At GTC Paris — held alongside VivaTech, Europe’s largest tech event last June — Nvidia founder and CEO Jensen Huang told a rapt audience: “AI is the essential infrastructure of our time, just as electricity and the Internet once were,” while calling on Europe to exercise “bold leadership.”
Clad in his signature leather jacket, Huang’s appearance in Paris — where he was treated as a rock star — was part of a trip to European capitals to pitch “sovereign” AI infrastructure to leaders eager to play catch-up. The proposed new data centers offering essential compute power within national borders will be powered by Nvidia GPUs, Nvidia software stacks (CUDA), and Nvidia-certified cloud partners — all of which remain under American corporate and legal control.
The Huang roadshow was, in many ways, a live illustration of the Trojan Horse thesis of a recent report by The Future Society (TFS) entitled “Beware of Geeks Bearing Gifts: Building True EU Frontier AI Sovereignty.” The report outlines how Europe’s most consequential AI dependencies may not arrive through coercion, but as gifts: infrastructure investments, model access agreements, or technology partnerships that appear beneficial in the short run but quietly erode Europe’s capacity for self-determination.
TFS’s report provides a framework to assess the trade-offs across sovereignty pillars and frontier AI stack components. Authored by TFS’s Executive Director Nick Moës and five co-researchers, it is the most forensic public audit yet of where the European Union actually stands in the global AI race — not only at the level of strategy documents and political communiqués, but also layer by layer, component by component, all the way down to the wet chemicals used to clean silicon wafers.
The picture it paints is sobering.
Frontier AI is not just another technology sector. Unlike prior general-purpose technologies, it is designed to act with increasing autonomy, embedding the values and priorities of its developers into everything it touches, notes the report. With a projected impact on global economic output of up to 15% over the next decade, it is already reshaping labor markets, democratic processes, and warfare.
Europe enters this transformation from a position of structural dependence: frontier AI models are almost exclusively produced in the U.S. and China, and Europe’s supercomputing capacity relative to global leaders is in consistent decline. Yet this capacity is important to develop or even just use frontier AI models at scale.
Following the call to action in the 2024 Draghi report, the European Commission has launched no fewer than 92 AI initiatives spanning compute infrastructure, data governance, regulatory reform, research funding, and skills.
What Europe needs, the authors argue, is not a list of initiatives but a coherent view of the stack: a clear understanding of which layers require European control for sovereignty to be meaningful, which can safely be sourced externally, and which represent the leverage points where targeted investment generates outsized returns.
Only then can the Continent craft a clear strategy. There is no silver bullet, Moës said in an interview with The Innovator, but “Europe is wasting its potential right now. It can take short- and long-term measures to structurally improve its responses and very concretely reach out to all nations that are being affected — not by pretending to be the alternative but to try to co-shape the alternative to the U.S. and China together.”
The Anatomy of an Infrastructure Europe Doesn’t Own
The report’s centerpiece is what the authors call the Frontier AI Stack: a 26-layer decomposition of everything required to build and deploy the world’s most powerful AI systems. It spans foundational resources (the critical raw materials, electricity, water, and land needed to house data centers), hardware (the sprawling semiconductor supply chain), software (operating systems, model architectures, applications), data (training sets, synthetic data, the feedback loops that improve deployed models), and what the authors term the “enabling environment” — talent, finance, governance, and the market conditions in which all of it operates.
The report documents the EU’s position in each.
Europe’s supporters proudly point to the Netherlands’ ASML. When it ships one of its extreme ultraviolet lithography machines, the transaction is a geopolitical event. Each machine — the size of a double-decker bus, worth upwards of $380 million — is the only device on earth capable of etching the most advanced chips onto silicon. No ASML, no cutting-edge semiconductors. No cutting-edge semiconductors, no frontier AI. The Dutch firm holds a near-monopoly, accounting for roughly 78% of global lithography revenue. It is, in the language of policy analysts, a choke point.
But Europe does not have many of these.
Take the hardware layer. In advanced CPU design — the general-purpose chips that manage AI data centers — the EU’s presence is described as “negligible.” The market belongs to Intel, AMD, Amazon, and Apple. In discrete GPUs, the workhorses of AI training, the EU has “no considerable share.” Nvidia dominates, with AMD and Intel holding much smaller positions. In memory chip design, including the high-bandwidth memory that has become a bottleneck for frontier AI workloads, the EU has “no major commodity memory chip company comparable to the global leaders.” South Korea’s SK Hynix and Samsung, alongside America’s Micron, have that market largely to themselves.
The pattern continues down into fabrication sub-components — the equipment and chemistry that manufactures chips. The EU is entirely absent from wafer and photomask handlers, ion implanters, and several classes of test equipment. In each of these categories, Japan and the United States divide the spoils.
There are bright spots. Beyond ASML’s dominance in lithography, the Netherlands’ ASM International holds roughly 55% of the atomic layer deposition market, a critical component in modern transistors. Austria’s EV Group commands 56% of the wafer bonding segment. Germany contributes through Siemens EDA (around 13% of chip design software), Carl Zeiss (optics inside ASML machines), and Linde (industrial gases). France’s Air Liquide is a significant supplier of the specialty gases that semiconductor fabs consume in enormous quantities.
But these islands of excellence exist within a landscape of dependence. “Sovereignty over frontier AI cannot be assessed through the technical value chain alone,” the authors write. “A state that controls chips and models, but depends on foreign energy, lacks capital markets, or cannot retain talent, does not hold meaningful sovereignty.”
The Energy Problem
Beneath the semiconductor story lies an even more fundamental constraint that the report addresses: electricity.
AI runs on power — enormous quantities of it. Data centers already account for about 1.5% of global electricity consumption, and that figure is expected to more than double by 2030. The compute required to train and run frontier models is only growing. And here, Europe faces a structural disadvantage with no easy fix.
In major European data center hubs — Frankfurt, Amsterdam, Paris, Dublin — grid connection queues average seven to ten years. That is not a bureaucratic inconvenience; it is a hard physical limit on how quickly new AI infrastructure can come online. Meanwhile, EU industrial electricity prices were 158% higher than in the U.S. in 2023, a consequence of Europe’s dependence on imported fossil fuels against America’s status as a net energy exporter.
The water problem is less discussed but equally real. Next-generation server racks are expected to require liquid cooling at densities approaching 600 kilowatts. Water scarcity already affects 28% of EU territory at least seasonally, and roughly 30% of people in southern Europe face permanent water stress. Spain and Greece rank among the highest-risk locations globally for data center water stress this decade. The Nordic countries, with their cold climates and clean energy, are the structural winners.
China, the report notes with some alarm, is rapidly expanding power generation capacity through large-scale clean-power deployment and an accelerating nuclear pipeline. The U.S. commands roughly sixteen times the EU’s AI supercomputing capacity. Further, only 15% of global hyperscale data center capacity sits within the EU, against 54% in the United States.
Software, Data, and the Invisible Dependency
Moving up the stack from hardware, the picture does not improve. In the software layer, the EU holds no significant position in the operating systems and frameworks that run AI workloads — those are American. Nor does it have anything comparable to the frontier model ecosystems built by OpenAI, Google DeepMind, Anthropic, or Meta. France’s Mistral AI is the only European developer publicly tracked alongside them, but trails on nearly every relevant metric, such as capabilities or active users. The Franco-German Frontier AI Initiative, launched in late 2025, represents an attempt to change this, but it is early days.
The data layer raises questions that go beyond market share. The report identifies five distinct data components: generic open data, specialized proprietary datasets, synthetic data, human data labor (the armies of annotators and evaluators that train and grade AI systems), and what the authors call the “deployment data flywheel” — the self-reinforcing feedback loop by which widely deployed models improve through use. On this last point, the EU faces a compounding disadvantage: models deployed at lower scale generate less feedback, which could mean they improve more slowly, attracting fewer users. Breaking into a market with established incumbent flywheels without a very large initial deployment or a distinctive advantage is difficult. The EU has neither at the frontier model level.
The enabling environment layer — talent, finance, standards, and market dynamics — reveals structural barriers for Europe. The EU is home to world-class AI researchers, but it loses a significant share of them to better-funded American and British companies. The venture capital ecosystem for AI is thinner than in the U.S., and the concentration of investment around a handful of U.S. hyperscalers has created what the report calls conditions of “sovereignty washing”: American companies offering European governments rebranded products that technically comply with data residency requirements while the underlying control and data exposure remain extraterritorial.
Three U.S. hyperscalers account for around 65% of the EU cloud market. The EU’s own cloud champions — IONOS, OVHcloud — are a fraction of the size.
A Framework for What’s Actually at Stake
What makes the report more than an inventory of deficiencies is its proposal for a unified framework that connects this technical map to a richer understanding of what sovereignty actually means.
European policy discourse, the authors argue, has fallen into a trap: equating sovereignty with economic competitiveness, and economic competitiveness with deregulation. This is analytically wrong and politically dangerous, says the report, as regulation reclaims some form of control over the technological trajectory and dominant business models. It proposes five distinct sovereignty pillars, each of which frontier AI affects differently and sometimes in direct tension with the others.
Economic Competitiveness is the most familiar: can Europe generate and capture economic value from AI? But it sits alongside Resilience and Preparedness — can European systems withstand shocks, from a chip export ban to a major cyberattack? Security and Defense — can Europe protect its citizens and infrastructure from adversaries using AI-enabled tools, from disinformation campaigns to semi-autonomous weapons? European Values and Identity — can Europe ensure that the AI systems shaping its citizens’ lives embody European commitments to human dignity, privacy, and democratic accountability? And Foreign Relations and Perceptions — can Europe set the terms of its own engagement with the world on AI, rather than accepting terms set by Washington or Beijing?
The tendency to treat AI sovereignty as a purely economic question is an error, says the report. It points out that the European Values and Identity pillar of the sovereignty framework is crucial. When AI systems increasingly shape hiring decisions, medical diagnoses, legal judgments, and political information environments, the question of whose values are embedded in those systems is a sovereignty question in the deepest sense. The EU’s answer to that question — human dignity, democratic accountability, the rule of law — is one of the few genuine differentiators it has. Treating it as a cost to be minimized in the pursuit of competitiveness would be, in the report’s framing, a category error of historic proportions.
The framework’s power lies in making visible the trade-offs between these pillars that current policy tends to paper over. Opening up proprietary health data for AI training, for instance, might boost economic competitiveness by enabling richer models — but at direct cost to European values and potentially to security, by increasing vulnerability to foreign influence. Aggressive deregulation to attract AI investment might improve short-term competitiveness while undermining the regulatory tools that give the EU global standard-setting and bargaining power. Partnering with U.S. hyperscalers for AI Gigafactories might close the compute gap while deepening the dependency it was meant to address.
The authors apply this framework to five current EU initiatives. The AI Regulatory Sandbox program scores well on Economic Competitiveness but raises questions about values when it systematically exempts companies from third-party scrutiny. InvestAI, the Commission’s funding vehicle, concentrates benefits but may not address structural barriers in the lower layers of the stack. The European Health Data Space creates access to valuable training data but requires careful handling of the security and values trade-offs. The Frontier AI Initiative is welcome but currently underpowered.
Most forensically, the report analyzes the EU’s flagship compute initiative: the AI Gigafactories.
The Trouble With Europe’s AI Gigafactories
The AI Continent Action Plan, released in April 2025, set out to establish up to five AI Gigafactories across the EU, each with computing capacity equivalent to 100,000 of Nvidia’s H100 processors. It is the most concrete attempt yet to close the compute infrastructure gap. The report’s treatment of this initiative is clear-eyed.
The Gigafactories address compute infrastructure — the eighth component in the Frontier AI Stack — but their effectiveness depends on layers the initiative does not control. If the factories run on U.S.-designed chips (which they almost certainly will), they might reproduce dependency at a different level. If they are powered by electricity that remains 158% more expensive than in the U.S., the economics of training large models in Europe remain unfavorable. If the networking hardware inside the factories comes primarily from American or Asian vendors, the infrastructure is sovereign in name only. And if the cloud software running on top of the compute is provided by hyperscalers whose legal obligations run to American courts, the sovereignty case weakens further.
None of this means the Gigafactory initiative is not needed. More European compute is genuinely important. But the report’s framework forces the question: what precise sovereignty problem are these factories solving, across which pillars, and what dependencies will they leave intact — or will they create new ones? Questions must also be raised about whether their scale is adequate, as what is planned is dwarfed by foreign buildouts. Without that analysis, there is a risk of what the authors might call sovereignty theater — large, visible investments that address a real gap while leaving the structural architecture of dependence unchanged.
The EU currently accounts for just 4.8% of global supercomputer performance, against 74.4% for the U.S. and 14.1% for China. The gap is vast. Closing it matters. But the manner of closing it matters just as much.
Creating the Ability to Set Terms
The report suggests that some of the generosity flowing toward Europe from American tech giants — the data center investments, the cloud agreements, the AI partnerships — comes at the cost of dependencies that are difficult to reverse once established.
But sovereignty does not mean isolation. The report quotes Canadian Prime Minister Mark Carney’s Davos remark that “a world of fortresses will be poorer, more fragile, and less sustainable.” The goal is not to build everything in Europe but to maintain what Carney called the ability to set terms.
There are concrete things that Europe can and should do now to ensure that it safeguards its values and that short-term measures do not hurt the long-term objectives, says Moës. These include:
- Treating compute infrastructure as the public infrastructure it is: given the broad industrial, governmental, and social dependencies and capital intensity, EU governments should apply at least the same caution — and scrutiny — for foreign ownership of domestic compute as they have for electricity grids, gas pipelines, or water management. More than just a defensive approach, this sets the foundation for leapfrogging toward grid computing paradigms at scale.
- Building a safety net. What happens should the U.S. or China suddenly cut off service or access to their tech? Forging alliances with so-called middle powers on critical components of the tech stack and creating backup plans is a must. Teaming with Japan on chips could be one avenue. Relying more on open-source solutions is another.
- Using collective bargaining power: Pool together all EU countries for commercial purposes to negotiate the pricing and terms of AI contracts, and consider adding other middle powers, such as Brazil, to the mix.
- Smartly negotiating public contracts to gain access to Mythos without giving in on the risks it poses to the EU: Mythos, Anthropic’s latest and most powerful AI model, claims to be able to autonomously find and exploit zero-day vulnerabilities in every major operating system. It is, from a national security perspective, a significant threat — raising questions about who gets access, under what terms, and what public authority governs private developers of such dangerous technologies.
- Encouraging technology transfers: Canadian AI company Cohere’s acquisition of German AI company Aleph Alpha in April is an example. The governments of both countries helped facilitate the deal. The new company plans to attract customers in business and government that are uncomfortable relying on American tech firms for artificial intelligence and other digital services.
- Providing incentives: Encourage European organizations not to give up independence in exchange for foreign technology or give away Europe’s valuable data to foreign entities for training purposes; for example, using very specific but still commercially attractive advanced market commitments and innovation procurement.
A Meiji Moment?
Europe woke up late to the Internet. It is trying not to repeat the mistake with AI. Whether the stack analysis this report provides translates into the kind of coherent, long-term industrial strategy Meiji reformers once executed — transforming Japan, within a single generation, from a feudal society to a modern industrial state capable of defeating both China (1895) and Russia (1905) in war — is an open question.
The lithography machine is a lesson. Own the choke points and the world comes to you. Europe has some. The question is whether it can find more — and whether it understands clearly enough what it is trying to protect.
This article is content that would normally only be available to subscribers. Become a subscriber to see what you have been missing
