Last week, a 5-year scenario about AI and Europe’s future was published and it could not have been more timely. The piece, titled Europe 2031, outlines how Europe could slide into global irrelevance, being left without leverage in a new AI-powered world where the US, and perhaps China, hold the only meaningful geopolitical say. Three days later the US government pushed Anthropic to block access to its most capable models (Fable & Mythos 5) for all foreign nationals - resulting in the first step backwards on the publicly accessible frontier of Artificial Intelligence. In the already gloomy Europe 2031 narrative, such US-imposed, country-based restrictions to the AI frontier only begin in 2029. It could well be that the current export control based framing of the Fable 5 ban was only the tool at hand to block the model entirely for now, but the ban imposed by the US government makes painfully clear how dependent Europe has become on the US in terms of AI.
Europe 2031 and previous analyses argue that the major driver for even further dependence is the divergence of AI computing power. For their scenario-building, Europe 2031 mapped all major AI compute factories that are currently planned to go online by 2031. One part of the story is that Europe is projected to stay at around 5% of the projected global AI compute capacity in terms of power, with a small temporary bump up to 8% in 2028. The other part of the story is how that compute build-out is spread over Europe. Looking at the top 7 AI compute builders in Europe (see graph above), Norway and other Nordic countries are currently leaders of this continent, but that picture changes clearly by 2031. If France executes on its current plans, the country will be home to around 2.4 GW of AI compute capacities - that would be 35.3% of the whole continent and as much as Norway, Germany, Portugal and Sweden together.
Of course, projections like this always come with limitations and the authors of Europe 2031 clearly acknowledge that it is very likely that the AI compute that actually is built by 2031 will differ from current official plans, but it is also a fact that building giant AI factories is a venture that needs ambition and commitment. The graph below shows the 101 AI compute projects in Europe that are publicly known.
Looking at these specific projects is helpful to realise that a few big and ambitious projects are driving a large share of the overall projection. Large projects such as the Softbank data centre and Mistral Campus in France often make up the majority of a country’s compute capacity and also across Europe the top 10 projects are responsible for more than half of the total planned capacity. Measuring AI compute capacity is hard and megawatts are only one proxy for data center performance, while most of the AI compute in 2026 will be outdated for training large AI models by 2031. Additionally, the details of many projects such as the EU Gigafactories still have to be determined and it could for example be that the Volt AI Gigafactory in The Netherlands will become one of the EU Gigafactories which could lead to a double-count in the forecast as of now.
However, the data still clearly shows very relevant data in terms of ambition. Within Europe, France clearly leads there and could serve as an example for other nations such as Germany. But globally the US still dwarfs the AI compute plans of the European continent with around 70% of global AI compute by 2031.
So what can be done? Europe 2031 recommends massive investments in AI compute and its supply chains. This needs simpler permitting procedures, energy policy and special economic zones. They recommend partnerships both with American compute providers for building infrastructure as well as with middle power governments on a geopolitical level. There are many details to be filled in here and time seems certainly to be running out.
The recently proposed EU Cloud and AI Development Act (CADA) goes a step in that direction by defining a structured sovereignty framework. The act aims to reduce supply chain vulnerabilities, but cannot promise it and as always risks being too little too late. The numbers of the currently planned AI compute build-out show that the US is on track to increase its computational AI power by 15-fold by 2031. This means that even the planned 15-fold AI compute increase in Europe will not reduce the relative gap, while the absolute gap is becoming larger and larger.
Scenarios like Europe 2031 and the Fable incident show why it is imperative to have less compute dependence, but actually delivering the necessary build-out in Europe will be an incredibly hard and ambitious task that needs serious commitment.