAP, has already expressed interest in bidding to host one such Gigafactory in Germany.65

However, the Gigafactory plan represents a high-risk, high-reward gamble on a specific thesis about the future of AI—namely, that "scale is all you need" and that training frontier models will continue to require ever-larger, monolithic compute clusters. This paradigm is already facing challenges from emerging research and companies like DeepSeek, which have demonstrated that innovation in model architecture and training efficiency can achieve state-of-the-art performance with significantly less computational power than previously thought.9 If this trend toward smaller, more efficient models continues, the EU risks investing billions in the AI equivalent of battleships just as more agile aircraft carriers become the dominant strategic asset.

More fundamentally, the Gigafactory plan is, for now, a financial projection rather than a funded reality. Unlike OpenAI's Stargate project, which was announced with named capital partners like SoftBank and Oracle58, the EU's plan relies on the theoretical appeal of its public de-risking model to attract the tens of billions in private investment needed to make these centers a reality. It is a far more uncertain proposition, dependent on convincing a risk-averse European private sector to co-invest in colossal, state-led infrastructure projects whose commercial viability is not yet proven. The Gigafactory remains a distant dream, a bold aspiration whose realization hinges on overcoming significant financial and strategic hurdles.

Section 3: The Ecosystem Engine: Lean Enabler or Bureaucratic Bottleneck?

Beyond the sheer scale of compute, the success of the AI Factories will be determined by their operational effectiveness. The EU has positioned them as agile, user-centric hubs designed to serve the fast-paced world of startups and SMEs. The central question is whether these publicly funded, consortium-run entities can deliver on the promise of being lean enablers, or if they will inevitably be constrained by the bureaucratic processes inherent in their structure. This section evaluates the operational model of the factories, from their "one-stop-shop" promise to their access mechanisms, to assess their suitability for the target user base.

3.1. The "One-Stop-Shop" Promise

The official vision for the AI Factories is ambitious and user-focused. They are explicitly designed to function as a "one-stop shop" for AI innovators.11 This model aims to provide a single, accessible point of contact for a comprehensive portfolio of services that extends far beyond raw compute. The intended offering includes expert guidance on technical challenges, specialized training programs, curated datasets, and access to optimized software and foundational models.11

To make this vision tangible, some factories plan to establish physical hubs that serve as community centers. For instance, the AI Factory Austria (AI:AT) plans to set up a hub with co-working spaces and a dedicated support staff of around 60 employees to assist users.67 Similarly, the LUMI AI Factory in Finland aims to create "dynamic co-working spaces tailored to diverse needs and goals" to foster a collaborative environment.32 The overarching goal, as articulated by the LUMI consortium, is to build a "thriving AI community that transcends borders and sectors".14

This "one-stop-shop" concept is a laudable and necessary attempt to replicate the integrated, seamless user experience that has been a key to the success of commercial cloud providers like AWS and Google Cloud. However, the federated, consortium-based structure of the AI Factories presents a significant challenge to achieving this ideal. A commercial cloud provider offers a unified, globally consistent interface; a user interacts with "AWS," not a regional franchise. In contrast, the EU's AI Factories are a network of distinct entities, each managed by different national institutions and consortia. For example, AI:AT in Austria is led by Advanced Computing Austria and the AIT Austrian Institute of Technology67; the JAIF factory in Germany is centered at the Jülich Supercomputing Centre11; and the LUMI AI Factory is managed by a consortium of Nordic countries led by Finland's CSC.14

This distributed model, while effective at securing broad member-state participation and funding, inherently introduces layers of administrative friction and potential inconsistency. A startup seeking to access resources on LUMI will have to navigate the procedures of the CSC-led consortium. If they later wish to use the unique capabilities of Leonardo in Italy, they will have to engage with a different set of processes managed by CINECA. While the EuroHPC JU provides a central portal for applications, the on-the-ground support, specific service offerings, and operational cultures will inevitably vary from factory to factory. This federated structure, therefore, creates more administrative complexity than a single, centrally managed commercial platform, posing a direct challenge to the "lean and efficient" ideal that is so critical for fast-moving startups.

3.2. Accessing the Machines: A Pathway for Startups or a Gauntlet of Red Tape?

Recognizing that traditional academic grant processes are ill-suited for commercial entities, the EuroHPC JU has designed specific access pathways tailored for the AI industry. Access to the AI Factories is managed through formal calls for proposals, but with dedicated tracks for startups and SMEs that are, crucially, offered free of charge.18 These "Industrial Innovation" tracks are tiered to accommodate projects of varying scale and maturity18:

  • Playground Access: Provides limited resources for entry-level users and initial experimentation.
  • Fast Lane Access: A lightweight process for users already familiar with HPC, offering up to 50,000 GPU hours for more substantial projects.
  • Large Scale Access: Caters to the development of large AI models and applications requiring more than 50,000 GPU hours.

All applications, regardless of the track, are subject to a peer-review process conducted by an independent Access Resource Committee (ARC). This is intended to ensure fairness, transparency, and the allocation of resources to projects with the highest potential for excellence and impact.71 The system aims for expediency, with a stated goal of providing access to resources within one month of a call's cut-off date.72

Table 3: EuroHPC AI Factory Access Tiers for SMEs and Startups

Access Tier Target User Resource Allocation Limits Cost Structure Application Process
Playground Entry-level users, startups exploring initial concepts Limited resources for experimentation Free of charge Lightweight application
Fast Lane Users familiar with HPC, SMEs developing prototypes Up to 50,000 GPU hours Free of charge Lightweight evaluation process
Large Scale Projects developing large AI models or applications More than 50,000 GPU hours Free of charge Full peer-review evaluation
Commercial Access Other industrial applications not for R&I Varies Pay-per-use Not specified

Note: This table summarizes the "AI for Industrial Innovation" track. A separate track exists for scientific and collaborative projects. Sources:18

Despite these well-intentioned efforts to streamline access, the project-based, peer-reviewed application model remains fundamentally misaligned with the agile, iterative, and often unpredictable nature of commercial startup development. Startups do not operate in fixed, pre-defined project cycles of one year.73 Their value lies in their ability to experiment rapidly, pivot strategy based on market feedback, and scale resources up or down on demand. The need to formulate a formal proposal, submit it by a specific cut-off date, and await a peer-review decision introduces a level of friction and delay that is entirely absent in the on-demand, credit-card-driven model of commercial cloud platforms. For a startup, the "cost" of using the AI Factories is not paid in euros but in agility and speed—two of its most critical competitive assets.

Furthermore, the "free compute" offering, while a powerful incentive to lower the barrier to entry, may create perverse incentives and fail to prepare startups for the realities of commercial scaling. This subsidy-based model risks fostering a class of "grant-preneurs" who become adept at navigating the EU funding landscape but not at building commercially sustainable products where compute is a major and unavoidable cost of goods sold (COGS). The model socializes the cost of research and development but does not adequately prepare companies for the privatized cost of production and scale. When a promising startup eventually outgrows the resource limits of the AI Factory program, it will be forced to transition to the commercial market, where it will face the full, unsubsidized price of compute. This can create a fatal financial shock for a business model that was not designed with this major operational expense in mind. The small provision for "pay-per-use" commercial access indicates an awareness of this cliff-edge, but the overwhelming focus of the initiative remains on subsidy, not on building a self-sustaining market.18

3.3. A Network of Hubs or a Collection of Silos?

The EU's vision is to create a truly pan-European network, with at least 15 factories and several "Antennas" expected to be operational and interconnected by 2026.2 There are documented plans for cross-border collaboration. For instance, the JAIF factory in Germany intends to establish formal links with partners in Italy, Spain, and other countries.11 A particularly ambitious plan involves a Franco-German collaboration to create a "virtual supercomputer" of 50,000 GPUs by federating their respective exascale systems, JUPITER and Alice Recoque (formerly Jules Verne).25 The "Antenna" model further extends this network, allowing countries without a full-fledged factory, like the post-Brexit UK, to establish a national gateway that partners with and provides access to a continental factory.23

However, the distributed, nation-centric ownership and funding model risks reinforcing the very market fragmentation the EU officially aims to overcome. Each factory is a point of national pride and investment, closely tied to the industrial strategies and priorities of its host country. The LUMI AI Factory, for example, is described as "very important to Finland as a whole"32, while Austria's AI:AT factory is explicitly focused on integrating AI into the nation's critical manufacturing sector.11

While pan-European collaboration is a stated goal, the underlying structure incentivizes competition among member states for high-profile projects, top talent, and international prestige. Instead of creating a single, seamless "European AI Cloud," the initiative risks establishing an "E-UN of Supercomputing"—a federation of powerful but distinct national assets, each with its own priorities and operational quirks. This could perpetuate silos of expertise and make it difficult to pool resources effectively for truly massive, continent-spanning projects that could rival the scale of US hyperscalers. The challenge for the EuroHPC JU will be to impose a layer of governance and technical interoperability strong enough to overcome these powerful national interests and forge a whole that is greater than the sum of its parts.

Section 4: The Strategic Crucible: Can the AI Factories Revitalize Europe's Ecosystem?

The AI Factories initiative represents a multi-billion-euro bet on Europe's technological future. Its ultimate success, however, will not be measured by the number of GPUs deployed or the PetaFLOPS achieved. It will be judged on its ability to catalyze a vibrant, self-sustaining, and globally competitive AI ecosystem. This final analytical section synthesizes the findings on strategy, compute, and operations to assess the initiative's prospects. It weighs the potential benefits against the significant structural, regulatory, and competitive challenges that define the European AI landscape, culminating in a nuanced verdict on whether this ambitious gambit can truly revitalize the continent's fortunes.

4.1. The AI Act Paradox: Fostering Trust or Stifling Innovation?

A defining feature of the AI Factories is their dual mandate: they must not only drive innovation but also champion the development of "trustworthy AI" that is fully compliant with the EU AI Act.2 The European Commission is actively building a support framework to help users navigate this complex regulatory landscape, including the creation of an "AI Act Service Desk" to provide guidance to businesses.3

This creates a fundamental paradox at the heart of the EU's strategy. The AI Act is widely regarded as the world's most stringent and comprehensive AI regulatory regime. While lauded for its focus on ethics and fundamental rights, it is also a source of significant concern for the technology industry. A recent report commissioned by AWS found that 68% of European businesses find the Act difficult to interpret, with compliance costs absorbing up to 40% of IT budgets. The report warns that this regulatory burden could lead businesses to reduce their AI investments by nearly 30%.18 Critics, particularly from outside the EU, have described its requirements for data quality, traceability, and bias elimination as "burdensome" and "technologically unrealistic" for many companies, especially startups.10 The Act's broad and sometimes vague definitions of risk are feared to create legal uncertainty that could stifle innovation and lead to "AI flight," where innovative companies choose to develop their products outside the EU to avoid the regulatory overhead.10 Even major US tech firms like Google and Meta have publicly criticized draft implementation codes for the Act as "unworkable" and a "step in the wrong direction".78

Viewed through this lens, the AI Factories initiative can be interpreted as a massive state subsidy designed to counteract the negative economic externalities of the EU's own signature legislation. The EU is simultaneously imposing a high regulatory "tax" on AI development through the AI Act while offering a direct subsidy in the form of free, state-of-the-art compute resources via the AI Factories. This creates a contradictory policy dynamic. The free compute effectively masks the true cost of operating under the AI Act, creating an artificial, sheltered environment for startups. This risks creating a "dependency trap": a company can develop a promising model using the subsidized infrastructure, but its business model may only be viable within that sheltered ecosystem. Once the startup outgrows the public program and must compete on the open market, it will face the full, unsubsidized burden of the AI Act's compliance costs, a potentially fatal shock to its commercial viability.

The EU's strategic response to this paradox is to attempt to turn a regulatory liability into a global competitive advantage. The core narrative is that by building AI in a compliant, "ethical," and "trustworthy" manner, European firms will offer a superior product that commands market preference and trust.2 The AI Factories are positioned as the crucibles where this new class of "Made in Europe" AI will be forged. This is a high-stakes bet on ethics as a key purchasing criterion in the global marketplace. The commercial success or failure of the startups that emerge from the AI Factory ecosystem will serve as the ultimate test of this thesis. If they can successfully compete on the global stage, the EU's dual strategy of regulation and subsidization will be validated. If, however, they are outcompeted on performance, cost, and speed by rivals from less-regulated jurisdictions, the AI Act will be proven to have been a net economic drag, one that even a multi-billion-euro infrastructure program could not overcome.

4.2. Beyond Compute: Addressing the Talent and Data Deficits

The EU's AI Continent Action Plan correctly recognizes that a successful ecosystem requires more than just hardware. The plan explicitly includes pillars for strengthening AI skills and increasing access to high-quality data.1 The AI Factories are designed to be central to this effort. They are envisioned as hubs that bring together not just computing power but also "data and talent".2 Most factory proposals include plans for extensive training programs, workshops, and expert consultation to upskill the European workforce.32 To address the data challenge, the plan calls for the establishment of "Data Labs" within the factories. These labs will be tasked with curating large, high-quality datasets and linking them to the broader Common European Data Spaces initiative, aiming to create a true internal market for data.3

Despite these well-laid plans, the initiative faces formidable headwinds. Europe continues to suffer from a significant "talent drain," with many of its best and brightest researchers and engineers being lured to the higher salaries and greater commercial opportunities in the United States.80 Furthermore, the corporate adoption of AI within the EU remains troublingly low, at just 13.5%, indicating a wide gap between the technology's potential and its practical implementation in the economy.3

This highlights the core challenge for the AI Factories: they address the most straightforward problem—access to hardware—but their success is ultimately hostage to the far more intractable issues of talent retention and data fragmentation. Building and procuring supercomputers is primarily a capital expenditure problem, one that public funding can solve. However, creating a vibrant, self-sustaining talent pool that chooses to stay and build in Europe is a deep-seated structural and cultural challenge that cannot be solved by infrastructure alone. Similarly, creating a unified, high-quality data market requires overcoming a complex patchwork of national regulations, privacy interpretations, and language barriers that have long hindered the creation of a true digital single market.

The training programs and Data Labs within the factories are positive and necessary interventions, but they are small-scale solutions to massive, continent-wide problems. There is a significant risk that the EU will succeed in building powerful machines that then remain underutilized, not for lack of ambition, but for a shortage of skilled people to operate them at their full potential and a lack of the large-scale, high-quality European datasets needed to train uniquely European AI models.

4.3. A Viable Path to Digital Sovereignty?

The ultimate objective of the AI Factories, and the broader AI Continent Action Plan, is to establish Europe as a "global leader in AI" and, in doing so, achieve a meaningful degree of "digital sovereignty" by reducing its dependence on foreign technology.1 This infrastructure push is complemented by other strategic initiatives, most notably the €43 billion European Chips Act, which aims to foster a domestic semiconductor industry capable of producing advanced processors, such as SiPearl's Rhea CPU, which is slated for inclusion in future exascale systems.1

Despite these ambitions, the reality remains that the compute and investment gap between Europe and the US is immense and growing (as detailed in Section 2). Critics forcefully argue that the EU's regulation-heavy approach has paradoxically weakened its own innovation capacity, creating a vicious cycle where a lack of domestic champions necessitates more regulation, which in turn further hinders the emergence of those champions.7

This leads to a final, critical assessment: the AI Factories initiative is a necessary but insufficient condition for achieving digital sovereignty. It correctly identifies and addresses a critical symptom—the lack of accessible, domestic compute—but it does not cure the root causes of Europe's technological malaise. These causes are a less dynamic and more risk-averse venture capital market, particularly at the late-growth stage, and a fragmented commercial ecosystem that lacks the scale and unity of the US or Chinese markets.

Digital sovereignty is ultimately an outcome of a thriving, competitive, and self-sustaining commercial market, not just of state-owned infrastructure. The United States leads in AI not because of government-run supercomputing centers, but because of a hyper-competitive market driven by private-sector giants like Google, Meta, and NVIDIA, which is in turn fueled by the world's deepest and most aggressive venture capital ecosystem. The EU's state-led model can provide subsidized R&D infrastructure and help incubate promising early-stage startups, as it has with companies like France's Mistral AI.6 However, the AI Factories do not, by themselves, create the pan-European capital markets or the unified consumer market needed for those startups to scale into global giants without eventually having to look to the US for growth capital and customers.

Therefore, the AI Factories initiative should be viewed primarily as a defensive strategy. It is a vital and necessary investment to prevent Europe from falling further behind in a technology that will define the 21st-century economy. It provides a crucial lifeline of compute resources to a nascent ecosystem that would otherwise be starved of this critical input. However, it is not an offensive strategy that can, on its own, create a European Google or NVIDIA. It is one piece—albeit a very large and expensive one—of a much larger and more complex puzzle.

Section 5: Strategic Recommendations and Outlook

The European Union's AI Factories initiative is a bold and necessary response to a clear and present strategic challenge. It represents a commendable pivot towards building the hard infrastructure required to compete in the age of artificial intelligence. However, as this analysis has demonstrated, the initiative faces significant hurdles related to scale, operational structure, and the broader European innovation ecosystem. For the AI Factories to succeed not merely as impressive engineering projects but as true catalysts for a European AI renaissance, policymakers and stakeholders must adopt a clear-eyed view of these challenges and pursue a series of targeted strategic actions.

Recommendations for EU and National Policymakers:

  1. Streamline and Harmonize the User Experience: The "one-stop-shop" vision is compromised by the federated nature of the AI Factories. The EuroHPC JU must enforce a set of common, simplified operational standards, application programming interfaces (APIs), and user support protocols across all factories. The goal should be to create a user experience that is as close to a single, unified cloud platform as possible, minimizing the administrative burden on startups navigating the network.
  2. Evolve Beyond Project-Based Access Models: The current access model, rooted in academic grant cycles, is a poor fit for commercial startups. The EU should pilot and scale a more flexible, on-demand access tier. This could involve allocating a portion of compute resources to a system that functions more like a commercial cloud, allowing startups to purchase or draw down resources dynamically based on real-time needs, rather than being locked into fixed, year-long project allocations. This would better align the infrastructure with the agile development methodologies that drive startup innovation.
  3. Address the "Subsidy Cliff": The reliance on free compute creates a dangerous cliff-edge for startups when they need to scale. The EU should develop a structured transition program that gradually introduces startups to the real costs of compute. This could involve a tapered subsidy model, where the percentage of free compute decreases as a company matures or secures private funding. This would force startups to build commercially viable business models from day one, preparing them for the realities of the market.
  4. Double Down on Talent and Data Mobility: The most powerful supercomputers are useless without skilled operators and high-quality data. The EU must intensify its efforts to create a true single market for tech talent and data. This requires more aggressive policies to harmonize regulations, simplify cross-border data sharing for AI training (within the bounds of GDPR), and create competitive compensation and equity incentive structures to retain top AI talent in Europe. The "Data Labs" are a good start, but they must be scaled and funded to become more than just appendages to the compute infrastructure.
  5. Focus the AI Act on genuine High-Risk Applications: The current broad-brush approach of the AI Act creates uncertainty and a high compliance burden that disproportionately harms startups. Regulators, guided by the new European AI Office, should work swiftly to issue clear, narrow guidelines that focus enforcement on genuinely high-risk use cases (e.g., in critical infrastructure, medical devices) while providing safe harbors and simplified compliance pathways for the vast majority of low- and medium-risk AI applications. This would reduce the regulatory "tax" on innovation without compromising on core European values.

Outlook:

The AI Factories initiative is not a silver bullet, but it is a critical first step. It successfully addresses the most immediate bottleneck for European AI development: the scarcity of accessible, high-performance computing. By providing this vital resource, the EU has the potential to nurture a new generation of AI startups that would otherwise have been unable to get off the ground. The initiative's success will hinge on its ability to evolve from a collection of publicly-funded supercomputing centers into a genuinely cohesive and user-friendly ecosystem.

The ultimate verdict on the AI Factories will likely be written in the late 2020s. If, by then, a handful of globally competitive, European-born AI companies have emerged from this ecosystem, having leveraged the factories to build their foundational models before successfully transitioning to the commercial market, the initiative will be deemed a success. It will have proven that a state-led, coordinated approach can, at least partially, substitute for the market dynamics of the US.

However, if the ecosystem fails to produce commercially viable champions, and the most promising startups continue to either fail or relocate to the US for growth capital, the initiative will be seen as a costly experiment that built impressive machines but failed to solve the underlying structural problems of the European tech landscape. The AI Factories have provided the European AI ecosystem with a powerful engine; the challenge now is to build the rest of the vehicle and ensure there is a clear road to a competitive global market. The future of Europe's digital sovereignty may well depend on it.

Works cited

  1. The EU AI Continent Action Plan - CSIS, accessed on June 22, 2025
  2. AI Factories - Shaping Europe's digital future - European Union, accessed on June 22, 2025
  3. Commission sets course for Europe's AI leadership with an ambitious AI Continent Action Plan, accessed on June 22, 2025
  4. Financing Infrastructure for a Competitive European AI - Groupe d'études géopolitiques, accessed on June 22, 2025
  5. BSA Leads Industry Call for EU to Prioritize Innovation in AI Regulation, accessed on June 22, 2025
  6. EU's InvestAI Initiative: Can €200 Billion Close the AI Gap with the U.S. and China?, accessed on June 22, 2025
  7. The EU's AI Power Play: Between Deregulation and Innovation, accessed on June 22, 2025
  8. European HPC efforts – ETP4HPC, accessed on June 22, 2025
  9. The European AI Continent Action Plan | Stibbe, accessed on June 22, 2025
  10. EU AI Act's Burdensome Regulations Could Impair AI Innovation - Bloomberg Law News, accessed on June 22, 2025
  11. The EuroHPC JU Selects Additional AI Factories to Strengthen Europe's AI Leadership, accessed on June 22, 2025
  12. EU boosts European AI developers with the AI Factories call for proposals, accessed on June 22, 2025
  13. Mapping the EU AI Landscape (Part 4): AI Factories | Nicolo' Brandizzi, accessed on June 22, 2025
  14. AI Factories Systems - EuroHPC JU - European Union, accessed on June 22, 2025
  15. European High-Performance Computing Joint Undertaking - Wikipedia, accessed on June 22, 2025
  16. A EuroHPC Success Story: Advancing Radiotherapy, accessed on June 22, 2025
  17. Discover EuroHPC JU | HPC Serbia, accessed on June 22, 2025
  18. EuroHPC JU access calls for AI Factories - LUMI supercomputer, accessed on June 22, 2025
  19. AI Factories Access Modes - EuroHPC JU - European Union, accessed on June 22, 2025
  20. EU's AI Continent Action Plan: a turning point for digital sovereignty - Telefónica, accessed on June 22, 2025
  21. Inside InvestAI — Europe's ambitious $20 billion plan to build AI gigafactories, accessed on June 22, 2025
  22. European Commission – Invest AI initiative - Knowledge Centre Data & Society, accessed on June 22, 2025
  23. Fresh UK-EU collaboration on AI to unlock new avenues for innovation and research, accessed on June 22, 2025
  24. AI Factory Antenna Applications Now Open - techUK, accessed on June 22, 2025
  25. Europe's AI factories, data centre development and the vital role of dispatchable power, accessed on June 22, 2025
  26. JUPITER – The New Dimension of Computing - Forschungszentrum Jülich, accessed on June 22, 2025
  27. NVIDIA Powers Europe's Fastest Supercomputer, accessed on June 22, 2025
  28. Thinking Outside Of The Box With The Jupiter Supercomputer Datacenter, accessed on June 22, 2025
  29. LUMI - Wikipedia, accessed on June 22, 2025
  30. LUMI - Sigma2 documentation, accessed on June 22, 2025
  31. MareNostrum 5, the eighth most powerful supercomputer | IQTC - The Institute of Theoretical and Computational Chemistry of the Universitat de Barcelona - IQTCUB, accessed on June 22, 2025
  32. LUMI AI Factory Launches as EU Hub for AI Talent, Data, and Compute - HPCwire, accessed on June 22, 2025
  33. Leonardo (supercomputer) - Wikipedia, accessed on June 22, 2025
  34. Leonardo | HPC Cineca, accessed on June 22, 2025
  35. Leonardo | HPC Cineca, accessed on June 22, 2025
  36. System overview - MareNostrum 5 | BSC Support Knowledge Center, accessed on June 22, 2025
  37. MareNostrum 5 | BSC-CNS, accessed on June 22, 2025
  38. Jean Zay supercomputer : France has increased its AI dedicated resources fourfold | CNRS, accessed on June 22, 2025
  39. Jean Zay: Eviden BullSequana XH3000, HPE SGI 8600 supercomputer - IDRIS, accessed on June 22, 2025
  40. Scalability Evaluation of HPC Multi-GPU Training for ECG-based LLMs - arXiv, accessed on June 22, 2025
  41. System overview - MeluXina User Documentation, accessed on June 22, 2025
  42. Vega supercomputer - HPC at the ICP, accessed on June 22, 2025
  43. Vega - izum.si, accessed on June 22, 2025
  44. gpu cluster (partition) - DISCOVERER SUPERCOMPUTER, accessed on June 22, 2025
  45. DISCOVERER SUPERCOMPUTER, accessed on June 22, 2025
  46. Our Supercomputers - EuroHPC JU - European Union, accessed on June 22, 2025
  47. Building Meta's GenAI Infrastructure - Engineering at Meta, accessed on June 22, 2025
  48. Meta to build 2GW data center with over 1.3 million Nvidia AI GPUs ..., accessed on June 22, 2025
  49. Meta Is Comfortable Betting the House on AI, and Investors Should Pay Attention, accessed on June 22, 2025
  50. Google touts AI infrastructure as 'powerful foundation' for growth, accessed on June 22, 2025
  51. AI Hypercomputer inference updates for Google Cloud TPU and GPU, accessed on June 22, 2025
  52. Performance per dollar of GPUs and TPUs for AI inference | Google Cloud Blog, accessed on June 22, 2025
  53. Ask a Techspert: What's the difference between a CPU, GPU and TPU? - Google Blog, accessed on June 22, 2025
  54. TPU vs GPU: What's the Difference in 2025? - CloudOptimo, accessed on June 22, 2025
  55. Dojo supercomputer explained: How Tesla plans to beat Nvidia at AI ..., accessed on June 22, 2025
  56. Tesla Dojo: A Breakthrough in AI and Self-Driving Tech, accessed on June 22, 2025
  57. OpenAI Is A Systemic Risk To The Tech Industry, accessed on June 22, 2025
  58. Announcing The Stargate Project | OpenAI, accessed on June 22, 2025
  59. Stargate LLC - Wikipedia, accessed on June 22, 2025
  60. Infrastructure - AI at Meta, accessed on June 22, 2025
  61. AI & Robotics | Tesla, accessed on June 22, 2025
  62. Closed Door or New Opportunity: Will Europe Survive New American AI Chips Export Rules? - Modern Diplomacy, accessed on June 22, 2025
  63. D1.5-36a Report on successful exploitation of the co- design between this project and EPI / (Pre-)Exascale Hosting Entities - EC Europa, accessed on June 22, 2025
  64. EU Sets Out Plan to Close the AI Gap with US and China - PYMNTS.com, accessed on June 22, 2025
  65. AI Data Centres to be Built by Deutsche Telekom & SAP | Technology Magazine, accessed on June 22, 2025
  66. EuroHPC Expands AI Infrastructure with Six New AI Factories Across Europe, accessed on June 22, 2025
  67. AI Factory Austria (AI:AT) strengthens national AI ecosystem - AIT, accessed on June 22, 2025
  68. AI Factory Austria (AI:AT) to boost national AI ecosystem - EODC, accessed on June 22, 2025
  69. EuroHPC JU access calls for AI Factories - LUMI supercomputer, accessed on June 22, 2025
  70. Supercomputers Access Policy and FAQ - EuroHPC JU - European Union, accessed on June 22, 2025
  71. EuroHPC JU to Enhance EuroHPC Supercomputers Access Platform, accessed on June 22, 2025
  72. EuroHPC JU Call for Proposals for AI and Data-Intensive Applications - European Union, accessed on June 22, 2025
  73. Access Policy and FAQ - EuroHPC JU - European Union, accessed on June 22, 2025
  74. Second wave of AI Factories set to drive EU-wide innovation, accessed on June 22, 2025
  75. Apply to become an AI Factory Antenna: expression of interest | UKTIN, accessed on June 22, 2025
  76. EU AI Act challenges 68% of European businesses, AWS report finds, accessed on June 22, 2025
  77. The EU AI Act: A Double-Edged Sword For Europe's AI Innovation Future - Forbes, accessed on June 22, 2025
  78. EU rules for advanced AI are step in wrong direction, Google says - Politico EU, accessed on June 22, 2025
  79. Shaping Europe's leadership in artificial intelligence with the AI continent action plan - European Commission, accessed on June 22, 2025
  80. Betting on EU AI Infrastructure: Navigating Fragmentation and the Race for Digital Sovereignty - AInvest, accessed on June 22, 2025
  81. EU Commission Publishes AI Continent Action Plan and Seeks Input - Data Matters, accessed on June 22, 2025