Europe's AI compute gap is a grid problem, not a money problem
The EU holds only around two gigawatts of AI compute — roughly five percent of the global total — and Bruegel argues the bottleneck is not capital, but permitting speed and grid access. Here is the diagnosis, the proposals, and what Finnish practice shows about how the constraints play out in reality.
The number that frames the debate
Bruegel, the Brussels-based think tank, published a policy brief by Bertin Martens and Tillman Schenk on 10 September 2026 with a stark finding: the EU has only around two gigawatts of AI compute, equivalent to five percent of the global total, and planned projects will barely change the EU's global share (Bruegel Policy Brief 18/2026).
The figure rests on the Europe2031.ai dataset (Juijn et al., 2026) — a list of 100 investment projects in AI infrastructure of at least ten megawatts. Notably, such datasets simply did not exist before 2026. This is the first comprehensive dataset on EU AI compute, according to the brief. Bruegel also points out itself that there is very little transparency about Chinese AI infrastructure, so the cross-regional basis for comparison carries real measurement uncertainty. Five percent is an estimate, not an exact fact — but it is the first attempt at a comprehensive mapping of European AI compute.
A lag across the entire stack
Compute capacity is not an isolated problem. According to the brief, the EU trails the US and China across the entire vertical AI stack, from chips and infrastructure to models and applications. At the model layer, the EU accounts for barely any of the more than 100 notable AI models (figures via Epoch AI), and has only one significant model developer: Mistral. On the chip side, global AI chip production is dominated by American companies, primarily NVIDIA, with China's Huawei on the rise.
This means Europe's vulnerability cannot be solved at one layer alone — but Bruegel's point is that at one of those layers, the barrier is unnecessarily narrow.
The capital exists. The power and the permission do not
The core of the brief's argument is that private capital is not the decisive constraint on the build-out of European AI compute. It is time to operation and grid access that bind.
The Commission has attempted to address the permitting side through the proposed Cloud and AI Development Act (CADA). But according to Bruegel, grid access has been kept outside CADA's scope — and that is precisely where, in the connection to the power grid, the real bottleneck sits. It is a specific, technical move in the legislative proposal with large consequences: even if permitting becomes faster, projects may still be queuing for power.
Bruegel's recipe
The brief proposes three main measures:
Redirected gigafactory subsidies into competitive grants. Regions would compete on permitting times — those that are fastest receive larger subsidies. The point is to make speed a competitive reward rather than a procedure everyone completes at their own pace.
Anticipatory grid investments. Grid build-out must be prioritised before demand is fully concrete — along with faster grid connections. This is the area CADA does not cover, and which Bruegel argues must be handled elsewhere in the regulatory framework.
Sovereignty on the output side, not the input side. Instead of strict local-content requirements — of the kind some proposed CADA sovereignty criteria represent — Bruegel proposes placing the conditions on the services side: voluntary "processed in the EU" labels on services sold to European customers, and requiring hyperscalers to disclose how many AI chips they have installed in the EU and how much workload is computed there. The EU could, according to the authors, incentivise American hyperscalers to install a larger share of their chip supply in European data centres.
The warning is equally concrete: strict local-content criteria could, according to Bruegel, increase costs and trigger a negative supply response — meaning providers would simply place capacity elsewhere.
The regulatory landscape Bruegel enters
The brief landed in the middle of an active policy period. On 16 September 2026, the European Commission, according to HackerNoon's analysis "The Geopolitics of AI in Q3 2026" (based on a quarterly compendium on AI geopolitics), welcomed an initiative from 19 member states to establish the first IPCEI project for AI — an "important project of common European interest" covering the entire AI stack and explicitly linked to compute governance, industrial products, applications, adoption, and European technological sovereignty. These are secondary sources, not documentation directly from the Commission, but they show the terrain in which the debate is taking place: EU awareness of compute as strategic infrastructure is established at the top level, while the bottleneck side — the grid — still lacks a dedicated legal act.
Finland as a microcosm
What does a national response look like in practice? Finland offers an example — with the caveat that the source material here comes through a single secondary channel.
Finland's power company Fingrid stated, according to Market Briefs (23 October 2026, citing Fingrid's own website), that the country's electricity consumption could rise by up to 40 percent in five years. As of mid-October, data centre projects of more than five gigawatts had secured formal grid connection agreements, including distribution-level connections. That stack slightly exceeds the roughly 4.4 gigawatts from Finland's five reactors. It illustrates both how large the pressure on the grid actually is, and why Bruegel's emphasis on grid access hits something concrete.
At the same time, the Finnish government moved on the sovereignty side: it is preparing legislation to prioritise grid access for critical infrastructure, and the same week it proposed tightened control of foreign direct investment — with permit requirements for, among other things, data centres of 100 megawatts and above, as well as AI and quantum data processing. That is precisely the tension Bruegel points to: faster connection on one side, stricter screening on the other.
What remains
It is worth underlining what this is: one think tank's argument in an ongoing debate, not established consensus. Bruegel's estimate rests on a dataset published the same year, and the comparison with China and the US carries significant uncertainty. The IPCEI reports and the Finnish figures come from secondary sources that are not independently confirmed in the available material.
The open questions are therefore concrete: Will CADA be extended to cover grid access, as Bruegel argues? Will competition over permitting times actually deliver faster build-out, or simply move the bottleneck one step down the process? And will voluntary "processed in the EU" labels be enough to shift hyperscalers' investment decisions — or will local-content requirements, with the costs Bruegel warns against, end up the politically most likely outcome anyway?
What does remain after reading is this: if the EU truly wants to change its five percent share, the battle is probably not over the budget — but over how quickly the power can be connected.