EU's new AI factories: a doppelganger network of AI hubs
One of the European Commission's initiatives to boost the continent's competitiveness in artificial intelligence is building public AI factories and public-private AI gigafactories across Europe. These infrastructures are data centers optimised for AI workloads, i.e. they are centralised clusters of GPUs. The Commission's vision is for them to be more than compute: the factories promise to create dynamic ecosystems for frontier AI development, bringing together compute, data, and talent.
A recent analysis by CEPS shows that the locations of the 19 AI factories are mostly outside of the European hubs leading in AI research and innovation. Only 3 of the factories: these in the regions of Paris, Stuttgart and Cologne, are built in 'hubs of AI excellence'. The rest, including the regions of Munich, Eindhoven, and Stockholm, won't have factories. The mismatch in location between the factories and the AI hubs won't, by default, impede the factories from fulfilling their primary purpose, because training AI models can be done remotely. Yet the (non-)existing ecosystems around them matter because innovation for complex technologies is concentrated. It tends to grow where there is existing expertise. Regions without (sufficient) AI capabilities may struggle to reach the critical mass to compete internationally. The AI factories could therefore benefit from being promoted to talent located in other regions (and countries) that are already leading in AI. At a technical level, ensuring seamless and flexible remote access will be essential for their success.
Source: Kyosovska and Renda, 2025
Notes from the referenced paper: the top 20 AI hubs based on the number of scientific publications in AI for 2021 to mid-2025, number of patents in AI for 2021–2024 and amount (in USD) of venture capital investments in AI start-ups for 2021 to mid-2025. The AI ecosystem index is the sum of the normalised (0-100) metrics. The names of the regions (y-axis) are colour-coded to denote any link with a factory: green is a match with a factory; yellow is a region in a country that hosts a factory; and grey is none of these. See interactive version here.
The AI factories must be a truly cloud-based offering, which does not only involve remote access, but inter-operability and federation. Given their relatively small size, pooling resources across sites may be critical for the most ambitious projects. Further, federation can achieve more than boosting the total compute available for a single user: it can be conducive to collaboration between multiple users. This can unlock a significant amount of potential for international cooperation. The dynamics of innovation show that more integrated networks of collaboration lead to higher success. The European innovation network is significantly fragmented, and more so than the US one, so siloing the AI factories could only exacerbate this issue.
Source: Kyosovska & Renda, 2025
Notes from the referenced paper: the metric shows how many patents were more or less co-created as a proportion of the expected number; for example, if the regions of London and Paris were partners in 20 patents, and the expected number was 10, the difference in RCA would be 20/10 – 1 = 1. A value of -1 means there were no observed collaborations; a value of 0 means expected and actual were the same. The positive values are in green, while the negative ones are in red. The lowest possible is -1, as collaborations cannot be negative. See interactive version for a) here and b) here.
The narrative around the AI factories is centered around their function as new ecosystems of AI excellence. The data on the structure and dynamics of the current network of AI R&I in Europe suggests that the strategy should refocus on supporting the current community and on building bridges.