Europa Nova

Executive Summary

Europe has one frontier AI laboratory. The United States has at least five. China has four and counting. Of the 36 AI models that have crossed the 10²⁵ FLOP training threshold since GPT-4 in March 2023, 28 came from American labs, six from Chinese labs, and two from Europe — both from Mistral. Industry now controls 80 per cent of global AI compute ^[8], up from 40 per cent in 2019, and 95 per cent of commercially available AI compute is operated by US or Chinese companies ^[8]. The EU's entire EuroHPC network amounts to approximately 57,000 high-end accelerators ^[3][7]. Mark Zuckerberg alone planned 1.3 million GPUs by the end of 2025. This is not a competitiveness gap. It is a structural absence — of institutions, of concentrated talent, of architectural ambition — at the centre of the most consequential technology race of this century.

The European Commission has responded with real ambition: nineteen AI Factories ^[3], five planned Gigafactories ^[4], €20 billion through InvestAI ^[1], and an AI Continent Action Plan that promises to triple data centre capacity within seven years. But three independent analyses — by CEPS, Interface, and the Bertelsmann Stiftung ^[6][7][8] — converge on the same structural diagnosis. The Gigafactories lack anchor customers with sustained frontier training demand; Europe has one frontier lab, and it does not yet operate its own compute at scale. The factories are geographically dispersed across member states rather than concentrated where energy is cheapest and talent is deepest. And they depend almost entirely on NVIDIA's proprietary technology stack — hardware, CUDA, networking, factory design — potentially deepening the dependencies they were created to reduce. The EU has built a compute layer. It has not built the research organisation that gives compute a purpose.

This paper proposes Europa Nova: a treaty-based research institution, modelled on CERN but with a commercial licensing arm, dedicated to developing safe artificial general intelligence. Established by treaty among five to eight nations — France, Germany, the United Kingdom, Norway, Sweden, the Netherlands, Switzerland, and Finland — it operates through three pillars. A Research Foundation governed by nine internationally recruited scientists, led by a Scientific Director with veto authority on technical decisions and compensation at €1-2 million. One concentrated research hub, not a network distributed across capitals. A Compute Infrastructure Division in Northern Scandinavia, running on 100 per cent renewable hydropower at €0.03/kWh, with Arctic cooling reducing energy costs by 30--60 per cent, and initial capacity of 500MW expandable to 2GW. OpenAI chose Narvik for Stargate Norway ^[15][16] for exactly these reasons. A Commercial Licensing Entity, wholly owned by the Foundation, licensing models and compute to European industry and the public sector. Revenue funds the mission. The mission generates the technology. Europa Nova's charter is protected by international treaty, not corporate bylaws amendable by a board vote.

Europa Nova is what the EU's Gigafactories currently lack: an anchor customer with sustained frontier training demand. It complements the Commission's compute infrastructure. It does not replace it. Six governance principles define the design, each a response to a documented failure in European collaboration. Concentration over distribution — because dispersed resources produce dispersed mediocrity. Scientific autonomy — because political direction of research produces incremental work. Market-rate compensation — because the alternative is not cheaper talent but no talent, as 57 per cent of the world's top AI researchers already work in the United States ^[24]. A built-in sunset clause at year three, with independent evaluation and provisions for orderly wind-down if results are insufficient — this is what makes the proposal credible to sceptics. No juste retour — because Airbus's workshare distortions demonstrate the cost of proportional return. And transparent democratic oversight without operational control, modelled on central bank independence.

€50 billion over ten years. Funded by four streams: sovereign treaty contributions of approximately €3 billion per year from five to eight nations; a 1-2 per cent allocation from Norway's Government Pension Fund Global, structured as infrastructure investment with residual value, not a grant; commercial licensing revenue from year four, projected at €500 million rising to €2-3 billion annually; and EIB co-financing for physical infrastructure. Context matters. OpenAI's Stargate is $500 billion ^[17]. Microsoft alone spent $80 billion on AI infrastructure in fiscal 2025 ^[18]. Alphabet, Meta, Microsoft, and Amazon collectively projected more than $380 billion in AI capex for the current year ^[18]. Europa Nova's €5 billion per year is roughly four times CERN's budget and two-thirds of ESA's. It is the minimum credible stake.

Europa Nova does not attempt to out-scale OpenAI on the existing Transformer architecture. That is a race Europe enters a decade late and a trillion dollars short. Instead, it invests in architectural innovation — state-space models, hybrid architectures, mixture-of-experts — that changes the scaling laws themselves, making each unit of compute yield more capability. Its data strategy converts Europe's multilingual diversity and structured public-sector records into a training advantage that monolingual English corpora cannot replicate. Its alignment approach — pluralistic, rooted in European traditions of navigating genuine value diversity — addresses the hardest unsolved problem in the field. Europe did not win the race to build the first particle accelerator. It built CERN, and CERN found the Higgs boson.

Convene a founding summit of heads of state from five to eight European nations within twelve months. Appoint sherpas with authority to negotiate a treaty. Begin the search for a Scientific Director. Assess sites in Northern Scandinavia. Engage the EIB. CERN's founding convention was negotiated in under eighteen months. Europa Nova should aim for the at least same pace. The technology will not wait for consensus among twenty-seven member states — which is why this proposal is a treaty among five to eight, not an EU regulation for all. The question is not whether artificial general intelligence will arrive. The question is whether Europe will have any voice in what it becomes — or whether it will, once again, write the rules for a technology it depends on others to build.

1 Europe’s Strategic AI Deficit

Artificial general intelligence is being built. It is being built in San Francisco, in Beijing and in Shenzhen. It is not being built in Europe.

This is not a statement about abstract technological competitiveness or innovation rankings. It is a concrete assessment of where the organisations, the compute infrastructure, and the capital capable of producing the most consequential technology of the twenty-first century are concentrated. The answer, for Europe, is uncomfortable. The continent that invented the scientific method, that hosts CERN and the European Space Agency, that designed the regulatory architecture now imitated worldwide, does not possess a single institution capable of training a frontier AI model at the scale required to matter. That gap is not closing. It is widening.

1.1  The Concentration of AGI Development

Frontier AI development is concentrated in fewer than a dozen organisations worldwide. In the United States, five labs dominate: OpenAI, Anthropic, Google DeepMind, Meta AI, and xAI. In China, DeepSeek, Baidu, Alibaba, and an emerging cluster of state-backed competitors are advancing rapidly despite semiconductor export controls. In Europe, there is one: Mistral, a French company founded in 2023, which has produced two models at the frontier scale of 10²⁵ floating-point operations. One lab. Against at least nine.

The scale of capital behind this concentration is staggering. The combined AI-related capital expenditure of Microsoft, Google, Amazon, and Meta exceeded $200 billion in 2024, with Microsoft, Meta, Amazon, and Apple earmarking over $300 billion in capex for 2025, primarily for data centre construction and IT equipment ^[18]. OpenAI's Stargate initiative, announced in partnership with SoftBank and Oracle ^[17], represents $500 billion in planned US AI infrastructure. The EU's InvestAI programme ^[1], announced by Commission President von der Leyen at the Paris AI Action Summit in February 2025, mobilises €200 billion — of which only €20 billion is public money. The asymmetry is not merely financial. It is structural.

The Interface and Bertelsmann Stiftung analysis ^[8], published in October 2025, documents the trajectory with precision. The industry share of global AI compute rose from 40 per cent in 2019 to 80 per cent in 2025, with 95 per cent of commercially available AI compute infrastructure operated by companies headquartered in the United States or China ^[8]. Europe's position in this landscape is marginal. Mistral, the continent's only frontier lab, does not operate its own compute infrastructure at scale. Its announced partnership with Fluidstack for a gigawatt-class campus in Bruyères-le-Châtel will not become operational until 2026 at the earliest, with full capacity projected for 2028. Until then, Europe's sole frontier AI company depends on external, non-European compute infrastructure to train its models.

The CEPS In-Depth Analysis by Renda and Kyosovska ^[6] adds a layer of structural concern. Europe's AI Gigafactories — up to five facilities with at least 100,000 advanced chips each — will rely overwhelmingly on Nvidia's vertically integrated technology stack: GPUs, the proprietary CUDA software platform, networking, and factory design. Nvidia is not a European company. Its chips are designed in the United States and manufactured by TSMC in Taiwan. The CEPS analysis identifies a pointed irony: infrastructure built in the name of European sovereignty may, in practice, deepen the very dependencies it purports to address.

The EuroHPC AI factories — thirteen sites announced before October 2025, each equipped with up to 25,000 GPU equivalents ^[3][7] — collectively amount to approximately 57,000 high-end AI accelerators. Mark Zuckerberg announced plans to have 1.3 million GPUs operating by the end of 2025. The scale mismatch is not a rounding error. It is orders of magnitude.

1.2  The Dependency Risk

The absence of sovereign AGI capability is not an abstract competitiveness concern. It is a dependency that reaches into every sector of the European economy and every function of the European state.

Consider the sectors. Financial services — algorithmic trading, credit scoring, fraud detection, regulatory compliance — depends increasingly on AI systems that European banks licence but do not control. Healthcare, where AI-assisted diagnostics and drug discovery are advancing fastest, will within a decade be shaped by models trained on data that European hospitals provided but American companies own. Defence and intelligence analysis, where autonomous systems and pattern recognition are transforming operational capability, will be determined by access to models developed under foreign jurisdiction. Public administration — welfare allocation, tax compliance, regulatory enforcement — is adopting AI systems whose decision logic European citizens cannot inspect because the models are proprietary and the companies that built them are not subject to European oversight in any meaningful operational sense. Industrial manufacturing, from predictive maintenance on automotive production lines to supply chain optimisation, is being rebuilt around AI platforms controlled from outside Europe.

The dependency runs deeper than sector-by-sector exposure. It is systemic. When a European government deploys an AI system for welfare eligibility determination or border security screening, the model's training data, its decision boundaries, and its failure modes are opaque not because of technical complexity alone, but because they are proprietary assets of a foreign corporation. When a European pharmaceutical company uses an AI platform for drug candidate identification, the intellectual property generated on that platform exists in a legal grey zone shaped by American contract law and American terms of service. When a European defence ministry integrates AI-assisted intelligence analysis, it does so on infrastructure that another government can, in principle, monitor, restrict, or shut down.

The obvious counterargument is that Europe can simply purchase access to these systems. Global markets function. American companies are happy to sell. This reasoning has three weaknesses.

The first is the export control precedent. The United States has already demonstrated its willingness to restrict access to advanced technology when strategic interests demand it. The October 2023 semiconductor export controls, expanded in January 2025, cut China off from the most advanced Nvidia GPUs — not because of a trade dispute, but because Washington judged that AI capability constitutes a national security asset. No formal restriction on European access exists today. But the legal and institutional machinery to impose one is in place, and the precedent has been set. A future US administration facing a trade confrontation with the EU, or seeking diplomatic concessions, would possess an instrument of extraordinary coercive power: the ability to throttle European access to the AI infrastructure on which European economies increasingly depend. The CEPS report ^[6] documents this risk explicitly, noting that Europe's gigafactories will be built on American hardware and that Europe has, at present, no credible alternative.

The second weakness is data sovereignty. European business data, government data, and citizen data processed on American cloud infrastructure falls under American jurisdiction. The theoretical protections of the EU-US Data Privacy Framework do not alter the fundamental asymmetry: data processed on systems controlled by Microsoft, Google, or Amazon is accessible to American authorities under FISA Section 702 and Executive Order 12333. The EU AI Act's transparency and oversight requirements — including obligations for high-risk AI systems to undergo conformity assessments and maintain human oversight — are difficult to enforce on systems whose training data, model weights, and inference infrastructure are controlled by entities outside European jurisdiction.

The third weakness is the risk of architectural lock-in. The CEPS analysis ^[6] devotes particular attention to Nvidia's role in the AI infrastructure supply chain. Nvidia does not merely sell GPUs. It provides a complete technology stack — hardware, the CUDA software platform, networking components, and full factory design services. The CEPS report warns that this vertical integration creates a dependency comparable to the Microsoft-Intel "Wintel" dominance of personal computing in the 1990s. Once European infrastructure is built on CUDA, switching to alternative platforms — whether AMD's ROCm, Google's TPUs, or future European or Asian alternatives — becomes technically and economically prohibitive. China has already responded to this risk. In 2025, Chinese companies began restricting purchases of Nvidia chips and accelerating domestic alternatives, including Huawei's CloudMatrix system and chips developed by Alibaba and Tencent. Europe, by contrast, is doubling down on the Nvidia stack.

The Interface/Bertelsmann report ^[8] frames the structural problem bluntly. Of the 36 AI models that have crossed the 10²⁵ FLOP threshold since GPT-4 in March 2023, 28 were trained by US-based labs. Six came from China. Two came from Europe — both from Mistral. Europe is not a participant in frontier AI development. It is a spectator with a regulatory clipboard.

1.3  What Europe Has — and What It Lacks

An honest assessment of Europe's position must begin with what Europe possesses. The deficits are real, but so are the assets. Confusing weakness with helplessness would be as dangerous as confusing ambition with capability.

Europe's scientific institutions remain world-class. CERN demonstrated that European nations, when they commit to concentrated investment in fundamental research, can lead the world. The European Space Agency operates the Copernicus earth observation programme and the Galileo navigation system — both sovereign European capabilities in domains that the United States and China also dominate. Max Planck, CNRS, INRIA, and the ETH system produce research of the highest calibre. ASML, headquartered in the Netherlands, manufactures the extreme ultraviolet lithography machines without which no advanced semiconductor on Earth can be produced. Carl Zeiss, in Germany, supplies the optical systems at the heart of those machines. Europe is not absent from the technology supply chain. It occupies chokepoints of extraordinary strategic value.

The talent pipeline is strong at the source. European universities train excellent AI researchers. The problem is retention. Yann LeCun, a Turing Award laureate and one of the founders of modern deep learning, is French. He works at Meta AI in New York. Demis Hassabis, a Nobel Prize laureate and co-founder of DeepMind, is British. DeepMind was acquired by Google in 2014 and operates under Alphabet's control. The Global AI Talent Tracker ^[24] records that while 16 per cent of the world's top AI researchers were trained in France, Germany, and the United Kingdom combined, 57 per cent of all top-tier AI researchers work in the United States. Europe exports its best minds and then purchases their output.

Renewable energy is an underappreciated European strength. Norway, Sweden, Finland, and Iceland possess abundant hydroelectric and wind power at costs that American and Chinese data centre operators have noticed before European policymakers did. OpenAI chose Kvandal in northern Norway for Stargate Norway ^[15][16] — its first European data centre — because of hydropower, not because of proximity to European AI talent. Microsoft, CoreWeave, and Brookfield collectively committed over $15 billion to Nordic AI infrastructure in 2025 ^[18]. Northern Europe's cool climate reduces cooling costs by 30 to 60 per cent compared to conventional data centre locations. The IEA's April 2025 report ^[10] confirms that global data centre investment surpassed $500 billion in 2024 and is projected to exceed $800 billion annually before 2030, with energy availability emerging as the binding constraint on further expansion. Europe has energy. It does not yet have the institutional architecture to convert that energy into sovereign compute.

Regulatory experience constitutes a genuine, if double-edged, advantage. The EU AI Act, which entered into force in 2024, is the world's first comprehensive legal framework for artificial intelligence. It establishes risk-based categorisation, conformity assessment procedures, and transparency obligations that other jurisdictions are studying. The risk is that regulation without production capacity amounts to setting the rules for a game Europe does not play. Regulatory authority divorced from industrial capability is influence on paper.

Against these assets, Europe's weaknesses are specific and identifiable. The venture capital gap is severe. US AI startups raised approximately $67 billion in 2024 ^[23]; European AI startups raised roughly $12 billion. AI adoption across European enterprises remains low: Eurostat data ^[25] show that only 13.5 per cent of European enterprises reported using AI technologies in 2024, far below the EU's own Digital Decade target of 75 per cent by 2030. The European paradox — strong in fundamental research, weak in commercial translation — persists. Europe produces scientific papers. America and China produce products.

Decision-making speed compounds the structural disadvantage. xAI built Colossus, a 100,000-GPU training cluster in Memphis, Tennessee, in 122 days. The EU's AI Gigafactory selection process — from the February 2025 announcement to the formal call for expressions of interest, the evaluation of 76 submissions, and the eventual site selection — will take considerably longer than that. European decision-making is designed for legitimacy and consensus. Those are virtues in treaty negotiation. They are liabilities in a technology race where eighteen months of delay can mean an entire generation of architectural progress missed.

The absence of concentrated compute infrastructure at frontier scale is the deficit that underpins all others. Without the ability to train models at the 10²⁵ FLOP threshold and beyond, European researchers cannot compete on architecture. Without competing on architecture, European institutions cannot attract and retain talent. Without talent, no amount of regulatory authority or renewable energy translates into sovereign capability. The deficits are connected, and compute is the load-bearing constraint.

The EU's response to date — the AI Continent Action Plan, the InvestAI programme, the AI factories and planned gigafactories — acknowledges the problem. It does not yet solve it. The Commission's approach remains principally an infrastructure deployment programme: build compute, and frontier AI will follow. The Interface/Bertelsmann analysis ^[8] challenges this assumption directly, arguing that the supply-side focus overlooks the demand side — Europe lacks not only GPUs, but the organisations capable of fully utilising them. The CEPS analysis ^[6] questions whether the factories are being built in the right locations, with the right architecture, and for the right kind of AI.

Both analyses ^[6][8] converge on a conclusion that this paper shares: infrastructure is necessary but not sufficient. Europe requires a research institution with the scientific ambition, the concentrated talent, the dedicated compute, and the governance architecture to operate at the frontier. Building factories without a research organisation capable of using them is like constructing a particle accelerator without a physics department.

The picture, however, is not static. Three developments are converging to create a window of opportunity that did not exist two years ago: a shift in the architectural frontier that favours innovation over brute-force scaling, a transformation in energy economics that turns Northern Europe's renewable abundance into a decisive competitive asset, and a geopolitical realignment that makes sovereign AI capability a prerequisite for strategic autonomy rather than a luxury. Each of these developments is examined in the chapter that follows.

2 The Window of Opportunity

Windows of opportunity in technology do not announce themselves. They are visible only to those who understand the structural conditions that create them, and they close faster than most institutions can respond. The founding of CERN in 1954 exploited a narrow post-war moment when the United States had not yet consolidated its dominance in particle physics and European governments were unusually willing to pool sovereignty. Airbus was created in 1970 precisely when Boeing's development costs were straining a single national champion model. In both cases, the opportunity was time-limited. Had Europe waited a decade, neither institution would have been possible.

Today, three converging factors create a comparable opening for European leadership in artificial intelligence. The technical architecture of frontier AI is in flux. Energy has emerged as the binding constraint on further progress. And the geopolitical landscape is shifting in ways that both necessitate and enable a European alternative. Each of these conditions is temporary. Within five years, the architectural landscape will consolidate, energy infrastructure investments will lock in geographic advantage, and the geopolitical settlement around AI governance will harden. Europe must act before the window closes.

2.1  The Architecture Is Not Settled

The Transformer architecture, introduced by Vaswani and colleagues at Google Brain in 2017 ^[19], is the foundation of every frontier AI model deployed today. GPT-4, Claude, Gemini, and Llama are all Transformer-based. This dominance has led to a widespread assumption that progress in AI is now primarily a function of scale: more parameters, more data, more compute. That assumption is wrong, or at best incomplete.

The scaling laws that governed the Transformer era are encountering diminishing returns at the frontier. The Chinchilla analysis by Hoffmann and colleagues ^[22] at DeepMind demonstrated that many models were dramatically over-parameterised relative to their training data. Training costs for each successive generation of frontier model have grown faster than the corresponding improvement in capability. GPT-4's training is estimated to have cost over $100 million. The next generation of models at comparable laboratories is projected to cost $1--5 billion. The energy consumed by large-scale training runs is now measured in tens of gigawatt-hours per run, a figure that the IEA projects will continue to rise sharply as the sector's total electricity consumption more than doubles by 2030. The Transformer's core mechanism — self-attention — scales quadratically with sequence length. This means that processing longer contexts requires exponentially more computation. For many real-world applications, from scientific document analysis to long-form reasoning, this is a fundamental bottleneck.

The technical community knows this. Research into alternative architectures has accelerated dramatically since 2023. The most significant challenge comes from the state-space model family. Mamba, developed by Gu and Dao ^[20] at Carnegie Mellon and Princeton, achieves linear-time sequence modelling — meaning computational cost grows proportionally with input length rather than quadratically. This builds on the earlier S4 architecture by Gu, Goel, and Ré ^[21] at Stanford, which demonstrated that structured state-space models could match Transformer performance on long-range tasks at a fraction of the computational cost. The Mamba-2 framework further unifies selective state-space models with structured attention, suggesting that the boundary between these architectural paradigms is more porous than it first appeared.

Hybrid architectures that combine selective attention mechanisms with SSM efficiency represent the most active research frontier. Jamba, developed by AI21 Labs, interleaves Transformer and Mamba layers within a single model. Several research groups are exploring continuous-time neural networks that process information as differential equations rather than discrete token sequences. The architectural landscape, in short, is open. No single approach has consolidated the advantages of Transformers with the efficiency of state-space models.

European research groups are well-positioned in this space. ETH Zürich's AI Center has published leading work on efficient attention mechanisms. INRIA's research teams in Paris and Grenoble are among the world's strongest in mathematical foundations of machine learning. The Max Planck Institute for Intelligent Systems in Tübingen has produced foundational contributions to neural architecture design. The University of Edinburgh and Cambridge maintain internationally competitive groups in probabilistic machine learning and Bayesian methods that are directly relevant to next-generation architectures.

The strategic implication is direct. A new entrant that bets on architectural innovation rather than brute-force scaling of the existing Transformer paradigm is not entering a race already lost. It is positioning for the next paradigm. This is precisely where a concentrated European research effort adds value that raw compute infrastructure alone cannot provide. Compute can be purchased. Architectural insight cannot.

2.2  Energy as the Decisive Constraint

The AI industry has entered an era defined not by the availability of algorithms or even processors, but by the availability of electricity. Training a single frontier model now requires power equivalent to what tens of thousands of households consume in a year. The IEA estimates that global data centre electricity consumption reached approximately 415 TWh in 2024 — roughly 1.5% of the world's total electricity demand — and projects this to rise to approximately 945 TWh by 2030 ^[10] in its Base Case scenario. That projected figure exceeds Japan's entire current electricity consumption.

This demand is not distributed evenly. It is concentrated in a small number of urban clusters that developed as data centre hubs over the past two decades. In Europe, the FLAP-D markets — Frankfurt, London, Amsterdam, Paris, and Dublin — host approximately 62% of the continent's data centre capacity ^[12]. These markets are now saturated. The IEA reports ^[10][11] that grid connection queues for new data centres range from three to five years in Spain, up to seven years in Germany and the United Kingdom, and up to ten years in the Netherlands. Dublin has paused new data centre connections entirely until 2030. Grid congestion volumes in Germany and Great Britain have increased year on year since 2019, even as congestion costs have fluctuated with gas prices.

The Ember analysis ^[12] of European grid capacity, published in June 2025, quantifies the structural shift underway. The FLAP-D markets' share of European data centre capacity is projected to fall from 62% today to 55% by 2030 and 51% by 2035. Of the five traditional hubs, only France — with its large nuclear fleet — retains relatively unconstrained grid capacity for new connections. Growth outside the FLAP-D markets is expected to be far stronger: data centre demand in non-FLAP-D countries will increase by 110% by 2030, double the 55% growth rate within the established hubs.

This is not a problem that better permitting will solve, though permitting improvements would help. The constraint is physical. Transmission lines take four to eight years to build in advanced economies. Order backlogs for power transformers grew by more than 30% in 2024, and the price index for transformers has increased by 1.5 times since 2020. The infrastructure bottleneck is deep and structural.

Northern Europe's position in this constrained landscape is distinctive. Norway generates approximately 98% of its electricity from renewable sources, predominantly hydropower. Long-term Nordic power purchase agreements are available at rates as low as €0.03/kWh — a third to a quarter of the industrial electricity prices in major Continental European markets. The cold climate of Scandinavia and the Arctic reduces cooling costs for data centres by 30--60% compared with conventional locations in Central and Southern Europe. Finland's main transmission grid achieved a reliability rate of 99.99995% in 2025, among the highest in the world, and Fingrid reports receiving over 100 GW of grid connection enquiries, more than half related to data centre projects.

The market has already validated this logic with capital commitments of historic scale. In July 2025, Nscale, Aker, and OpenAI announced Stargate Norway ^[15][16] — OpenAI's first European data centre, located near Narvik in northern Norway. The facility is planned to deliver 230MW of initial capacity with ambitions to expand by an additional 290MW, targeting 100,000 NVIDIA GPUs by the end of 2026. The site at Kvandal was selected for its abundant hydropower, low local electricity demand, cool climate, and established industrial infrastructure from Norway's aluminium and fertiliser sectors. Microsoft, CoreWeave, and Brookfield have collectively committed over $15 billion to Nordic AI infrastructure in 2025, driven by the same calculus.

The Nordic countries also benefit from deliberate planning by their transmission system operators. Statnett in Norway is planning for a tripling of data centre electricity demand by 2030. Energinet in Denmark began building high-voltage substations for future data centre connections as early as 2017. Fingrid in Finland has accelerated its investment programme, commissioning the Aurora Line cross-border connection between northern Finland and northern Sweden ahead of schedule in late 2025. This is not accidental advantage. It is the product of long-term infrastructure policy that continental Europe failed to match.

The strategic implication should be stated bluntly. Europe's perceived weakness — the absence of concentrated technology infrastructure in established data centre hubs — becomes a strength when the binding constraint shifts from software to energy. Northern Europe possesses the resource that will define the next decade of AI infrastructure: abundant, affordable, reliable, and clean electricity. A European AGI initiative anchored in the Nordics starts with a structural energy advantage that no American or Chinese competitor can replicate.

2.3  The Geopolitical Window

Energy and architecture define the technical opportunity. Geopolitics defines the urgency. Three dynamics are converging to create a narrow period in which a European initiative is not merely desirable but necessary.

US export controls as precedent and risk. The United States has already established the principle that advanced computing technology is a strategic asset subject to export control. The semiconductor restrictions imposed on China since October 2022, progressively tightened through 2023 and 2024, culminated in the January 2025 Framework for Artificial Intelligence Diffusion, which established tiered access controls for AI chips and model weights based on a country's perceived alignment with US interests. Europe falls into the most favourable tier, but the precedent is set: access to frontier AI capability can be restricted unilaterally by the country that controls the supply chain. European dependence on American AI systems — not just hardware but models, APIs, and cloud infrastructure — carries sovereign risk that grows with each generation of capability. The risk is not that the United States will restrict European access tomorrow. The risk is that it could, and that this possibility alone should reshape European strategic planning.

China's acceleration. The release of DeepSeek's V3 model in December 2024 demonstrated that Chinese laboratories can produce frontier-competitive models at a fraction of the training cost of their American counterparts. Huawei's Ascend AI chips offer an alternative to NVIDIA hardware for training workloads, reducing China's vulnerability to US semiconductor controls. The Chinese state has directed massive compute infrastructure buildout, with data centre electricity consumption growing at 15% annually over the past decade. The emerging bipolar structure of AI competition — a US-China duopoly with all other nations as consumers — is the scenario that European policymakers should find most alarming. In a bipolar AI world, Europe's regulatory influence, its values around privacy and human rights, and its economic interests would be marginalised. A European third path is not a luxury. It is a geopolitical necessity.

NATO and AI as security infrastructure. The Nordic countries — Norway, Sweden, Finland, and Denmark — are NATO members with stable democratic governments and strong rule of law. Sweden and Finland joined NATO in 2023 and 2024 respectively, completing the Alliance's northern flank. AI compute located in allied, democratic nations with robust cybersecurity frameworks is a security asset, not merely a commercial one. As AI becomes embedded in defence systems, intelligence analysis, and critical infrastructure protection, the geographic location of training and inference infrastructure acquires military significance. European AI compute in Northern Europe serves both commercial and strategic purposes simultaneously.

The UK factor. Any serious European AI initiative must include the United Kingdom. DeepMind, based in London, is one of the world's three or four leading AI research laboratories. ARM's chip architecture underpins a significant share of global computing. The UK AI Safety Institute, established in 2023, has developed evaluation methodologies that inform international AI governance discussions. Brexit removed the UK from EU institutional structures, which means that any initiative built within the EU framework automatically excludes Britain. Europa Nova's treaty-based structure — modelled on CERN, which includes non-EU members Switzerland and the United Kingdom among its member states — solves this problem naturally. The UK can participate as a founding member without any requirement for EU accession or special bilateral arrangements.

These four dynamics are not permanent features of the international landscape. US export control policy may evolve. China's technological trajectory could plateau. NATO's strategic priorities will shift. The UK's relationship with European institutions will continue to develop in ways that are difficult to predict. The current configuration — in which all four factors simultaneously favour a European initiative — is a conjunction, not a steady state. Conjunctions pass.

*  *  *

The European Union has recognised, to its credit, that AI infrastructure requires urgent attention. The AI Continent Action Plan of April 2025 set a goal of tripling data centre capacity within five to seven years. The AI Factories and Gigafactories initiatives allocate substantial public funding to compute infrastructure. The European data centre market, valued at $47 billion in 2024, is projected to reach $97 billion by 2030 ^[13]. Recognition is not the same as effective execution. The current approach addresses one element of the problem — compute — while leaving the research organisation, architectural ambition, and governance design that would translate infrastructure into capability entirely unaddressed. Chapter 3 examines why.

3 Why Current EU Initiatives Are Insufficient

The European Commission has correctly identified the problem. Europe lacks the compute infrastructure to train frontier AI models. The Commission has responded with speed and ambition: nineteen AI factories, five planned gigafactories, €20 billion in public investment through InvestAI, and a broader AI Continent Action Plan that promises to triple European data centre capacity within seven years. This is the most aggressive public intervention in computing capacity that Europe has ever attempted.

The question is whether the chosen instruments can solve it.

Three independent analyses published in late 2025 — by the Berlin-based think tank Interface, by Interface and the Bertelsmann Stiftung jointly, and by the Centre for European Policy Studies in Brussels ^[6][7][8] — arrive at the same conclusion through different methodologies. The EU is building a compute layer without building the research organisation that gives compute a purpose. This is not a failure of ambition. It is a structural gap. Compute infrastructure is one layer in a multi-layer system, and without the layers above it — architectural research, data curation, alignment science, concentrated scientific talent — the hardware serves whoever arrives first to use it. On current trajectory, that will not be a European organisation.

3.1  The AI Factories: A Research Success That Cannot Scale

The AI factory programme is best understood as an upgrade to Europe's existing high-performance computing network. Nineteen sites, each equipped with up to 25,000 H100 GPU equivalents, are built around EuroHPC supercomputers. They offer tiered access — from playground allocations delivered within two working days to large-scale grants exceeding 50,000 GPU hours over twelve months — at no cost to researchers, startups, and SMEs. Three of the host machines rank in the global top ten: Jupiter in Jülich, LUMI in Kajaani, Leonardo in Bologna. For a machine learning researcher in Athens or Bratislava, this represents genuine access to infrastructure that was previously available only through American hyperscalers or well-funded Western European universities. That achievement matters.

It also marks the boundary of what the factories can do. The Interface analysis ^[7] of the thirteen factories selected before October 2025 found that while they support research in training medium-sized AI models, they are not sufficient to drive commercial AI innovation across the EU at scale. Swiss researchers recently trained a 70-billion-parameter model on the Alps supercomputer in Lugano, which is equipped with 10,752 specialised chips. That result demonstrates what public compute can achieve. It also shows the distance to the frontier: current state-of-the-art models contain hundreds of billions to trillions of parameters and train on clusters of 100,000 GPUs or more, consuming electrical power equivalent to a small city. The AI factories are approximately four times too small to enter that territory.

The institutional composition deepens the constraint. Ten of the thirteen factory consortia include at least one university or research institution. Only two lack an academic partner entirely. Globally, the picture is inverted: private industry controls 80% of AI-specific compute, up from 40% just six years ago. The factories were designed as research infrastructure. They function well as research infrastructure. They cannot, without fundamental restructuring, serve as the engine of commercial AI development the Commission envisions.

The missing element is not larger GPUs. It is the organisation that decides what to train, on what data, for what purpose, and with what safeguards. A particle accelerator without physicists is a tunnel. An AI factory without a research programme is a server farm.

3.2  The Gigafactories: Three Structural Weaknesses

The gigafactory initiative addresses the scale problem directly. Five sites, each hosting at least 100,000 H100 GPU equivalents, would give Europe installations comparable in raw compute to those being built by frontier AI laboratories in the United States. The €20 billion InvestAI fund covers roughly one-third of capital expenditure per site, with private partners providing the rest. Seventy-six expressions of interest were submitted in mid-2025, with sixty sites proposed across sixteen member states. The ambition is real. So are three structural weaknesses that independent analysts have identified in the design.

Who will use them? The Interface/Bertelsmann analysis ^[8] delivers the sharpest critique. The gigafactory initiative, they argue, places too much weight on Europe's past compute limitations and overlooks a critical factor: demand.

Every large-scale compute installation in the private sector is anchored by a single, dominant customer with sustained training demand. OpenAI anchors the $500 billion Stargate project. xAI built its 200,000-chip Colossus cluster in Memphis to train its own models. Europe has one frontier AI laboratory: Mistral, in Paris. Julia Hess, co-author of the Interface/Bertelsmann report ^[8], stated plainly that the stated goal of training frontier models on the gigafactories was never realistic given this constraint. Without an anchor customer, the gigafactories must assemble demand from hundreds of smaller users — fine-tuning runs, academic projects, specialised industrial models — to justify 100,000-GPU installations.

That aggregation model is viable. It is also the exact business model of neoclouds like CoreWeave and Lambda, which already offer flexible GPU access with hourly rental, surrounding services, and the ability to scale workloads dynamically. Whether publicly subsidised gigafactories can compete on flexibility with purpose-built commercial providers is an open question. It is made more awkward by the fact that several neocloud operators already operate on European soil: Nebius, headquartered in Amsterdam, provides GPU access within EU jurisdiction without public subsidy. If the gigafactories end up competing with private providers for the same pool of mid-tier customers, the €20 billion investment becomes a subsidy for cloud computing rather than a vehicle for sovereign AI.

The obvious counterargument is that public investment creates the conditions for demand to emerge. Build the infrastructure, the reasoning goes, and the users will come. European industrial history offers some support for this view — but also a cautionary precedent. Airbus succeeded because it concentrated design capability in a single commercial entity and paired it with guaranteed procurement from national airlines. It did not succeed by building five fuselage factories across five countries and hoping aerospace companies would materialise to fill them. The gigafactories, without an institutional customer with the mandate and capability to conduct frontier training, risk the latter outcome.

Where are they being built? The CEPS analysis by Renda and Kyosovska ^[6] examines the geography of the factory programme using data on patents, scientific publications, startup investment, AI job vacancies, and electricity prices. The findings are uncomfortable. The factories are being built mostly outside of AI hubs of excellence. The regions that score highest on CEPS's AI ecosystem index — Île-de-France, London, the Randstad, Munich, Stockholm — do not, in most cases, host a factory. The regions that do host factories often lack the surrounding talent pool, startup density, and industrial absorptive capacity to translate compute access into AI capability.

Energy costs sharpen the problem. Training a frontier AI model requires sustained electrical power measured in hundreds of megawatts over weeks or months. The IEA estimates ^[10] that the most power-intensive training run in its dataset drew approximately 154 MW at peak, consuming some 310 GWh. Both CEPS and Interface ^[6][7] conclude that gigafactories should be concentrated in northern states — Sweden, Finland, Norway — where abundant renewable energy, established AI talent, and cold climates that reduce cooling costs converge. CEPS finds ^[6] that only factories in Sweden and Finland benefit from energy prices comparable to those in the United States and China.

The political logic runs in the opposite direction. Distributing sites across member states satisfies cohesion objectives. It also dilutes critical mass. CERN's founding governments faced the same tension in the 1950s and chose concentration: one site, one director-general, one scientific programme, with governance structures that gave all member states voice without fragmenting the physics. That choice is why CERN discovered the Higgs boson rather than producing thirteen national reports about why the Higgs boson had not yet been found. The gigafactories, dispersed across sixteen candidate countries, risk producing what CEPS titles its analysis ^[6]: not sanctuaries of innovation, but cathedrals in the desert.

Whose technology are they built on? The third weakness is the one the Commission discusses least. The gigafactory architecture, as currently configured, rests heavily on Nvidia's vertically integrated ecosystem. This is not merely a hardware dependency. Nvidia's CUDA software platform, its InfiniBand and NVLink interconnects, its DGX and HGX server reference designs, and its data centre layout specifications constitute a full-stack architecture. Europe's publicly funded compute installations, intended to reduce strategic dependence on American technology, are being built on an American technology stack from the chip to the cooling system.

The short-term logic is sound. No alternative supplier delivers comparable performance at comparable scale. AMD's MI300X and Intel's Gaudi 3 compete on specific workloads but lack Nvidia's ecosystem integration. European chip designers — SiPearl, which is developing processors for EuroHPC — remain years from production-ready AI accelerators at frontier scale. Europe cannot delay its compute buildout until European silicon catches up.

But the long-term trajectory matters. Nvidia has announced a roadmap to build twenty factories in Europe, including five gigafactories, in partnership with European firms. If both the public gigafactories and the Nvidia-led private installations run on the same proprietary stack, Europe's compute infrastructure becomes a distribution network for American technology rather than a platform for European innovation. CEPS recommends that the EU diversify GPU supply, prioritise open-source software architectures, and invest in alternative computing approaches that do not depend on GPUs. These are the right recommendations. They are not yet reflected in procurement strategy or selection criteria for the gigafactory programme.

3.3  The Missing Layer

The three weaknesses — demand, location, dependency — are symptoms. The underlying condition is the absence of a European research organisation with the mandate, the talent, and the institutional authority to use frontier compute for frontier science.

Consider what it takes to train a frontier AI model. Compute is the most visible requirement but only one of five. You also need architectural research — the theoretical and experimental work that determines what kind of model to build, what training methodology to use, and how to achieve maximum capability per unit of compute. You need data curation at the scale of trillions of tokens: assembly, cleaning, annotation, legal clearance, and the deliberate inclusion of multilingual and domain-specific corpora. You need alignment and safety research — the technical discipline that ensures a model behaves as intended and does not produce harmful, deceptive, or ungovernable outputs. And you need talent concentration: not a distributed network of principal investigators at twenty universities, but 200--400 researchers working under unified scientific leadership in a single institution, with the density of interaction and speed of iteration that produces breakthroughs.

The EU's current initiatives address the first requirement. They do not address the remaining four. No European institution holds the mandate to conduct architectural research at the frontier. No pan-European programme curates training data at foundation-model scale. No European organisation conducts alignment research with the resources or independence of Anthropic, DeepMind, or OpenAI's safety teams. And no mechanism exists to concentrate the researchers required for a frontier programme in a single institution operating under a single director.

This is the gap the gigafactories cannot close by getting larger. A 100,000-GPU cluster without an in-house research organisation is a data centre. An expensive one, capable of renting capacity to anyone willing to pay, but not a vehicle for sovereign AI capability. It will train Mistral's next model. It will fine-tune American models for European languages. It will run inference for commercial customers. These are useful functions. They are not what €20 billion in public investment was promised to deliver.

Europa Nova is designed to fill this structural gap — not to replace the gigafactories but to make them matter. The relationship is complementary. The gigafactories provide processing power. Europa Nova provides what the processing power is for: a concentrated, treaty-based research institution with the mandate to conduct architectural research, curate multilingual training data, develop alignment methodology, and train European foundation models at the global frontier. Equally important, Europa Nova would be the anchor customer that the gigafactories currently lack — the institution with sustained, high-volume training demand that justifies 100,000-GPU installations and gives them a clear commercial rationale.

Without such an institution, Europe's compute investment risks repeating a pattern familiar from three decades of digital economy: building infrastructure that others exploit. European motorways carry American and Asian goods. European 5G networks run American and Chinese applications. European cloud capacity hosts American platforms. The gigafactories, absent a European research organisation to anchor them, may become the compute substrate for the next generation of American AI models deployed on European soil, trained on European electricity, subsidised by European taxpayers, and governed by American corporate boards.

The structural gap between infrastructure and ambition requires a new kind of institution. Chapter 4 presents Europa Nova's design.

4 Europa Nova — Design and Structure

Every structural choice in Europa Nova's design addresses a specific failure mode observed in European technology policy. The institution is not a thought experiment. It is an engineering response to decades of evidence about what works and what does not when Europe attempts large-scale scientific and industrial projects. Where CERN succeeded, Europa Nova borrows. Where Airbus compromised, Europa Nova refuses to. Where the EU's AI Factories risk becoming what CEPS has called ^[6] "cathedrals in the desert," Europa Nova is designed to ensure that compute infrastructure serves a coherent research programme rather than existing as an end in itself.

The design rests on three pillars: a treaty-based research foundation with concentrated scientific authority, a dedicated compute infrastructure division located where physics and economics demand, and a commercial licensing entity that transforms research output into revenue and industrial capability. Each pillar has its own governance, its own operational logic, and its own success criteria. Together, they form an institution that can do what no existing European structure can: conduct frontier AI research, operate the infrastructure to support it, and deliver the results to European industry and society.

4.1  Institutional Form: Treaty-Based Foundation

Europa Nova is established by international treaty between five to eight nations, not by EU regulation. This is not a slight against the European Union. It is a recognition that the EU's legislative machinery—designed for consensus among twenty-seven member states with divergent interests and capabilities—is structurally unsuited to creating an institution that must move at the speed of frontier AI development. Qualified majority voting under Article 187 TFEU could theoretically enable a Joint Undertaking, but the political reality of assembling such a majority for a concentrated, non-distributed research institution would dilute the design beyond recognition. Treaty-based foundations avoid this.

The precedents are well established. CERN was founded in 1954 by twelve states through a convention signed under UNESCO auspices. The European Space Agency was established in 1975 by ten founding members. The European Southern Observatory was created in 1962 by just five nations—Belgium, France, Germany, the Netherlands, and Sweden. All three are treaty-based. All three achieved world leadership in their domains. The pattern is clear: a smaller coalition of willing and capable nations, bound by treaty obligations rather than regulatory compromises, can make harder choices and sustain them across political cycles.

Treaty protection provides institutional independence that no EU agency enjoys. The Scientific Director of Europa Nova does not answer to a rotating Council presidency. The research programme is not subject to amendment by agricultural or cohesion policy negotiations. The budget is a treaty obligation, not an annual appropriation vulnerable to fiscal consolidation. This insulation is not a luxury. It is a prerequisite for any institution expected to retain world-class talent in competition with organisations that can offer researchers total freedom from political interference.

The founding coalition should comprise nations that combine demonstrated AI research capability with the political will and fiscal capacity to sustain a multi-decade commitment: France, Germany, the United Kingdom, Norway, Sweden, the Netherlands, Switzerland, and Finland. Not all eight need to sign simultaneously. ESO demonstrated that five founding members can establish an institution that others join later. The critical requirement is that the founding members set the institutional culture—the DNA—before accession opens to a broader group.

One distinction from CERN must be made explicit from the outset, because it shapes every subsequent design choice. CERN produces knowledge. Its output is published papers, trained physicists, and the occasional transformative spin-off. CERN did not set out to create the World Wide Web. It emerged because CERN created the conditions for it. Europa Nova must do something harder. It must produce both knowledge and technology—and it must deliver that technology to European industry and society deliberately, not accidentally. This is why a commercial licensing arm is built into the institutional design from day one, not grafted on as an afterthought a decade later when politicians notice that public money has produced private value elsewhere.

4.2  Pillar 1: The Research Foundation

The Research Foundation is the institutional core of Europa Nova. Its mandate is fundamental research in AI architectures, alignment and safety, data curation methodology, and model development. This is not a policy research centre. It is not an ethics board. It is a laboratory where researchers build the next generation of AI systems—systems designed to be more compute-efficient, more aligned with human values, and more transparent than those produced by the current generation of US and Chinese labs.

Governance. The Foundation is governed by a Board of nine members, internationally recruited from the top tier of AI research and technology leadership. No member serves as a national representative. Selection is conducted by an independent search committee composed of senior figures from the global scientific community, not by national nomination. Participating nations have observer status at Board meetings. They receive full transparency on research directions, budget allocation, and capability assessments. They do not vote on technical decisions.

Scientific Director. The Scientific Director is the single most consequential appointment Europa Nova will make. This individual sets the research agenda, recruits the senior team, and holds veto authority on all technical decisions. The Scientific Director reports to the Board, not to any political body, any national ministry, or any intergovernmental committee. Total compensation is set at €1--2 million annually. This will provoke objections from those accustomed to public-sector salary scales. The objection is understandable and wrong. The alternative to paying market rates is not finding a cheaper candidate. The alternative is that every serious candidate takes a position at Anthropic, Google DeepMind, or OpenAI—where base compensation for research directors exceeds $1 million, with equity packages several times higher. Europa Nova cannot match equity. It can match base compensation and offer something the private sector cannot: genuine research freedom unconstrained by quarterly product cycles.

Research programme. The research agenda is defined by the Scientific Director and approved by the Board. It is not determined by national industrial policy, sectoral lobbies, or commission work programmes. This autonomy is the source of Europa Nova's scientific credibility. If the Scientific Director concludes that the most promising path to safer AGI lies in state-space models rather than Transformer architectures, the institution pivots. If alignment research requires compute resources that compete with model training, the Scientific Director allocates accordingly. Political bodies that attempt to steer the research programme through funding conditions will find that the treaty protects against precisely this.

Publication policy. Fundamental research is published openly. This is non-negotiable. Open publication is what attracts top researchers, enables global scrutiny of safety work, and establishes Europa Nova's legitimacy in the international scientific community. A security-restricted exception exists for capabilities above a defined threshold—determined by the Board on the Scientific Director's recommendation—where publication could enable dangerous applications. This threshold is calibrated, reviewed annually, and made public. The principle is stratified openness: open by default, restricted by exception.

Physical location. The research hub is located in one major European city that combines scientific excellence, access to a deep AI talent pool, and quality of life attractive to senior researchers with families. Zürich, Amsterdam, and Copenhagen are the leading candidates. The choice is made on merit, not on political negotiation between capitals. There is a maximum of one secondary research node. There is no network of distributed centres of excellence in every participating nation. The history of European collaboration is unambiguous on this point: Airbus distributes production across Toulouse, Hamburg, Broughton, and Cádiz for political reasons, and pays a permanent efficiency penalty. CERN concentrates in Geneva, and leads the world.

4.3  Pillar 2: The Compute Infrastructure Division

The separation of the research hub from the compute facility is not administrative convenience. It reflects a physical reality. Frontier AI compute requires three things in abundance: cheap electricity, cooling capacity, and space. None of these is optimally available in the European cities where researchers want to live and where talent pools are deepest. The solution is the same one that physics adopted decades ago: theorists work at universities, particle accelerators are built where the physics demands. CERN's theorists sit in offices overlooking Lake Geneva. The Large Hadron Collider runs in a tunnel one hundred metres underground, straddling the Franco-Swiss border. The locations serve different functions and need different things.

Location. Europa Nova's primary compute facility is located in Northern Scandinavia—the Narvik region of Norway or Norrbotten county in Sweden. The market has already validated this geography. In July 2025, OpenAI announced Stargate Norway ^[15][16], its first European data centre initiative, located in Kvandal outside Narvik. The project targets 100,000 NVIDIA GPUs with an initial capacity of 230MW, expandable to 520MW. OpenAI's choice was driven by the same factors that drive Europa Nova's: abundant hydropower, energy prices well below the European average, a cool climate, and established industrial infrastructure.

Energy. The facility runs on 100% renewable hydroelectric power, secured through long-term power purchase agreements. Northern Norway's electricity prices for energy-intensive industry averaged approximately 42 øre/kWh (roughly €0.04/kWh) in the third quarter of 2025. Long-term PPAs can secure rates at approximately €0.03/kWh—a fraction of the €0.10--0.15/kWh that data centres pay in Frankfurt, Amsterdam, or Dublin. The CEPS analysis ^[6] confirms that among all AI factory locations in Europe, only those in Sweden and Finland benefit from energy prices comparable to those in the United States and China. Over the operational life of a facility consuming 500MW continuously, the difference between €0.03/kWh and €0.10/kWh amounts to approximately €300 million per year. That is not a rounding error. It is the difference between financial sustainability and permanent subsidy dependency.

Cooling. The facility employs closed-loop direct-to-chip liquid cooling, supplemented by natural Arctic conditions. Ambient temperatures in Narvik regularly fall below −20°C in winter. Liquid cooling is no longer optional for frontier AI clusters—the power density of current-generation accelerators demands it—but Nordic temperatures reduce the energy required for heat rejection by 30--60% compared to temperate locations. Waste heat from the facility is integrated with local district heating networks and industrial processes, turning an environmental liability into a community asset. Stargate Norway's design incorporates the same approach.

Scale. Initial installed capacity is 500MW, designed from the outset for expansion to 2GW and beyond. To put this in context: the entire EuroHPC network of thirteen AI factories accounts for approximately 57,000 high-end AI accelerators. A single 500MW Europa Nova facility, depending on hardware generation, can support 100,000--200,000 accelerators. Meta planned to deploy infrastructure equivalent to nearly 600,000 NVIDIA H100 GPUs by the end of 2024. Mark Zuckerberg announced plans for 1.3 million GPUs by the end of 2025. Europa Nova's initial scale is not extravagant. It is the minimum required for credible frontier research.

Hardware strategy. Europa Nova adopts a multi-vendor procurement strategy from day one. The facility is designed for architectural flexibility—able to support GPUs, TPU-equivalent accelerators, and novel architectures as they emerge. This directly addresses one of the most serious risks identified in the CEPS analysis ^[6]: Europe's near-total dependency on NVIDIA for AI accelerators. NVIDIA's full-stack offering—hardware, networking, software libraries, and cloud services—creates a vendor lock-in that extends far beyond chip procurement. Europa Nova's compute division is mandated to maintain compatibility with at least two hardware vendors at all times and to invest in open-source software stacks that reduce switching costs.

Connectivity. Dedicated high-capacity fibre links connect the compute facility to the research headquarters and to major European internet exchange points in Frankfurt, Amsterdam, and London. Latency between the research hub and the compute facility is managed through network design, not proximity. Modern distributed training systems are designed for exactly this kind of separation. The link between the researcher's workstation and the training cluster does not require the researcher to sit next to the cluster, any more than a particle physicist needs to sit next to the accelerator.

Security. Norway is a NATO member state with stable democratic governance, a mature regulatory framework, and established physical security infrastructure for critical installations—a legacy of decades of North Sea energy operations. The compute facility is classified as critical national infrastructure under Norwegian law, with corresponding protections. This matters for an institution that will, over time, develop AI capabilities of strategic significance.

4.4  Pillar 3: The Commercial Licensing Entity

The Commercial Licensing Entity is what distinguishes Europa Nova from CERN and from every previous European "big science" project. It exists because AGI is not fundamental physics. The Higgs boson has no market value. A frontier AI model does. An institution that produces world-class AI research but fails to deliver that research to European industry has failed half its mission.

Ownership and governance. The entity is 100% owned by the Research Foundation. Revenue serves the mission; the mission is not distorted to serve revenue. This is a structural firewall, not a statement of principle. The entity has its own board, recruited for commercial expertise—technology licensing, enterprise sales, intellectual property management—but bound by the Foundation's charter. The charter prohibits exclusive licensing arrangements, prohibits the sale of the entity or any subsidiary, and requires that a minimum of 70% of net licensing revenue be returned to the Foundation for reinvestment in research.

Business model. The entity licenses Europa Nova's models, tools, fine-tuning capabilities, and compute access to European and global enterprises. Pricing follows a tiered structure: preferential rates for European SMEs, public-sector institutions, and academic researchers; standard commercial rates for large European enterprises; and premium rates for non-European corporations. This is not protectionism. It is the natural pricing structure of any institution whose costs are borne primarily by European taxpayers and whose primary obligation is to European competitiveness.

Revenue trajectory. Conservative projections estimate €500 million in annual licensing revenue by year five, rising to €2--3 billion by years eight to ten. These figures are not aspirational. They are derived from observable market economics. Microsoft alone spent $80 billion on AI infrastructure in fiscal year 2025 ^[18]. The combined AI capital expenditure of Alphabet, Meta, Microsoft, and Amazon exceeded $380 billion in calendar year 2025 ^[18]. Any institution that produces frontier AI models and infrastructure at scale will find a market. The question is whether Europa Nova captures European demand that currently flows to American providers, or whether European enterprises continue to pay for AI systems designed in California and governed by American corporate priorities.

Precedents. Three models inform the design. ARM Holdings designs chip architectures and licenses them globally—ARM does not manufacture chips, but its designs power virtually every smartphone on Earth. The Fraunhofer Institutes operate Europe's largest application-oriented research organisation, with an annual budget of €3.6 billion, of which €867 million comes directly from industry contracts and €164 million from licensing and intellectual property sales. And CERN itself, while not commercially oriented, produced the World Wide Web—arguably the most economically significant technology transfer in history—as an incidental by-product of its research environment. Europa Nova takes the Fraunhofer principle of market-driven research transfer and applies it at the scale and ambition of CERN.

Strategic purpose. The licensing entity resolves a tension that has defined—and damaged—the most prominent AI organisation of the past decade. OpenAI was founded as a non-profit research lab. It became a capped-profit corporation. It is now restructuring as a full for-profit entity. At each transition, the mission shifted. Europa Nova's structure makes this trajectory impossible. The Foundation owns the entity. The entity cannot acquire the Foundation. The charter is protected by international treaty, not by corporate bylaws amendable by a board vote. Revenue sustains the mission. The mission generates the technology. The technology generates the revenue. The cycle is designed to be self-reinforcing and structurally stable.

4.5  What Europa Nova Is Not

Misunderstanding the proposal is the fastest route to opposing it. Several potential misreadings must be addressed directly, because each represents an objection that will be raised in policy debate.

Europa Nova is not a competitor to EU AI Factories or Gigafactories. The Commission's AI factories provide compute infrastructure. That infrastructure is necessary. But compute without a coherent research organisation to direct it, without architectural ambition beyond running NVIDIA's software stack, and without a governance structure that can make hard scientific choices, is an expensive commodity. The CEPS analysis ^[6] is blunt on this point: the factories are being built mostly outside of Europe's AI hubs of excellence, in locations that do not cooperate with each other, and with insufficient attention to energy costs or vendor diversification. Europa Nova is the organisational and research layer that gives compute infrastructure strategic purpose. It complements the factories. It does not replace them.

Europa Nova is not a regulatory body. That function belongs to the AI Act and its implementing authorities. Europa Nova engages with regulation as a subject, not an author. Its researchers comply with applicable law. Its commercial entity operates within the regulatory framework. Its safety research informs regulatory decisions by providing the technical evidence that regulators need. But it does not set rules for others. The independence required for frontier research is incompatible with regulatory authority.

Europa Nova is not a European Google or OpenAI. It is not a technology company. It does not ship products to consumers. It does not run a chatbot or a search engine. It is a research institution with a commercial arm—closer in concept to a Max Planck Institute or a Fraunhofer Institute, but at unprecedented scale and with a mandate that extends from fundamental research to commercial licensing. The distinction matters because technology companies are governed by market incentives. Research institutions are governed by their charter and their scientific culture. Europa Nova is designed to be the latter.

Europa Nova is not a distributed network. It is concentrated by design. This will be the single most politically difficult aspect of the proposal, because European institutional culture defaults to distribution. Every nation wants a share. Every capital wants a centre. The pressure to create a "node" in each participating country will be immense. It must be resisted absolutely. Frontier AI research is produced by concentrated teams with shared infrastructure, not by networks of loosely connected groups across twelve time zones. Google DeepMind is in London. OpenAI is in San Francisco. Anthropic is in San Francisco. None of them operates as a distributed network, because distribution is incompatible with the speed and integration that frontier research demands.

Europa Nova is not a short-term project. The treaty establishes a minimum ten-year operational horizon. A rigorous evaluation at year three determines whether the institution is on a credible trajectory—if it is not, the treaty includes provisions for orderly wind-down. This is not a concession to sceptics. It is a commitment to excellence over permanence. But if the evaluation is positive, Europa Nova operates on the time horizon that matters: the decade or more required to build an institution capable of sustained frontier research. The alternative—three-year funding cycles, annual reporting to twenty-seven national capitals, perpetual justification of existence—is a recipe for mediocrity disguised as accountability.

The design presented in this chapter is detailed because detail is where institutional proposals succeed or fail. A vague commitment to "European AI excellence" costs nothing and achieves nothing. A treaty-based foundation with a specific governance structure, a defined location strategy, a commercial arm with explicit revenue obligations, and a willingness to shut down if results do not materialise—that is a commitment that can be assessed, funded, and held to account. The next chapter examines what it will cost and how it can be financed.

5 Funding: Ambitious but Realistic

5.1  Total Investment Requirement

Europa Nova requires a committed investment of €50 billion over ten years, averaging €5 billion annually. This is a large number. It is meant to be. The development of safe artificial general intelligence is the most capital-intensive scientific undertaking since the Manhattan Project, and unlike that programme, Europa Nova is designed to generate commercial returns that progressively reduce the burden on public treasuries.

The number must be placed in context. OpenAI's Stargate initiative represents $500 billion in planned US AI infrastructure investment ^[17]. Microsoft alone announced $80 billion in AI data centre capital expenditure for its fiscal year ending June 2025 ^[18]. In their most recent earnings guidance, Alphabet, Meta, Microsoft, and Amazon collectively projected capital expenditure exceeding $380 billion for the current fiscal year ^[18]. The EU's own InvestAI initiative ^[1] aims to mobilise €200 billion, of which €20 billion is public funding earmarked for AI gigafactories. Against this backdrop, €50 billion over a decade is not extravagant. It is the minimum credible stake in a game where the buy-in rises every quarter.

Two institutional comparisons sharpen the picture. CERN operates on an annual budget of approximately CHF 1.2 billion. ESA's annual budget stands at €7.68 billion. Europa Nova at €5 billion per year sits between these two institutions in scale — roughly four times CERN, roughly two-thirds of ESA. The comparison is instructive. ESA is a mature operational agency with launch vehicles, satellite programmes, and ground stations across three continents. CERN is a pure research institution with no commercial mandate. Europa Nova occupies a novel position: a research institution with commercial revenue that, at maturity, could approach ESA-scale operations funded substantially from its own income.

The €50 billion breaks down across four categories. Approximately 40 percent — €20 billion — is allocated to compute infrastructure: construction of the primary facility in Northern Scandinavia, hardware procurement, energy contracts, cooling systems, networking, and rolling maintenance. This is the physical foundation. Roughly 35 percent — €17.5 billion — funds research operations: personnel costs for 400--800 researchers and support staff at market-competitive compensation, laboratory equipment, computing time for experiments, and operational overhead. Fifteen percent — €7.5 billion — supports the Commercial Licensing Entity: product development, go-to-market operations, customer infrastructure, legal and regulatory compliance. The remaining 10 percent — €5 billion — constitutes a strategic reserve for technology pivots, currency hedging, cost overruns, and unforeseen opportunities.

The reserve is not padding. It is an explicit acknowledgement that ten-year cost estimates in a field where hardware generations turn over every three years carry inherent uncertainty. Any funding model that claims precision to the last million is either dishonest or naïve. Europa Nova's budget is designed to be robust, not optimistic.

5.2  Funding Sources

Europa Nova draws on four distinct funding streams, each with a different risk profile, time horizon, and governance structure. The design is deliberate: no single stream bears the full weight, and the mix shifts over time from predominantly public to increasingly commercial.

A. Sovereign Treaty Contributions: approximately €3 billion per year

The foundational funding comes from treaty obligations of the 5--8 participating nations. This is the structural innovation that separates Europa Nova from EU programmes dependent on multiannual financial frameworks and annual budget negotiations. A treaty contribution is a sovereign commitment, ratified by national parliaments, with the predictability that frontier research demands. CERN has operated on this model since 1954. The principle is proven.

Contributions are distributed according to GDP and anticipated compute utilisation, with the precise formula to be negotiated during the treaty process. An illustrative allocation across the founding coalition: Germany at €800 million per year, France at €600 million, the United Kingdom at €500 million, Norway at €400 million, Sweden at €250 million, the Netherlands at €200 million, Switzerland at €150 million, and Finland at €100 million. These are indicative figures, not demands. They establish the order of magnitude.

The proportions are modest relative to national budgets. Norway's contribution of €400 million per year equals approximately 0.08 percent of GDP — comparable to its annual EEA financial mechanism contribution. Germany's €800 million amounts to less than 0.2 percent of federal expenditure, roughly what it spends annually on the Helmholtz Association of research centres. France's €600 million is less than its annual contribution to ESA. For every participating nation, the Europa Nova commitment falls within established precedent for large-scale scientific infrastructure.

The treaty structure carries a second advantage beyond predictability: it creates a credible multi-year commitment that enables Europa Nova to sign long-term power purchase agreements, hardware procurement contracts, and employment packages that would be impossible under annual funding cycles. The inability to offer multi-year certainty is precisely why EU-funded research programmes struggle to recruit top-tier talent from industry labs that offer equity and long-term compensation.

B. Norwegian Sovereign Wealth Fund Allocation: €15–30 billion over ten years

Norway's Government Pension Fund Global — the world's largest sovereign wealth fund — held assets exceeding $2.2 trillion at the end of 2025, having posted a record return of 15.1 percent for the year. A dedicated allocation of 1--2 percent to Europa Nova infrastructure, structured as an investment with expected commercial return, would provide €15--30 billion in anchor capital. This is substantial but not unprecedented for a fund of this scale. The fund already invests in unlisted renewable energy infrastructure and real estate across Europe and globally. AI compute infrastructure in Norway, powered by Norwegian hydropower, generating licensing revenue from a European customer base, represents a logical extension of the fund's existing investment mandate.

This is the most politically sensitive element of the funding proposal, and it deserves direct treatment. The fund's mandate, set by the Norwegian parliament, is to generate financial returns for future generations. It is not a development fund. It is not a strategic investment vehicle. The obvious objection is that a single-institution allocation of this size concentrates risk in a way the fund's diversification mandate is designed to prevent.

Three responses. First, the investment is in physical infrastructure — data centres, power connections, cooling systems, fibre networks — with residual value regardless of Europa Nova's research outcomes. If the programme underperforms, the infrastructure functions as European data centre capacity and can be sold, leased, or repurposed. This is not a venture capital bet that goes to zero on failure. Second, the Commercial Licensing Entity provides a return mechanism: revenues from model licensing, compute-as-a-service, and technology transfer flow back to the fund on commercial terms. Third, the strategic logic is compelling. Norway's wealth derives from fossil fuels. The transition to an AI-driven economy will erode the value of hydrocarbon assets over the coming decades. Converting a small fraction of fossil-fuel wealth into AI infrastructure represents exactly the kind of intergenerational value transfer the fund was created to facilitate.

The political path requires a parliamentary mandate to expand the fund's investment universe to include strategic European AI infrastructure. This is a significant political step. It is also a step that positions Norway at the centre of Europe's most consequential technology programme, with governance influence proportionate to its financial commitment.

C. Commercial Revenue: growing from year 4–5

Europa Nova's Commercial Licensing Entity is projected to generate €500 million in annual revenue by year 5, rising to €2--3 billion by years 8--10. These projections are conservative relative to the commercial AI market's growth trajectory, but they are intended to be. Revenue streams include four categories: licensing of Europa Nova foundation models and specialised models to European enterprises and public institutions; compute-as-a-service for organisations that need frontier-scale processing without building their own infrastructure; technology transfer agreements with industry partners; and consulting and integration services for governments and regulated industries.

The target market is large. Europe's 23 million SMEs represent the world's largest pool of potential AI adopters that currently lack access to domestically-developed frontier models. European public sector institutions — healthcare systems, tax administrations, judicial systems, defence ministries — face growing pressure to deploy AI while maintaining data sovereignty. The Interface and Bertelsmann Stiftung analysis ^[8] of EU AI gigafactories identifies precisely this demand gap: a large, heterogeneous base of users with moderate AI workloads that neither hyperscalers nor neoclouds are well-positioned to serve. Europa Nova's advantage is that it produces models and tools, not merely raw compute. It sits higher on the value chain, with a differentiated offering that does not compete on GPU-hour pricing.

At full maturity, commercial revenue could make Europa Nova financially self-sustaining, progressively reducing the need for sovereign contributions. This is the trajectory that transforms Europa Nova from a cost centre to an asset. ESA's recent ministerial in Bremen secured €22.3 billion in commitments over three years precisely because member states recognised the strategic return on space investment. Europa Nova's commercial arm makes the return explicit and quantifiable.

D. European Investment Bank Co-Financing: €3–5 billion for physical infrastructure

In December 2025, the European Investment Bank signed a Memorandum of Understanding ^[5] with the European Commission to provide advisory support and potential financing for AI gigafactory projects. The EIB's flagship TechEU Programme commits €70 billion over 2025--2027 for disruptive technologies and enabling infrastructure, with a target of mobilising €250 billion alongside financing partners. Europa Nova's compute infrastructure — data centres, power connections, fibre links, cooling systems — falls squarely within the EIB's existing investment frameworks.

EIB financing would be structured as loans, not grants, repaid from commercial revenue. This is the model the EIB has applied successfully to major European infrastructure: Galileo navigation, Trans-European Transport Networks, offshore wind farms. The principal advantage is cost of capital: EIB loans carry interest rates below commercial markets because of the EU's collective credit standing. For infrastructure with a 20--30 year useful life, the difference between EIB financing and commercial rates represents hundreds of millions in savings.

Europa Nova's infrastructure qualifies under the InvestAI framework, which earmarks €20 billion in public funding for up to five AI gigafactories. The important distinction: Europa Nova is not an AI gigafactory. It is a research institution that operates AI gigafactory-scale infrastructure. The infrastructure qualifies for EIB co-financing; the research programme does not and should not. Keeping these funding streams distinct preserves Europa Nova's scientific independence while capturing the financial benefits of EU infrastructure policy.

5.3  Risk Analysis

A funding proposal of this scale without an honest risk assessment would not be credible. Five categories of risk require direct treatment.

Termination risk. The Europa Nova treaty includes a year-3 evaluation gate with an independent external panel empowered to recommend continuation or orderly wind-down. If Europa Nova is terminated at year 3, total expenditure is approximately €12--15 billion. This is a significant sum. It is not, however, a sunk cost in the way that failed software ventures are. The compute infrastructure — data centres, power connections, cooling systems — retains value as European data centre capacity. Northern Scandinavia's energy and cooling advantages ensure strong demand from commercial operators. The infrastructure can be sold, leased, or transferred to another European programme. The worst-case scenario is not €15 billion lost; it is €15 billion partially recoverable through asset disposal and a hard lesson learned about what it takes to build an AGI research institution.

Hardware depreciation. GPU generations have approximately three-year useful lifespans for frontier training workloads. The compute budget must reflect rolling renewal, not one-time procurement. Europa Nova's €20 billion infrastructure allocation assumes three full hardware cycles over the ten-year programme, with each successive generation delivering substantially more performance per watt. This is built into the cost model. Older-generation hardware retains value for inference workloads and can be redeployed to the commercial licensing operation or sold to European research institutions at reduced cost. Hardware depreciation is a managed cost, not a risk — provided the budget acknowledges it from the outset.

Technology pivot risk. What if the Transformer architecture proves sufficient and the architectural innovation thesis that partly motivates Europa Nova proves wrong? This is a real possibility, and it would change the research programme significantly. It would not, however, invalidate the institution. The compute infrastructure serves any architecture. The research team would pivot from architectural innovation to training methodology, alignment research, or domain-specific model development — all areas where Europe has legitimate competitive advantages. The risk is in the specific research bet, not in the institutional design. Europa Nova's governance structure, with its empowered Scientific Director and year-3 evaluation gate, provides the mechanism for course correction.

Currency and energy price risk. Compute hardware is priced overwhelmingly in US dollars. A ten-year European programme denominated in euros faces real currency exposure. The strategic reserve includes an allocation for currency hedging, and large-scale procurement contracts can be structured with currency provisions, but the risk is inherent and must be managed actively. Energy price risk is lower: long-term power purchase agreements with Nordic hydropower producers lock in rates at approximately €0.03/kWh for 10--15 year terms. The stability of Nordic hydropower costs is one of Europa Nova's structural advantages and a primary reason for the Scandinavian location.

Demand risk for commercial licensing. The revenue projections assume growing European demand for domestically-developed AI models and compute services. The Hess and Sieker analysis ^[8] of EU AI gigafactories raises a legitimate concern about European AI compute demand: without anchor customers consuming large training workloads, large-scale facilities risk underutilisation. Europa Nova's position is different. As a research institution producing its own frontier models, Europa Nova is its own anchor customer for training compute. The commercial licensing operation targets a different segment: the vast European market of moderate-demand users — SMEs, public institutions, research organisations — that need models and services, not raw GPU hours. This is higher on the value chain and more defensible than competing with neoclouds on compute pricing. The risk is not zero. If European AI adoption lags global trends, commercial revenue will fall short of projections. The treaty contribution structure ensures Europa Nova survives a revenue shortfall; it does not depend on commercial income for its first five years.

The funding model is designed for resilience, not optimism. Treaty contributions provide a stable foundation. The sovereign wealth fund allocation provides scale. Commercial revenue provides a path to self-sufficiency. EIB co-financing reduces the cost of physical infrastructure. No single stream is essential beyond the first five years; together, they create a funding structure that can absorb shocks, adapt to changing circumstances, and sustain a programme ambitious enough to matter. The question is not whether €50 billion is a lot of money. It is. The question is whether it is proportionate to the stakes. When a single American technology company spends €80 billion in a single year on AI infrastructure, a ten-year European commitment of €50 billion — shared across eight nations, backed by the world's largest sovereign wealth fund, generating its own commercial revenue — is not an act of extravagance. It is the price of remaining at the table.

6 Governance: Designed for Excellence, Not Compromise

The single greatest risk to Europa Nova is not technical. It is not a shortage of compute, a failure to recruit, or an unforeseen breakthrough by a competitor. The single greatest risk is that the institution is designed, like so many European initiatives before it, to satisfy political constituencies rather than to produce results. Governance is the difference between CERN and Galileo, between an institution that wins Nobel Prizes and one that misses deadlines by a decade. Every governance choice in this chapter is a direct response to a specific, documented failure in European large-scale collaboration.

6.1  Lessons from European Institutional History

Europe has built institutions that changed the world. It has also built institutions that consumed billions and delivered little. The difference between these outcomes is not talent, funding, or ambition. It is governance design. Four cases illustrate the pattern with precision.

CERN: What concentrated authority produces

CERN is the most successful international research institution in history. Established by treaty in 1954, it has produced twelve Nobel Prizes, the World Wide Web, and the confirmation of the Higgs boson.

Three structural features explain this record. First, concentration: CERN operates from a single site straddling the Franco-Swiss border near Geneva. There are no satellite centres distributed across member states to satisfy national pride. The particle physics is done where the accelerators are. Second, scientific authority: the Director-General controls the research programme, and the CERN Council—composed of national representatives—approves budgets but does not direct experiments. Third, long-term commitment: member states contribute on a treaty basis, insulating the budget from annual political cycles. The Large Hadron Collider took twenty years from conception to first beam. No annual appropriations process would have sustained that timeline.

The result is an institution where a Greek postdoctoral researcher and a Swedish engineer work on the same detector, and neither government demands that the detector be partially assembled on its soil. This is the model Europa Nova must replicate.

Airbus: The cost of juste retour

Airbus is a technical success. It is Europe's only credible competitor to Boeing in commercial aviation, and the A320 family is the best-selling aircraft in history. But Airbus was built under the principle of juste retour—the requirement that each participating nation receive industrial work proportional to its financial contribution. The consequences were predictable.

Wings are manufactured in the United Kingdom, fuselage sections in Germany, final assembly in France, and tail units in Spain. These decisions were driven not by manufacturing logic but by political negotiation. The A380 programme alone required over fifty formal workshare renegotiations between 2000 and 2006. Every time a new variant is proposed, the industrial geography must be renegotiated. This adds cost, introduces delay, and forces suboptimal engineering choices.

Airbus succeeded despite juste retour, not because of it—and it succeeded in a market where development cycles are measured in decades and customers are patient. AI operates on a different clock. A governance structure that adds even twelve months of negotiation to a research programme will be fatally slow. Europa Nova explicitly rejects the juste retour principle. Participating nations contribute according to capacity. They receive the strategic benefit of European sovereign AI capability. They do not receive guaranteed contracts or facility placements.

Galileo: What ambiguity destroys

The Galileo satellite navigation programme is the textbook case of governance ambiguity killing a technology project. Launched in 2003 as a public-private partnership, Galileo was meant to give Europe independence from the American GPS system. The original concession was awarded to a private consortium in 2005. By 2007, the consortium had collapsed, unable to agree on risk allocation, revenue sharing, or technical authority. The project reverted to full public funding and was placed under direct European Commission management.

The European Court of Auditors documented the result: massive cost overruns, a decade of delays, and a system that did not reach initial operational capability until 2016—thirteen years after launch. The core problem was not technical. European industry had the engineering capacity to build a satellite navigation system. The problem was that no one had clear authority. The PPP structure created a governance vacuum in which every decision required negotiation between the Commission, ESA, the concessionaire, and member state governments. When the concessionaire failed, there was no fallback plan because no single entity owned the project.

The lesson for Europa Nova is unambiguous: every function must have a single point of authority. The Scientific Director controls the research programme. The Board controls the institution. The Advisory Council provides oversight. There is no shared authority, no co-management, and no governance structure that requires consensus among entities with different objectives.

EuroHPC and the AI Factories: The limits of distributed infrastructure

The EU's AI Factories initiative represents the most recent attempt to build European AI infrastructure at scale. The European Commission has selected thirteen AI Factories across the continent, each offering publicly subsidised compute to researchers, startups, and SMEs. A further round of AI Gigafactories—each with at least 100,000 advanced chips—was announced in early 2025 as part of the AI Continent Action Plan.

Two independent analyses—one by Interface, one by CEPS ^[7][6]—have identified structural problems. The Interface analysis ^[7] of all thirteen factories found that partnership consortia are dominated by research institutions rather than commercial actors, with only two of thirteen lacking a university or academic partner. Nine of thirteen factories target at least five sectors, revealing a generalist approach that rarely reflects regional industrial strengths. The CEPS analysis ^[6] goes further: AI factories are geographically dispersed across multiple EU countries, while leading European AI hubs are concentrated in a handful of regions—Inner London, Île-de-France, Oberbayern, Noord-Brabant, and Stockholm. The overlap between factory locations and genuine AI hubs is minimal.

This is not a criticism of the individuals running these programmes. It is a structural observation. When infrastructure is distributed across member states to ensure political balance, when consortia must include academic and public-sector partners to qualify for funding, and when no anchor commercial customer drives utilisation, the result is predictable: publicly funded compute that serves researchers well but does not produce the concentrated, commercially oriented research environment that frontier AI requires.

The Interface report ^[7] states the conclusion directly: AI factories provide useful public research infrastructure, but they are not sufficient to drive commercial AI innovation across Europe at scale. The CEPS analysis ^[6] warns that broad geographical dispersion risks creating not sanctuaries of innovation but "cathedrals in the desert"—impressive facilities in regions that lack the surrounding talent, capital, and industrial base to make them productive.

Europa Nova is designed as the structural antithesis of this approach. One research hub. One compute facility. Researchers selected by merit, not by nationality. Commercial licensing from day one. The AI Factories serve a legitimate purpose—democratising access to compute for European researchers and startups. But they are not, and cannot be, a vehicle for frontier AGI research. That requires a fundamentally different institutional design.

6.2  Europa Nova's Six Governance Principles

Each governance principle below responds to a specific, identified failure mode. These are not aspirational statements. They are engineering decisions.

Principle 1: Concentration over distribution.

Distributed research produces distributed mediocrity. This is the central finding of both the CEPS and Interface analyses of European AI infrastructure, and it is confirmed by seventy years of evidence from physics, biotechnology, and semiconductor research. Critical mass—the density of talent interacting daily in the same corridors, the same canteens, the same late-night debugging sessions—is not a luxury. It is a precondition for frontier research.

Europa Nova operates from one primary research hub and one primary compute facility. There are no "centres of excellence" distributed across participating nations. There is no secondary campus established to satisfy a government that contributed heavily to the treaty fund. Resources go where the physics—or in this case, the mathematics—demands. CERN works because it is in Geneva. Europa Nova works because it is concentrated.

Principle 2: Scientific autonomy.

When political bodies direct research priorities, the result is invariably safe, incremental work that offends no constituency and threatens no incumbent. The Scientific Director and the nine-member Board control Europa Nova's research programme. Participating nations have observer status at Board meetings and receive full briefings on research direction, but they hold no vote on technical decisions. This mirrors the CERN Convention, which vests the Director-General with authority over the scientific programme subject only to the Council's budgetary approval.

The obvious objection is democratic accountability: public money should be subject to public oversight. This is correct, and Section 6.2's Principle 6 addresses it directly. But oversight and control are different things. The public has a right to know what Europa Nova is doing and why. The public does not have the competence—through elected representatives—to decide whether to pursue state-space models or hybrid attention mechanisms. Every breakthrough institution in history—Bell Labs, Xerox PARC, CERN, the original DARPA programmes—gave researchers genuine autonomy from their funders. Europa Nova does the same.

Principle 3: Market-rate compensation.

Europe produces world-class AI researchers and then loses them. The Global AI Talent Tracker ^[24] reports that 57% of the world's top AI researchers work in the United States; France, Germany, and the United Kingdom combined account for 16%. This is not a patriotism gap. It is a compensation and freedom gap. A senior research scientist at Google DeepMind or Anthropic earns between $500,000 and $3 million in total compensation. A full professor at ETH Zürich—one of Europe's finest institutions—earns approximately CHF 220,000.

Europa Nova's Foundation structure—distinct from any national public agency—enables compensation at €500,000 to €2 million for senior researchers. This will provoke objections. Public-sector salary norms in most European countries cap academic and research compensation well below these levels. The response is straightforward: Europa Nova is not a public agency. It is a treaty-based foundation, and its compensation structure reflects the market in which it competes for talent. The alternative is not lower salaries. The alternative is that every viable candidate accepts an offer from San Francisco, and Europa Nova is staffed with the researchers who could not get those offers. That is a recipe for a second-tier institution, and a second-tier institution is not worth €50 billion.

Principle 4: Built-in termination.

Failed public projects persist because no one has the authority or the political incentive to end them. Galileo consumed thirteen years and billions of euros before delivering a minimally functional system, yet at no point did any authority have the mandate to terminate. The sunk-cost fallacy operates with particular force in public institutions, where admitting failure carries personal and political consequences.

Europa Nova's treaty includes an explicit sunset mechanism. At year three, an independent external panel—composed of recognised AI researchers with no institutional affiliation to Europa Nova—conducts a comprehensive evaluation. If the panel concludes that Europa Nova is not on a credible trajectory toward frontier results, the treaty provides for orderly wind-down: asset disposition, researcher transition support, and return of uncommitted funds to participating nations. No extensions are granted on political grounds. No "further study" is commissioned to delay the decision.

This is the mechanism that venture capital uses to discipline investment. A startup that has not demonstrated product-market fit by Series B does not receive Series C funding. Public institutions rarely adopt this discipline, which is precisely why they rarely achieve venture-capital-level results. Europa Nova is designed to succeed or to fail fast. It is not designed to persist indefinitely in a state of comfortable mediocrity.

Principle 5: No juste retour.

Participating nations contribute to Europa Nova on a capacity basis—a formula weighted by GDP and national wealth, comparable to existing treaty-based contribution models at CERN and ESA. In return, they receive the strategic benefit of sovereign European AGI capability, preferential licensing terms for their public sectors, and the prestige of participation in a world-class institution.

What they do not receive is a guaranteed proportional return in jobs, contracts, or infrastructure placement. If the best location for the compute facility is Northern Norway, it is built in Northern Norway—regardless of whether Norway's financial contribution is 8% or 18% of the total. If the strongest candidates for a research division are disproportionately French and Dutch, they are hired—regardless of whether France and the Netherlands have contributed proportionally more or less than Germany.

The Airbus precedent is instructive in both directions. Juste retour ensured that every participating nation had skin in the game and a tangible industrial benefit, which sustained political support for decades. But it also produced an industrial geography that added an estimated 10--15% to production costs and made every programme decision a diplomatic negotiation. In AI, where the competitive cycle is measured in months rather than decades, this overhead is not a nuisance. It is fatal. Europa Nova opts for the CERN model: nations contribute because the mission matters, not because they receive a proportional share of the factory floor.

Principle 6: Transparent democratic oversight without operational control.

Democratic accountability is non-negotiable. An institution that spends €5 billion per year of public money and develops technology with profound societal implications must answer to elected representatives. But democratic accountability and democratic management are different things, and conflating them has destroyed more European technology projects than any technical failure.

Europa Nova establishes an Advisory Council composed of parliamentarians from each participating nation. The Council receives full transparency: quarterly briefings on research progress, annual capability assessments, real-time access to safety and alignment decisions, and the right to question the Scientific Director and Board in formal sessions. These sessions are public. The Council has the right to information, the right to question, and the right to publish its assessment. It does not have decision authority over research direction, personnel, or operations.

The model is central bank independence, adapted for a research institution. Elected officials set the mandate: develop safe artificial general intelligence that reflects European values. Experts execute it. The Maastricht Treaty established this principle for monetary policy in 1992, and it has proven remarkably durable. The same logic applies with equal force to an institution whose decisions are, if anything, more technically complex than interest-rate setting.

6.3  Leadership Recruitment

Everything described above depends on one appointment. The Scientific Director of Europa Nova sets the research vision, recruits the senior team, and establishes the institutional culture that will determine whether €50 billion produces a world-class research institution or an expensive bureaucracy. This is a Demis Hassabis-level appointment—a person who has published at the highest level, who has built and led a research organisation operating at scale, and who understands both the science of AI and the engineering of its industrialisation.

The profile is specific. The Scientific Director must have a publication record in top-tier AI venues—NeurIPS, ICML, ICLR, Nature, or Science—that demonstrates genuine research contribution, not merely managerial oversight of others' work. The Director must have led a team of at least fifty researchers producing original work, because managing a laboratory of five hundred begins with knowing how to manage fifty. And the Director must understand the path from research to product, because Europa Nova's commercial licensing arm requires that research be designed from the outset with deployment in mind.

The recruitment process must be insulated from national politics. An international search committee—composed of five to seven recognised figures in AI research, drawn from both academia and industry, with no more than one member from any single participating nation—conducts a global search. The committee's composition matters as much as the candidate it selects: when the appointment is announced, the names on the search committee must lend immediate credibility. If the committee includes a former director of a leading AI laboratory, a Fields Medal or Turing Award recipient, and a senior figure from European science policy, the appointment carries authority from day one.

Compensation for the Scientific Director must be set at €1--2 million in total annual compensation, including base salary, performance-linked incentives, and a retention package structured over the initial five-year term. This will attract criticism from those who compare it to prime ministerial salaries. The comparison is misplaced. The relevant comparison is what Google DeepMind, OpenAI, and Anthropic pay for equivalent talent. At those organisations, total compensation for a research director exceeds $2 million, and for a CEO or chief scientist, it exceeds $10 million. Europa Nova is not competing with the civil service. It is competing with the most generously compensated research organisations in the world.

The treaty should specify that the Scientific Director is appointed for a single five-year term, renewable once by the Board on the recommendation of the external evaluation panel. A single term is too short to build an institution; unlimited terms risk the stagnation that comes with entrenchment. Two terms—ten years—is the correct horizon. It is long enough to set a research direction, recruit a team, and see the first results through to maturity. It is short enough to ensure accountability and renewal.

Getting this appointment right is not one priority among many. It is the priority. A mediocre Scientific Director with perfect governance will produce a mediocre institution. An exceptional Scientific Director with imperfect governance will find ways to make it work. The history of CERN, of Bell Labs, of every institution that has produced transformative research confirms this: institutions are made by the people who lead them, and governance exists to give those people the authority and resources to do their work.

7 Technical Strategy: Architecture, Data, and Alignment

Europa Nova is not a compute procurement programme. It is a research institution. The distinction matters because it determines what kind of intelligence Europe can realistically hope to build. The EU's existing AI Factory and Gigafactory initiatives address a genuine infrastructure deficit, but infrastructure without a research agenda is an expensive commodity service. Europa Nova's technical strategy rests on three pillars: architectural innovation that changes the economics of AI, a data strategy that converts European diversity into a training advantage, and an alignment programme rooted in European philosophical traditions. Each is designed to produce capability that cannot be replicated simply by spending more money on the same approach.

7.1  Architecture Strategy: Beyond Transformers

The Transformer architecture, introduced by Vaswani and colleagues at Google in 2017 ^[19], has dominated AI for nearly a decade. Every frontier model from OpenAI, Anthropic, Google DeepMind, and Meta is built on some variant of this design. The Transformer's self-attention mechanism is powerful, but it comes with a fundamental constraint: computational cost scales quadratically with sequence length. Doubling the context a model can process roughly quadruples the compute required. This is the scaling law that has driven the arms race in GPU procurement — and it is the race Europa Nova should not attempt to win on the same terms.

The reasoning is straightforward. OpenAI, Google, and Meta each spent more than $50 billion on AI infrastructure in 2024. Europa Nova's total budget of €50 billion over ten years cannot match a single year of spending by any one of these companies. Competing on scale within an identical architecture means accepting permanent inferiority. The alternative is to invest in architectural research that changes the scaling laws themselves — making each unit of compute yield more capability.

This is not speculative. The research frontier is active and moving fast. State-space models, particularly the S4 and Mamba families developed by Gu and Dao ^[20][21], achieve linear-time sequence modelling — processing information in time proportional to sequence length rather than the square of it. The IEA's 2025 analysis ^[10] of AI energy consumption found that Mixture-of-Experts architectures already reduce electricity consumption by roughly 40% compared to dense models of equivalent parameter count, because they selectively activate only the relevant portion of the model for each input. Hybrid architectures that combine selective attention with SSM efficiency are being explored by multiple research groups. Jamba, developed by AI21 Labs, demonstrated that merging Transformer layers with Mamba components reduces key-value cache memory by an order of magnitude while maintaining performance.

The research directions relevant to Europa Nova's mandate include five distinct lines of investigation. State-space models offer the most mature alternative to Transformer attention, with demonstrated advantages in long-sequence processing and inference efficiency. Hybrid architectures explore how to combine the strengths of attention mechanisms with the computational efficiency of recurrent and state-space approaches. Mixture-of-experts at unprecedented scale — activating perhaps 20 billion parameters from a model containing 1 trillion — could deliver frontier capability at a fraction of the energy cost. Neuromorphic and event-driven computing, which processes information through spike-based signals rather than continuous matrix multiplication, represents a fundamentally different compute paradigm with potentially transformative energy efficiency. Continuous-time models, which process information as flows rather than discrete tokens, may prove better suited to modelling physical systems and temporal reasoning.

Europe has genuine research strength in several of these areas. ETH Zürich's AI Center and EPFL's Machine Learning Laboratory have produced foundational work on efficient architectures and learning theory. INRIA's research teams in Paris and Grenoble have contributed to neural ODE and continuous-time modelling. The Max Planck Institute for Intelligent Systems in Tübingen is a global leader in probabilistic machine learning and model efficiency. The University of Edinburgh's School of Informatics and Cambridge's Machine Learning Group are among the strongest in Europe for theoretical AI research. These groups collectively publish at a rate competitive with any American institution — but they lack the compute infrastructure to validate their architectural ideas at scale.

This is precisely where Europa Nova fills a gap that the EU's current initiatives do not. The CEPS analysis ^[6] of AI Factories and Gigafactories, published in October 2025, identified a structural problem: the Gigafactory model rests heavily on Nvidia's vertically integrated stack, from GPUs through CUDA software to cloud orchestration layers. Andrea Renda and Nicoleta Kyosovska at CEPS ^[6] warned that this dependency could constrain which types of AI Europe can build — if the infrastructure is optimised exclusively for Transformer training on Nvidia hardware, alternative architectures that might run better on different hardware are effectively foreclosed.

Europa Nova's compute infrastructure must therefore be multi-vendor and architecture-flexible from the outset. This is a design requirement, not an aspiration. The compute facility in Northern Scandinavia should support Nvidia GPUs for near-term workloads, but its interconnect and orchestration layers must be designed to incorporate AMD accelerators, custom ASICs, and novel hardware — including neuromorphic processors like Intel's Loihi 2 or the European SpiNNaker2 system developed at TU Dresden — as they mature. The CEPS analysis ^[6] noted that even among existing EU Factories, incompatible hardware-software stacks create interoperability problems: developers building on Finland's LUMI system (which uses AMD ROCm) face substantial difficulty porting code to Nvidia-based systems elsewhere in Europe. Europa Nova avoids this trap by designing for heterogeneity rather than standardising on a single vendor.

The CEPS report, along with analysis from the European Parliament Research Service and the Interface network ^[6][9][7], converged on a related criticism: the EU is betting too heavily on generative AI and should support more diverse, trustworthy approaches to artificial intelligence. Europa Nova's mandate addresses this directly. The institution's goal is not generative AI. It is the development of general intelligence through whatever architecture proves most effective — whether that turns out to be an evolved Transformer, a state-space model, a neuromorphic system, or something not yet invented. This architectural agnosticism is not indecision. It is the correct research strategy for an institution operating on a ten-year horizon in a field where the dominant paradigm may shift more than once.

7.2  Data Strategy: European Data Curation

Europe's data position is better than most commentary suggests. The standard narrative emphasises American dominance in web-scale data. This is true for English-language internet text. It is far less true for the multilingual, structured, and domain-specific data that increasingly determines model quality.

Consider the assets. National libraries across Europe have digitised material in all 24 official EU languages, as well as Norwegian, Swiss, and other non-EU European languages. Europeana, the EU's digital cultural heritage platform, provides access to over 58 million digitised objects — texts, images, audio, and video — from more than 3,600 cultural institutions across Europe. Public administration records across EU member states contain decades of policy documents, legal proceedings, regulatory filings, and administrative case law in every European language. The European scientific publishing infrastructure — from CERN's arXiv contributions to the Max Planck Society's open-access publications — generates a continuous stream of high-quality technical text.

The multilingual dimension is not merely a matter of coverage. Research on multilingual large language models demonstrates that training on diverse languages improves cross-lingual transfer and generalisation. Models trained on typologically diverse languages develop more robust internal representations — what researchers describe as language-agnostic semantic spaces — that support better reasoning across all languages, including English. Shi and colleagues showed in their ICLR 2023 work that multilingual chain-of-thought reasoning emerges as a capability at scale, with performance improving even in languages underrepresented in training data. Europe's linguistic diversity, often framed as a fragmentation problem, is in fact a training advantage for building general intelligence.

GDPR, too, is an asset rather than merely a constraint. Europa Nova can build a data governance framework with legal legitimacy that American firms operating under different legal regimes cannot easily replicate. Models trained on European data under European data protection law can be deployed in European healthcare, financial services, and public administration — domains where American cloud-based AI faces persistent legal uncertainty. A Europa Nova model with demonstrably compliant data provenance has a commercial advantage in precisely the markets where AI adoption generates the most economic value.

Europa Nova's approach to training data composition follows several principles grounded in current evidence.

On synthetic data, the evidence supports a clear distinction between pre-training and post-training roles. In pre-training, synthetic data should constitute no more than 10--15% of the corpus. Beyond this threshold, models risk what researchers term "model collapse" — a progressive degradation of output quality as the model increasingly trains on its own artefacts rather than genuine human expression. In post-training — the alignment, instruction-tuning, and reinforcement learning phases — synthetic data is not merely acceptable but essential. 80--90% of post-training data in current frontier models is synthetically generated, because the specific behaviours being trained require precisely controlled examples that are expensive to produce through human annotation alone.

Code should comprise 15--20% of the pre-training corpus. The evidence for this proportion is robust: multiple research groups have shown that code training improves reasoning capabilities across domains, including mathematics, logical inference, and structured problem-solving. Code's formal syntax and logical structure appear to teach models something transferable about precision and sequential reasoning.

On the question of bias in training data, Europa Nova takes a position that differs from both the American and Chinese approaches. The institution does not attempt to artificially balance training data or strip cultural context from sources. Instead, it annotates data sources with provenance, perspective, and context as structured metadata. A legal opinion from the German Constitutional Court is tagged as such. An editorial from Le Monde is tagged as such. A scientific paper from Nature is tagged as such. The model learns that perspectives exist and differ — that a French court and a Polish court may reach different conclusions on the same legal question, that an economist at the Bundesbank and an economist at the Banca d'Italia may assess the same fiscal policy differently. It does not learn that all perspectives are equally prevalent or equally supported by evidence. This produces calibrated understanding rather than the engineered neutrality that characterises American corporate models or the directed consensus that characterises Chinese state models.

7.3  Alignment and Safety: A European Approach

Alignment — the discipline of ensuring that AI systems behave in accordance with human values and intentions — is where Europa Nova has the most distinctive contribution to make. This is not because European researchers have a methodological advantage in reinforcement learning from human feedback or constitutional AI. It is because alignment is ultimately a question about values, and European political philosophy offers a framework for navigating value pluralism that neither American nor Chinese traditions adequately provide.

The dominant American approach to alignment reflects a particular intellectual heritage: constitutional principles applied through corporate governance. OpenAI, Anthropic, and Google each define alignment largely as adherence to a set of rules — a model constitution — crafted by a small team of researchers and policy staff in San Francisco. This works tolerably well for a product serving a global market from a single jurisdiction. It does not answer the deeper question of whose values an artificial general intelligence should embody when values genuinely conflict.

Europa Nova's approach is pluralistic alignment. The goal is not to encode one correct value set but to build systems capable of navigating genuine value diversity — understanding that a Finnish perspective on privacy differs from a French perspective on free expression, that Polish Catholic social teaching and Dutch secular liberalism produce different but legitimate frameworks for ethical reasoning. This reflects the institutional reality of the European Union itself: 27 nations, multiple legal traditions, no single constitutional authority, and a political culture built on negotiated difference rather than imposed consensus.

The obvious counterargument is that pluralistic alignment is too complex to implement. A model that tries to respect every perspective risks respecting none — producing outputs that are either incoherent or evasively neutral. This is a genuine risk. Europa Nova addresses it through two design choices. First, the model is trained to be explicit about the perspective it is applying in a given context. When advising on German employment law, it applies German legal reasoning. When generating a medical assessment, it applies evidence-based clinical standards, not cultural relativism. Second, the institution distinguishes between empirical questions (where evidence determines the answer), normative questions (where values determine the answer), and questions that combine both. For empirical questions, the model is trained to converge on evidence. For normative questions, it is trained to present the reasoning behind different positions without false balance. This is calibrated pluralism, not relativism.

Europa Nova adopts a framework of stratified openness that reflects the different risk profiles of different types of AI output. Fundamental research — mathematical results, architectural innovations, training methodologies — is always published openly. This serves both the global scientific community and Europa Nova's own recruitment, since top researchers will not work at an institution that suppresses publication. Models below a defined capability threshold are released with open weights, enabling European SMEs, universities, and public sector institutions to build on Europa Nova's work without licensing fees. Models above the threshold — those capable of tasks that pose genuine misuse risks — are available through controlled licensing via the Commercial Licensing Entity. Alignment and safety research is always open, regardless of the capability level it addresses. This is a global public good. Restricting safety research to gain competitive advantage is strategically foolish and morally indefensible.

On institutional collaboration, Europa Nova's Safety Lab should work directly with the UK AI Safety Institute, which has established itself as the most operationally advanced government safety evaluation body globally. The outcomes of France's AI safety summit process, including the international commitments made at the 2024 Seoul and 2025 Paris summits, provide a diplomatic framework within which Europa Nova's safety work gains international legitimacy. Europa Nova does not duplicate these efforts. It provides the technical research capacity — the ability to actually build and test frontier systems — that government safety institutes lack.

The final dimension of alignment is democratic. What values should a European AGI system embody? This is not a question that a Scientific Director or a nine-member board should answer alone, however brilliant they may be. It is ultimately a democratic question. Europa Nova addresses this through structured processes for involving broader society in alignment decisions — citizens' assemblies, expert panels, and deliberative processes modelled on the Irish citizens' assembly format, which has demonstrated that complex ethical questions can be addressed through well-designed democratic deliberation.

The process must be designed to produce actionable input, not endless deliberation. Europa Nova's Advisory Council of parliamentarians from participating nations provides democratic accountability, but the Council does not vote on architecture choices or training methodologies. The democratic input is channelled specifically to the questions that are genuinely normative: what topics should the model refuse to engage with? How should it handle conflicts between national legal frameworks? What level of transparency about its own reasoning processes should it maintain? These questions have right answers only in a democratic sense — they reflect the considered preferences of the societies the system serves.

Alignment is where Europa Nova can lead. Not because European values are superior, but because the European institutional tradition — pluralistic, negotiated, legally structured, democratically accountable — maps more naturally onto the alignment problem than any alternative. Building an AGI that can navigate genuine moral complexity without either imposing a single worldview or collapsing into relativism is the hardest technical and philosophical challenge in the field. It is also the challenge for which European intellectual and political traditions are best prepared.

8 The Path Forward

The preceding chapters have established the diagnosis: Europe lacks sovereign AGI capability, the window to build it is narrowing, and existing EU initiatives do not fill the gap. They have described the remedy: Europa Nova, a treaty-based research institution with concentrated scientific authority, dedicated compute infrastructure, and a commercial licensing arm. What remains is the sequence of decisions required to move from paper to institution. This chapter sets out the timeline, the immediate actions, and the conditions under which Europa Nova either succeeds or is shut down.

The timeline is aggressive by the standards of European institution-building. It is not aggressive by the standards of the technology it must compete with. OpenAI went from a research paper to a $150 billion valuation in seven years. DeepSeek went from founding to frontier capability in eighteen months. Europa Nova has roughly ten years to achieve operational maturity. Every month of delay in the founding phase compounds into years of lost research output.

8.1  Timeline

Phase 0: Coalition Building (Now — 12 Months)

The first year is diplomatic, not technical. Its purpose is to assemble the founding coalition: five to eight nations willing to commit political capital and sovereign funds to a treaty-based institution. The most probable founding group includes France, Germany, Norway, Sweden, and the United Kingdom, with the Netherlands, Switzerland, and Finland as strong candidates for early accession. Each brings something distinct: France and Germany provide political weight and research depth; Norway brings the Government Pension Fund and abundant hydroelectric power; Sweden offers an advanced energy grid and the Nordic data infrastructure tradition; the United Kingdom contributes DeepMind-adjacent expertise and post-Brexit appetite for new European partnerships.

Three actions define this phase. First, the appointment of a sherpa team — four to six senior officials with direct authority from their heads of state or government — to negotiate the treaty's institutional architecture, funding formulae, and governance design. CERN's founding convention was negotiated in under eighteen months from the first intergovernmental meeting to signature. Europa Nova should aim for the same pace. Second, formal but confidential engagement with three to five candidates for the role of Scientific Director. The calibre of this appointment will determine whether the institution attracts world-class researchers or becomes another administrative body. The search must begin before the treaty is signed, not after. Third, the European Investment Bank should be engaged under the existing TechEU framework ^[5] to define co-financing parameters for physical infrastructure. The EIB's involvement signals institutional seriousness to private investors and sovereign wealth funds.

Site assessment for the primary compute facility begins in parallel. The analysis in Chapter 4 identified Northern Scandinavia as the optimal location on energy, climate, and grid capacity grounds. Phase 0 narrows this to two or three candidate sites through joint assessment with energy providers, grid operators, and infrastructure developers already active in the region.

Phase 1: Foundation (Months 0–24)

Phase 1 begins with the signing of the Europa Nova Convention. The convention establishes the legal personality of the institution, its governance structure, and the binding financial commitments of each signatory. Treaty ratification by national parliaments proceeds in parallel with operational preparations — the convention should include provisional application clauses that allow recruitment and site preparation before all ratifications are complete.

The Scientific Director is formally appointed within six months of signing. By month twelve, the Director has assembled an initial leadership team of five to eight senior researchers, recruited from the world's best AI labs with compensation packages competitive with DeepMind and OpenAI. This team defines the institution's first research programme, specifying which architectural directions Europa Nova will pursue and what compute configurations are needed to support them.

The commercial licensing entity, Europa Nova Innovations, is incorporated during this phase as a wholly owned subsidiary under the jurisdiction of a single signatory nation — most likely the Netherlands or Luxembourg, both of which have well-established frameworks for international research commercialisation. Site selection for the compute facility is finalised. Environmental impact assessment and construction permits are initiated. These permitting processes are the most likely source of delay; the convention should include host-nation commitments to expedited review, as ESA negotiated for its Kourou launch facility.

Phase 2: Build (Months 12–36)

Construction of the primary compute facility begins, targeting initial operational capacity of 500MW within twenty-four months of groundbreaking. This is an industrial construction timeline, not a research campus timeline. Modular construction techniques allow phased deployment: the first 100MW tranche can be operational while subsequent phases are still under construction.

Recruitment accelerates. The target is 200 to 400 researchers by month thirty-six, organised into the research divisions defined by the Scientific Director's programme. The critical mass threshold for productive frontier research — based on the experience of Google Brain, DeepMind, and FAIR — is approximately 150 researchers with full compute access. Falling below this number risks producing an institution that publishes papers but cannot train and evaluate models at scale.

First architectural research results should be published by month twenty-four. These will not be foundation models. They will be proof-of-concept demonstrations that Europa Nova's chosen architectural directions deliver measurable efficiency gains over pure Transformer baselines. Publication establishes scientific credibility and attracts additional talent. Partnerships with European universities are formalised through joint research agreements and compute-access programmes.

Year 3: Evaluation Gate

At month thirty-six, Europa Nova faces its first existential test. An independent external review panel — composed of internationally recognised AI researchers, none of whom are employed by or affiliated with Europa Nova — conducts a comprehensive assessment. The panel's mandate is binary: recommend continuation with full funding, or recommend orderly wind-down.

The evaluation criteria must be defined in the founding convention, not negotiated after the fact. They focus on scientific progress: quality and impact of publications, demonstrated architectural innovations and their measured efficiency advantages, ability to attract and retain world-class researchers, and progress toward training models that perform competitively on recognised benchmarks. Political metrics — number of member-state partnerships formed, diversity of hiring, geographic distribution of spending — are explicitly excluded. These are governance concerns, not scientific ones, and conflating the two is how European research institutions lose their edge.

The wind-down option is not a failure state to be avoided at all costs. It is the mechanism that gives Europa Nova credibility. A €50 billion programme that cannot be stopped is a political liability. A programme that subjects itself to rigorous external evaluation and accepts termination if it fails to deliver is one that policymakers and taxpayers can support with confidence. The Airbus consortium did not include such a gate, and the political impossibility of shutting down underperforming divisions led to decades of workshare distortions. Europa Nova learns from that mistake.

Phase 3: Scale (Months 36–84)

Assuming the evaluation gate is passed, Phase 3 scales Europa Nova from a research institution to a productive force in the global AI landscape. Compute infrastructure expands from 500MW to gigawatt capacity, with hardware refreshes incorporating next-generation accelerators and the heterogeneous architectures described in Chapter 7. Research staff grows to 800--1,200. The institution's first foundation models are released — not as commercial products, but as open-weight releases for European research and industry, licensed through Europa Nova Innovations.

Commercial licensing operations begin generating revenue. European SMEs, public-sector organisations, and research institutions access Europa Nova's models and compute through tiered licensing arrangements. Revenue targets are conservative in the first two years of commercial operation — the primary purpose is to establish market position, not to maximise short-term income. By month eighty-four, commercial revenue should cover 15 to 25 percent of annual operating costs, with a clear trajectory toward higher ratios.

Phase 4: Frontier (Months 84–120)

In its final phase within this paper's planning horizon, Europa Nova operates at the frontier of AGI research. Its models are competitive with the best produced anywhere in the world. Its architectural innovations have demonstrably changed the efficiency frontier for AI training and inference. Commercial revenue approaches levels sufficient to sustain research operations without continued growth in treaty contributions — though sovereign funding remains the backbone of basic research, as it does for CERN.

Europe, at the end of this decade, possesses what it lacks today: a sovereign institution capable of producing, understanding, and governing artificial general intelligence. The technology does not depend on American cloud providers, Chinese hardware, or the strategic decisions of any single foreign corporation. This is the objective. Everything in this paper is in service of it.

8.2  Immediate Next Steps

Five actions in the next twelve months determine whether Europa Nova becomes an institution or remains a policy paper.

First, a founding summit. The heads of state or relevant ministers of five to eight European nations convene to issue a joint declaration of intent. The summit's purpose is political commitment, not technical specification. Its deliverable is a single document: a declaration that the signatories intend to negotiate a treaty establishing an independent European research institution for artificial general intelligence, funded at a level commensurate with the strategic stakes. The declaration need not specify budgets, governance, or location. It must specify intent and urgency.

Second, the appointment of sherpas. Each founding nation appoints a senior official — at the level of a state secretary or permanent secretary — with authority to negotiate institutional structure, funding commitments, and governance design on behalf of their government. These are not advisory roles. The sherpas must have mandates to make binding commitments, subject to parliamentary ratification. The treaty preparation team should aim to have a draft convention ready for signature within twelve to eighteen months.

Third, a formal site assessment. A joint technical team, working with energy providers, grid operators, and infrastructure developers already active in Northern Scandinavia, conducts a formal assessment of two to three candidate sites for the primary compute facility. The assessment covers power availability and reliability, grid connection capacity, cooling infrastructure, construction logistics, and transport links. The European Spallation Source in Lund provides a relevant precedent for multinational site selection in Scandinavia.

Fourth, constitution of a Scientific Director search committee. An international committee of five to seven recognised leaders in AI research is convened to identify and approach candidates. The committee's composition should span American, European, and Asian AI research communities — demonstrating that Europa Nova seeks the world's best, not the best available within existing European networks. The committee operates confidentially until a shortlist is agreed.

Fifth, engagement with the EIB. Under the existing TechEU framework, the European Investment Bank has committed indicative allocations for compute infrastructure co-financing. Europa Nova's infrastructure team engages the EIB to define co-financing parameters — loan terms, equity structures, and disbursement schedules — for the compute facility's physical construction. Early engagement signals institutional credibility and unlocks private co-investment.

These five actions are not preconditions for each other. They proceed in parallel. The founding summit enables the sherpa appointment. The site assessment and Scientific Director search can begin before the summit, provided there is informal agreement among the founding group. The EIB engagement requires only a credible institutional counterparty — which the sherpa team provides. Twelve months is sufficient if the political will exists. If it does not, no amount of technical preparation will compensate.

8.3  The Decision

The analysis is complete. The strategic deficit is documented. The technical opportunity is identified. The institutional design is specified, the funding model is constructed, the governance is architected, and the research strategy is defined. What remains is not a question of evidence or argument. It is a question of political will. Europe has built CERN, Airbus, ESA, and the single market — each time against scepticism that cooperation at this scale was possible, each time because a small number of leaders decided that the cost of inaction was greater than the risk of ambition. Artificial general intelligence is the defining technology of this century. Europe can build it, or Europe can depend on others who do. The window is open. It will not stay open long.

APPENDIX A

References

Official EU and EuroHPC Documents

1. [1] European Commission, “InvestAI: Stepping Up Investment in AI,” February 2025.
2. [2] European Commission, “AI Continent Action Plan,” April 2025.
3. [3] EuroHPC Joint Undertaking, AI Factories selection decisions, December 2024, March 2025, October 2025.
4. [4] Council of the EU, “Artificial intelligence: Council paves the way for the creation of AI gigafactories,” January 2026.
5. [5] European Investment Bank, “EIB Group and European Commission join forces to finance AI gigafactories,” December 2025.

Independent Policy Analyses

1. [6] Renda, A. and Kyosovska, N., “EU Plans for AI (Giga)Factories: Sanctuaries of Innovation, or Cathedrals in the Desert?,” CEPS In-Depth Analysis, November 2025.
2. [7] Lemke, N. and Schneider, C., “The European Union’s AI Factories,” Interface, October 2025.
3. [8] Hess, J. C. and Sieker, F., “Built for Purpose? Demand-Led Scenarios for Europe’s AI Gigafactories,” Interface / Bertelsmann Stiftung, October 2025.
4. [9] European Parliament Research Service, “Artificial Intelligence in the EU: State of Play,” EPRS Briefing, 2025.

Energy and Infrastructure

1. [10] International Energy Agency, “Energy and AI,” IEA Special Report, June 2025.
2. [11] IEA, “Overcoming energy constraints is key to delivering on Europe’s data centre goals,” Commentary, 2025.
3. [12] Ember, “Grids for data centres: ambitious grid planning can win Europe’s AI race,” June 2025.
4. [13] S&P Global, “European data center power demand to double by 2030,” July 2025.
5. [14] Fortum, “The real foundation for data centres in the Nordics is energy stability,” 2025.

Industry and Infrastructure Announcements

1. [15] OpenAI, “Introducing Stargate Norway,” July 2025.
2. [16] Nscale, Aker ASA, and OpenAI, “Stargate Norway” joint press release, July 2025.
3. [17] SoftBank, Oracle, and OpenAI, “The Stargate Project,” January 2025.
4. [18] Microsoft, “The Golden Opportunity for American AI,” January 2025.

Technical Research

1. [19] Vaswani, A. et al., “Attention Is All You Need,” NeurIPS 2017.
2. [20] Gu, A. and Dao, T., “Mamba: Linear-Time Sequence Modeling with Selective State Spaces,” ICLR 2024.
3. [21] Gu, A., Goel, K. and Ré, C., “Efficiently Modeling Long Sequences with Structured State Spaces,” ICLR 2022.
4. [22] Hoffmann, J. et al., “Training Compute-Optimal Large Language Models,” arXiv:2203.15556, 2022.
5. [23] Stanford HAI, AI Index Report 2025, Stanford University, April 2025.

Data and Statistics

1. [24] Macro Polo, Global AI Talent Tracker, Paulson Institute, 2025.
2. [25] Eurostat, “Artificial intelligence — statistics on the use by enterprises,” 2025.

APPENDIX B

Institutional Comparison

APPENDIX C

Illustrative Treaty Structure

The following outlines the indicative structure of the Europa Nova Convention, drawing on the CERN Convention (1953), the ESA Convention (1975), and the ESO Convention (1962) as precedents.

Preamble. The Contracting States, recognising that artificial general intelligence represents a technology of transformative consequence for European security, prosperity, and values; affirming that no single European nation possesses the resources to develop sovereign AGI capability independently; and determined to establish an institution of the highest scientific calibre, have agreed as follows.

Article I — Establishment. An international organisation, to be known as Europa Nova, is hereby established. Europa Nova shall have legal personality and the capacity to contract, acquire and dispose of property, and institute legal proceedings.

Article II — Purpose. The purpose of Europa Nova is to conduct fundamental and applied research in artificial general intelligence; to build and operate the compute infrastructure required for such research; and to license the resulting technology for the benefit of European industry, public institutions, and citizens.

Article III — Membership. The founding members are [list]. Additional states may accede to this Convention upon invitation by the Board, subject to ratification by the applicant state’s parliament and approval by a two-thirds majority of existing members.

Article IV — Organs. Europa Nova shall comprise: (a) a Board of nine members, serving as the governing body; (b) a Scientific Director, serving as the chief executive and scientific officer; (c) a Commercial Licensing Entity, established as a wholly owned subsidiary; and (d) an Advisory Council of parliamentarians from Contracting States.

Article V — The Board. The Board shall consist of nine members appointed by the search committee for terms of five years, renewable once. Board members serve in their personal capacity, not as national representatives. The Board approves the annual budget, appoints and may dismiss the Scientific Director, and sets the strategic direction of the institution.

Article VI — Scientific Director. The Scientific Director is appointed by the Board following an international search. The Scientific Director has sole authority over the research programme and holds veto power on all technical decisions. The Scientific Director reports to the Board.

Article VII — Finance. Each Contracting State shall contribute to the budget in proportion to a formula based on gross domestic product and national wealth, as specified in Annex I. Contributions constitute treaty obligations and are not subject to annual appropriation. The budget period is five years, with mid-term review.

Article VIII — Privileges and Immunities. Europa Nova, its officials, and its property shall enjoy such privileges and immunities as are necessary for the fulfilment of its purpose, as specified in a Protocol annexed to this Convention.

Article IX — Evaluation and Termination. At the end of the third year following the Convention’s entry into force, an independent external panel shall evaluate Europa Nova’s scientific progress. Should the panel recommend termination, the Board shall initiate orderly wind-down proceedings as specified in Annex II. Assets shall be disposed of and proceeds returned to Contracting States in proportion to their cumulative contributions.

Article X — Amendment and Duration. This Convention shall remain in force for an initial period of fifteen years. Amendments require approval by a two-thirds majority of Contracting States and ratification by national parliaments.

APPENDIX D

Financial Model Detail

The following tables present the detailed financial assumptions, national contribution allocations, ten-year cash flow projections, risk analysis, and sensitivity scenarios underpinning the €50 billion programme envelope described in Chapter 5. All figures are in €M unless otherwise stated.

D.1  Sovereign Treaty Contributions — Illustrative National Allocation

Contributions are structured as treaty obligations (folkerettslig bindende), not subject to annual appropriation. The allocation formula is based on GDP and national wealth, with host-nation premium for Norway and Nordic co-host credit for Sweden. All contributions represent less than 0.1% of national budgets — comparable in scale to existing multilateral research commitments.

D.2  10-Year Cash Flow Model

Model design: Treaty contributions provide a stable base. GPFG investment provides scale and front-loads infrastructure spend. Commercial revenue provides a path to operational self-sufficiency. EIB co-financing reduces capital cost for physical infrastructure. Commercial revenue covers 68% of operating costs by Year 10 (baseline). Full operational self-sufficiency requires approximately €3.7B/yr — achievable in the optimistic scenario by Year 10–11.

D.3  Risk Analysis

Net assessment: Residual infrastructure value of €8–10 bn even in worst-case termination provides downside protection. The Year 3 evaluation gate is a feature, not a risk — it is what makes the proposal credible to sceptics and fiscally responsible to taxpayers.

D.4  Sensitivity Analysis — Net 10-Year Surplus (€M)

How does the net surplus change under different assumptions for GPFG allocation and commercial revenue scale? Base case (highlighted): GPFG = €22.5 bn, Commercial Yr 10 = €2.5 bn. Fixed inputs: Treaty = €30 bn, EIB = €4 bn, Expenditure = €50 bn.

Reading the table: The highlighted cell (€16,050M) is the base case — GPFG at €22.5 bn and commercial revenue reaching €2.5 bn/yr by Year 10. Red/parenthesised values indicate scenarios where total funding falls below €50 bn expenditure, requiring additional sources or scope reduction. Even at minimum GPFG (€12 bn) and weak commercial (€1 bn/yr), the deficit is only €4 bn — coverable by expanded EIB lending or additional member nations. Key insight: GPFG allocation is the dominant variable; each €1 bn increase adds €1 bn to the net surplus directly.

Building on Existing Work

Europa Nova does not emerge in a vacuum. The case for a European CERN for AI has been made by others, with greater expertise and over a longer period. This paper builds on their work and attempts to answer the operational design questions that remain open.

CAIRNE — The Confederation of Laboratories for AI Research in Europe has advocated for a CERN for AI since 2018, mobilising the European AI community and pressing the case with EU policymakers. Their analysis of what a pan-European AI institution requires — autonomy, longevity, trust in science — informs Europa Nova’s governance design. cairne.eu

Centre for Future Generations — CFG’s January 2025 blueprint “Building CERN for AI” and their December 2025 geopolitics report (co-authored with RAND Europe, MIT, Oxford, and Max Planck) provide the most detailed institutional analysis to date. Europa Nova’s emphasis on trustworthy AI and competitive compute reflects their work. cfg.eu

CEPS — Andrea Renda’s analyses — the July 2024 scoping paper on a European large-scale AI initiative and the November 2025 “Sanctuaries of Innovation or Cathedrals in the Desert?” — document the structural gaps in the EU’s current approach. Europa Nova is designed to fill those gaps. ceps.eu

Interface and Bertelsmann Stiftung — The October 2025 policy briefs on AI Factories and AI Gigafactories provide the empirical evidence for Europa Nova’s central argument: that compute infrastructure without a research organisation is insufficient. interface-eu.org

EDIRAS — The European Distributed Institute for AI in Science, proposed by the EU’s Scientific Advice Mechanism in 2024, represents a different model: distributed and science-focused. Europa Nova disagrees with the distributed approach but shares EDIRAS’s concern for European research sovereignty.

Europa Nova’s contribution to this landscape is operational specificity: a treaty structure outside the EU framework, physical concentration of compute in Northern Scandinavia based on energy economics, a commercial licensing entity, financing through Norway’s sovereign wealth fund, and a year-3 evaluation gate. These are the design choices that remain unresolved in the broader debate. This paper is an attempt to resolve them — and an invitation to those who have been working on this longer to tell us where we are wrong.