The AI Gigafactory or: Ηow the EU learned to stop worrying and love the hyperscaler

Acknowledgements The event was sponsored by the Amsterdam Centre for European Studies and specifically the theme group on Technology, Power, and Policy. The blogpost was written based on notes taken by Marcella Lucardie, who worked as a research intern at the Institute for Information Law. The authors would like to thank the panelists for their contributions; editorial decisions were made to reflect the authors’ views. EU’s push for scale At the 2025 Paris AI Action Summit, European Commission President Ursula von der Leyen unveiled the construction of five AI Gigafactories across the EU under the InvestAI framework and tied to a €20 billion public-private funding scheme. These new facilities are aimed at training industrial-scale “AI frontier models,” that is, very large, resource-intensive general-purpose AI models. In a way, AI Gigafactories signal the EU’s ambition to compete with US hyperscalers like Amazon, Microsoft, and Google, by building on existing supercomputing infrastructure. This is part of the EU’s digital sovereignty agenda, focused on advancing public and private investments in the EU’s own AI models and the wide-scale integration of AI technologies across all sectors, as well as developing and deploying trustworthy and human-centric AI governed through laws such as the AI Act. However, it is hard to ignore the inherent inconsistencies of the agenda. While the AI Gigafactories are not designated yet, recent reporting suggests that the project is plagued by financial and political hurdles that have thrown a wrench in its works, undermining the realisation of the project even before its official launch. Against that background, on 10 March 2026, we convened an expert roundtable in Amsterdam with Dr Charlotte Ducuing , Dr Fabbian Ferrari , Anushka Mittal and Leevi Saari to discuss these challenges. This essay reflects on the discussion and advocates for a shift of focus regarding the development of digital infrastructures. The hard(ware) truth One of the starkest realisations of our discussion was that, in the absence of a home-grown industry, Gigafactories risk being entirely reliant on hardware suppliers which are almost exclusively non-European large tech companies. Already, a vast majority of the European AI start-ups’ later-stage funding comes from American investors, including Big Tech . Even from a business case perspective, all of our experts agreed that there was no clearly justified and attractive business proposal as there seems to be anaemic industry demand in the EU for such infrastructures. With the EU strategy seemingly being ‘build it and they will come’, it is questionable whether there is even a “mature appetite among startups to become foundation model developers” (Mittal), which is the presumed use case for AI Gigafactories. Likewise, as Saari warned, AI engineers working in European startups are more likely to value efficiency and ease-of-use over ambitious normative goals related to sovereignty. In this light, some policy analysts have argued for carving a coherent and centralised demand-driven strategy that would coordinate and bundle the needs of research labs and start-ups. Pooling demand to leverage the EU’s consumer power and steer it to a direction that is more publicly and socially beneficial is crucial. However, even in this scenario, the EU would be effectively subsiding “ the computing costs of companies that the EU is seeking to reduce its dependency on .” A different approach could be a “federated, modular, and open-source European cloud compute system” (Saari) wherein Gigafactories and providers would pool computing capacity into a shared infrastructure. Still, such a system would be riddled with technical challenges and, perhaps more importantly, political-economic barriers, including potentially competing with incumbents who already do similar work in the EU like NVIDIA. Another critical element that is often neglected is the materiality of the hardware that goes into these infrastructures. Data centres depend on powerful Graphic Processing Units (GPUs), which however age quite rapidly. And since these facilities will probably not be operational – at best – until 2028, hardware procured for them now risks being outdated before they even get online. Due to the meteoric rise of GPUs’ costs owing to Generative AI’s gargantuan uptake, Saari noted, it is likely that a significant amount of the Gigafactories’ budget will go to buying chips that, in the meantime, will have been devalued both in terms of performance but also cost. More importantly, as argued by Ferrari, the EU is betting on the current dominant paradigm of large-scale transformer-based Large Language Models . In doing so, we might be precluding smaller-scale alternatives and other technical architectures, while tying the bloc’s infrastructural future to an industrial model that is extractive and stress-testing our planetary boundaries. Consider, for example, xAI’s “Colossus,” a supercomputer of 200,000 GPUs in Memphis which was built in just 122 days thanks to regulatory facilitations and a complete disregard for environmental harms , massively degrading the lives of people in its vicinity. Is it really in the best interest of the EU to engage with such infrastructures in a race to the bottom? Public money, public interest? Let us not forget how the aforementioned billions are public European money that the Commission is deciding to allocate through investment mechanisms into supercomputing infrastructures. The latter will be regulated through contractual agreements with the private companies that will co-finance the AI Gigafactories. However, contractual agreements and investment mechanisms “do not seem well-suited to supporting an infrastructure-oriented approach that would shape demand and access in a deliberate way” (Ducuing). Instead, they take control away from public institutions and shift public money, resources, and control towards the hands of private entities. Unlike the EU's smaller AI Factories , which are coordinated by the Commission and aimed at supporting scientific research, Gigafactories are massive public-private partnerships. Under the new legal framework , the EU’s contribution is strictly capped at 17% of the capital costs, a figure matched by the host member state, while the private consortium covers the remaining two-thirds of the construction costs (approx. 66%) and all of the operational costs. Accordingly, the private partners will get the lion's share of access to the Gigafactories’ computing capacity. Put simply, the EU becomes a minority tenant in infrastructure using significant sums of public money . The co-financing “reduces the Commission’s ownership and its ability to influence [control over] these infrastructures, fundamentally altering who controls access, how it is used, and the balance between public versus private interests in European infrastructures” (Mittal). In addition, the Commission has been rolling back legal safeguards in the name of competitiveness and innovation. The Digital Omnibus’ simplification/deregulation agenda is looking to limit and delay obligations for AI systems, as well as to redefine scientific research, in a way that “blurs established boundaries, often equating it with innovation” (Ducuing). This risks further diminishing the public-interest use of these computing infrastructures. Reports suggest that the Commission is considering specific environmental obligation carveouts for the Gigafactories, while environmental safeguards, particularly on data centre permits and information regarding their environmental footprint, are already being lowered, serving questionable interests . Democratic control over scale In light of all these challenges, we unanimously agreed on the need for more democratic control. This entails, among others, identifying use cases for AI that serve the public interest, with public institutions and societal stakeholders playing a key role in co-shaping them. Computing made available through the AI Gigafactories could be conditioned on and evaluated against such public interest goals. To that end, stricter legal frameworks for governance and procurement of (super)computing should take into consideration the actual social and environmental needs and concerns . Moreover, slower and more peer-reviewed approaches to developing AI with stronger guardrails should remain possible. Not all applications require increasing scale or intensity, for example in the case of chatbots, so trade-offs can be made. Furthermore, lessons can and must be learned from the past and adapted to the future, for instance in terms of technological diversification. Perhaps not at the scale required for “ Big AI ,” there have been alternatives such as computing grids and peer-based systems. Such solutions won’t be able to compete with hyperscalers, which implies a paradigm shift in what the EU is currently pursuing: we need to imagine alternatives that are attuned to collectively agreed upon needs. At a time when the EU is claiming to pour public money into AI infrastructures to achieve digital sovereignty – evidently a misguided endeavour due to all the aforementioned challenges – performance politics is not sufficient: “the EU is doing Gigafactories so that politicians can say they are doing something on AI. It is a spectacle” (Ferrari). Indeed, the instrumentalisation of the project reflects a broader pattern in the EU’s approach to the digital economy: an agenda that prioritises winning the AI race against the US and China through competitiveness and growth above all. Striving for digital autonomy and sovereignty, however, should be less about boosting the bloc’s industrial competitiveness and more about establishing democratic governance to steer control towards infrastructures that benefit the broader public based on evidence of current needs rather than speculative bets.