A central part of Europe 2031 concerns compute: the AI data centres and chips required to train and run frontier AI systems. We believe compute is rapidly becoming one of the world’s most important strategic resources, because the amount of compute a country or continent controls increasingly determines its ability to build advanced AI systems, deploy them at scale, capture the resulting economic gains, and maintain geopolitical leverage. If a continent has 50% of the world’s global compute, a company or country will definitely think twice about restricting your access.
To ground the scenario in concrete assumptions, we developed detailed forecasts for how AI compute capacity in Europe, the United States, China, and the rest of the world could evolve through 2031, both under Europe’s current trajectory and under a far more ambitious industrial mobilisation scenario. Our default estimate suggests Europe is currently on track to host only around 3-5% of global AI compute by the end of the decade, compared to roughly 70% in the United States and 15% in China. The following sections outline the assumptions behind these projections.
Our forecast is anchored in projections by the AI Futures Project, to be published soon as part of their forthcoming AI 2030 scenario. These projections are an update to their previous compute forecast from April 2025, which estimates the global stock of AI compute based on public information about supply chain bottlenecks. Their reasoning is described here and does not presuppose their predictions about future AI progress.
1. Global AI compute
The AI Future Project’s most recent forecast estimates global AI compute to reach 300 GW by 2030. We apply a logistic fit to extend this estimate by another year, yielding 373 GW in 2031.
This is well within the upper bound of what the world’s EUV lithography machines – the most complex tool required to produce cutting-edge chips – would be physically capable of producing by 2031. Building on a recent estimate by chip researcher Dylan Patel, we assume that by 2031, there will be around 700 EUV machines, in theory capable of producing another 780-870 GW of AI compute from 2026 onwards.
In practice, the effective output will be much less, as there are bottlenecks to chip production other than EUV (e.g. high-bandwidth memory or advanced packaging) and some EUV machines are used to produce non-AI chips. Assuming that 50% of all EUV capacity goes into AI chip production, these machines would produce 390-430 GW of AI compute between 2026 and 2031, slightly above our estimate of 373 GW.
2. Relative shares in the default branch
The AI Futures Project assumes that, in 2026, 87% of global AI compute is in the US and 11% in China, leaving around 2% for the rest of the world, including Europe. By the end of the decade, the US share decreases slightly to 84% while China’s share approaches 16%, effectively crowding out all other world regions.
Assuming a Chinese share of 11% today is plausible and fits the recent 11.5% estimate from Epoch AI’s chip owners database (when measured in GW)[1]. We also follow the AI Futures Project in assuming that China’s share will increase to around 15% by 2031. However, we depart from their forecast in assuming that the Western compute stack will be entirely in the US.
Instead, we think that Europe’s share is around 5% today, which will briefly increase to 8% by 2028 before decreasing again to slightly below 5% in 2031. This follows a bottom-up estimate of European compute buildout over the next 5 years, based on public information about existing and announced AI data centers. Our database of publicly known AI data centers in Europe shows that its compute base will likely reach 2.3 GW by the end of 2026 and over 17 GW by 2031.
While European AI compute is mostly concentrated in the Nordics today, our data indicates a shift by 2031, with the three largest compute owners by then being France (24%), Norway (11%) and Germany (10%). The largest single AI data centers by 2031 include GW-scale projects in France and Portugal, and >500 MW facilities in the Netherlands, Romania, Norway, Germany, Italy and the UK. Our default estimate assumes that privately and publicly announced projects will mostly be realized and that European compute policy retains its current level of ambition.
We assume that, by default, the compute share of the rest of the world (ex-US/China/Europe) will exceed Europe’s share and reach around 10% by the end of the decade. GW-scale announcements in several countries – UAE (5 GW), Saudi-Arabia (1.5 GW, 1 GW, 1 GW), India (3 GW, 1 GW, 1 GW), South Korea (3 GW, 1 GW), Canada (2.4 GW, 1.2 GW) and Brazil (1.5 GW, 1 GW) – alone make up a total of over 23 GW, thereby exceeding Europe’s share of 17.5 GW by 2031. Even if multi-GW projects face higher execution risk than in the US or Europe, the rest of the world’s share is still very likely to be greater than in Europe, especially when taking into account that several data centers in the multiple-hundred MW range are also planned in countries such as Australia, Malaysia or Japan.
If China has around 15% of global compute, Europe has 5% and the rest of the world (ex-US) has 10%, this leaves around 70% for the US in the default scenario.
3. Relative shares in the retrospective branch
The retrospective branch is our account of what happened in a world where Europe successfully changed course on AI, including by pursuing a much more ambitious compute policy. In such a world, we assume that Europe would aim to reach a 15% share of global AI compute by 2031, which would triple its current 5% share. In the medium term, it would put Europe on track to reach a compute share roughly proportional to its global GDP share (~25% currently).
Since building large AI data centers has multi-year lead times even in optimistic scenarios, most of this additional capacity would come online in 2030/2031, assuming Europe decided in 2026 to onshore a greater compute share. Building enough data centers would be a European effort, making use of suitable sites and grid connections wherever they are available, though compute would plausibly be concentrated in regions with stable, fossil-free energy sources: the Nordics (hydropower), France (nuclear) or Iberia (solar). Countries with decommissioned industrial areas, such as Germany or Poland, will also play a major role, as some of these sites have existing, GW-scale grid connections that could be repurposed for large AI data centers.
Hosting more AI compute in Europe likely won’t affect the world’s total compute capacity: the supply chain for AI compute is already churning out chips at maximum capacity, irrespective of where they end up. The difference to the default branch thus only concerns relative shares. We assume that an ambitious European compute strategy would mostly reduce the US share (by around 9pp), as Europe makes it more attractive for US AI companies to invest and leverages its supply chain position to obtain more US-made chips for sovereign compute projects in Europe. China’s share will also decrease, as more US chip exports will go to Europe rather than China, though only to a small degree (around 1.5pp), as China is actively decoupling its compute stack from the West. Finally, the rest of the world (ex-US/China) will also have a slightly lower share, as chip exports that would have otherwise gone to countries like the UAE or Saudi-Arabia would go to Europe instead.
- Note that measuring compute in GW rather than H100-equivalents overstates China’s compute share, as Chinese hardware is less power-efficient than US alternatives. For example, when measured in H100-equivalents, China’s current share of global compute is just around 5%, compared to 11.5% when measured in GW. We still use GW as a metric for two reasons. First, it’s the more relevant metric for policymakers, who need to think about AI compute buildout in the broader context of energy policy and grid planning. Second, we’re mostly interested in how the relative shares of the US and Europe shift in the two scenarios. Here, GW is perfectly fine as a metric, since the US and Europe rely on the same, equally power-efficient hardware.