On July 30, the EuroHPC Joint Undertaking published a call for tenders to select consortia that will build and operate up to seven AI Gigafactories across the European Union. The submission deadline is November 12. Award decisions are expected in early 2027. Once contracts are signed, successful consortia will have eighteen months to complete construction. The language is precise and the timeline, by European institutional standards, is ambitious. It is also the most concrete expression to date of a bet Europe has been placing for the past two years: that sovereign computing infrastructure, publicly anchored and privately scaled, can close the gap between European AI ambition and American and Chinese AI reality.
The gap is real and the numbers that describe it are not comfortable. Europe currently hosts roughly 5 percent of the world’s AI computing capacity. The United States hosts around 35 percent. China, depending on the metric used, is somewhere between 20 and 30 percent. The existing network of EuroHPC AI Factories — twenty-five facilities distributed across the continent, from Barcelona to Sofia, from Athens to Kajaani in northern Finland — contains three of the top ten supercomputers in the world by raw performance, including Jupiter in Jülich and LUMI in Kajaani. What they do not contain, in sufficient quantity, is the specific GPU cluster architecture that training and running large-scale AI models requires. Supercomputing performance and AI training capacity are related but not identical, and Europe has historically invested in the former while the latter grew at a pace its existing infrastructure was not designed for.
The Gigafactory procurement is structured in two tiers. Four medium-scale facilities will each be eligible for up to €100 million in EU funding in Phase I, with the requirement that each deploys at least as many advanced AI processors as are currently installed in Europe’s most powerful AI factory. Phase II unlocks an additional €400 million per project, contingent on tripling capacity. Three larger facilities receive up to €200 million in Phase I and up to €800 million in Phase II. The joint public funding from the EU and member states is designed to function as an anchor customer, de-risking the capital required and unlocking expected total private investment of more than €20 billion across the EU. Eighteen member states — including Germany, France, Italy, Spain and Poland — have already signed a joint procurement agreement with EuroHPC JU. Consortia may source hardware from European or allied suppliers, and the Commission has signed letters of intent with AMD, Nvidia and Qualcomm to ensure hardware access following the EU-US trade deal.
That last detail is worth pausing on. The letters of intent with AMD, Nvidia and Qualcomm are a tacit acknowledgment of something that European technology sovereignty advocates have not always been willing to say plainly: the chips that will run Europe’s sovereign AI infrastructure are, for the foreseeable future, going to be made in America or by American-designed processes in Taiwan. Officials are realistic about this. Future hardware upgrades are expected to involve a larger number of suppliers as competition increases, but chips alone will not solve the problem. Europe’s semiconductor ambitions — the European Chips Act, the ASML ecosystem, the nascent efforts to develop European AI accelerator chips — are moving on a timescale measured in decades. The Gigafactories will be operational, if all goes to plan, in 2028 or early 2029. The processors inside them will be Nvidia H-series or AMD Instinct parts, not European-designed silicon.
This is not a criticism of the initiative. It is a description of where the initiative sits in the longer arc of what European technological sovereignty actually requires. The Gigafactory programme is solving a real and urgent problem: European researchers, startups, universities and public institutions currently have inadequate access to the compute capacity needed to train, fine-tune and run competitive AI models. With 76 expressions of interest already submitted from European cities and regions hoping to host one of the seven facilities, the demand for this infrastructure is not in doubt. The question is whether infrastructure, in isolation, is sufficient.
Henna Virkkunen, Executive Vice-President for Tech Sovereignty, Security and Democracy, described the call as “a milestone in our AI Continent ambitions,” and said that “access to the raw scale of computing power within AI Gigafactories is a strategic necessity for Europe as AI development accelerates.” The framing is accurate. It is also, inevitably, partial. Raw computing power is necessary but not sufficient for AI leadership. What converts compute into capability is talent, data, software ecosystems, and the capital structures that allow the combination of those ingredients to be assembled and sustained over the years required to build genuinely frontier models. Europe has the first of these in greater abundance than is sometimes acknowledged — its universities produce excellent AI researchers who then, in large numbers, move to American or British labs where salaries and resources are better. It has significant advantages in the second — European data, particularly in healthcare, public administration and industrial processes, is among the richest and most structured anywhere. It has a software ecosystem that is technically sophisticated but commercially fragmented. And it has capital markets that are improving but remain structurally less willing to fund long-duration, high-uncertainty bets than their American counterparts.
The Gigafactory programme addresses the compute gap. It does not address talent retention, data governance, software fragmentation or capital access. Those require different instruments, most of which are either under construction or still in political negotiation. The AI Act’s transparency rules entered force on August 2 — the same week the Gigafactory tender was published — and the full enforcement architecture comes online this month. The Cloud and AI Development Act aims to triple European data centre capacity over seven years. The Frontier AI Grand Challenge, launched in February, will fund one project to train a frontier AI model. Part of the Gigafactory financing will come from the current MFF; the bulk will depend on the 2028-2035 budget framework, whose negotiations are still ongoing.
That last dependency is not a minor administrative detail. The most capital-intensive phase of the Gigafactory programme — Phase II, where each of the seven facilities receives between €400 million and €800 million in additional EU co-funding — is contingent on a budget that has not been agreed. The October European Council is the next substantive milestone in MFF negotiations. The gap between the ambition expressed in the Gigafactory tender and the financial reality of a budget framework still being argued over in the Council is one that the programme’s architects are managing carefully and not discussing loudly.
There is a broader question about the initiative that the technical literature on AI infrastructure raises but that the Commission’s communications do not. AI capabilities are currently advancing at a pace where the hardware assumptions embedded in a 2026 tender may be substantially different from the hardware reality of 2028-2029, when the facilities are expected to become operational. The AI processors that will be installed in the Gigafactories are the best available today. Nvidia’s next-generation architectures, and whatever follows them, will be available before the construction is complete. The procurement framework includes provisions for hardware upgrades, but the capital-intensive nature of large-scale compute infrastructure means that yesterday’s frontier cluster can become today’s mid-range facility faster than anyone planning a multi-year construction programme would prefer.
None of this makes the programme less necessary. The alternative — continuing to rely on American hyperscaler cloud infrastructure for the computing that trains and runs European AI — carries its own risks, which the Anthropic hearing at the European Parliament three weeks ago illustrated with some precision. A continent whose AI research, whose public sector AI applications, and whose competitive industrial AI capabilities all depend on AWS, Azure and Google Cloud is a continent whose AI autonomy is a function of the commercial terms those providers choose to offer and the export control decisions the US government chooses to make. The Gigafactories are a structural hedge against that dependency. Whether they are sufficient to build the kind of AI capability that makes European autonomy meaningful in practice — rather than simply in procurement documents — is a question that will not be answered by the tender’s November deadline.
The facilities are expected to be located in at least seven member states, with the possibility of some operating as cross-border collaborations that deploy infrastructure across national boundaries. The geographic distribution question will, when the award decisions arrive in early 2027, be as politically charged as any technical assessment of the proposals. Large member states with existing HPC infrastructure — Germany, France, Spain, Italy — will have strong bids. Smaller member states that have invested in their EuroHPC footprint — Finland, Luxembourg, Bulgaria — will have arguments about distributed resilience. The Commission will have to balance technical merit against the political reality that seven facilities distributed across eighteen co-funding member states must be distributed in a way that is defensible to all of them.
That is, in microcosm, the challenge that European AI policy as a whole faces: building something genuinely excellent within a political architecture that requires it to be genuinely inclusive. The tension between those two imperatives is not unique to AI. It runs through every major European infrastructure project ever built. Managing it is what European institutions spend most of their energy doing. The Gigafactory programme is the latest and most consequential test of whether they can manage it fast enough for it to matter.
Subscribe to EuroLuminant for independent European journalism.
