RAND Europe’s response to the EU Cloud and AI Development Act consultation is a compelling and at times blunt strategic diagnosis of the European Union’s position within the global AI race. It offers a vision that moves away from romanticised ideals of full technological sovereignty and instead embraces the colder, more transactional logic of strategic indispensability.
At the heart of the analysis is the “compute gap” — a term that encapsulates Europe’s lag in access to the massive computational resources required to train and run advanced AI systems. As of early 2025, the EU holds a mere 4.8% of global AI supercomputing capacity, in stark contrast with the United States’ commanding 74.4% and China’s 14.1%. This asymmetry, far from being a statistical oddity, is interpreted as a structural disadvantage with cascading implications: economic, strategic, technological, and normative.
The authors dissect these implications with clarity. The inability to access sufficient compute is not simply a bottleneck; it risks relegating European firms to the role of consumers rather than creators of cutting-edge AI. This in turn undermines competitiveness, hinders technological diffusion across the single market, and strips the EU of the leverage necessary to shape global AI norms. More provocatively, RAND warns that compute access could become a geopolitical bargaining chip in an era where AI capabilities are becoming synonymous with state power — a point illustrated by references to possible future Artificial General Intelligence (AGI) scenarios, which may concentrate power in the hands of a few compute-rich actors.
Rather than propose a utopian vision of European self-sufficiency “across the stack” — from chip design to model development — the paper urges a pivot towards what it terms “strategic indispensability.” That is, Europe should build enough AI-critical infrastructure to ensure it cannot be easily excluded from the global value chain. The location of compute becomes a point of leverage in itself. If Europe hosts data centres at scale, even if powered by US-designed chips or models trained elsewhere, it gains legal jurisdictional control, operational resilience, and a stake in future negotiations over access, safety and governance.
This logic underpins RAND’s recommendations, which are ambitious but framed in a starkly pragmatic tone. First, the EU should set compute targets relative to global benchmarks (e.g. 15% of global AI compute capacity by 2030), rather than merely aiming to triple its current capacity — a goal judged insufficient and internally oriented. Second, the authors propose the creation of “Special Compute Zones” — expedited permitting environments in energy-rich regions such as the Nordics or France, which would accelerate infrastructure deployment and circumvent the EU’s notoriously slow and fragmented approval processes. Third, they call for robust security standards — including RAND’s own tiered framework — to be applied to AI data centres, especially to those exceeding 50MW, given the rising threat from nation-state adversaries and the strategic value of these assets. Finally, the importance of inference compute (as opposed to training compute) is highlighted as an area where Europe can build scalable, broad-based AI deployment capacity without directly engaging in the highest rungs of the arms race for frontier model training.
One of the more striking aspects of the paper is its willingness to critique some of the foundational assumptions behind the EU’s proposed Act. For instance, the belief that simply localising infrastructure (i.e. building data centres on EU soil) will deliver “tech sovereignty” is described as misleading. Without control over upstream technologies — like advanced chips, many of which are designed and produced outside Europe — localisation may not yield real autonomy. Worse, it risks being performative. Instead, the RAND authors suggest that Europe use what leverage it can realistically build: chips located within EU territory can, by their physical presence, be partially subjected to EU law and policy, especially if operated under stringent security and governance standards.
Yet, the paper also exhibits a sense of constraint — an awareness that its recommendations, while technically feasible, are politically and institutionally challenging. It acknowledges, for example, that any EU-wide strategy will require coordination among diverse member states, some of which may be reluctant to cede permitting power or adopt a harmonised approach. It also admits that current trends in AI architecture (e.g. more efficient algorithms) could potentially reduce the raw compute requirements over time, complicating long-term infrastructure investment plans.
Still, the underlying message is unambiguous: if Europe wishes to retain any credible claim to leadership in the AI era — not just regulatory but also technological and strategic — it must act now, decisively, and at scale. There is a sense of urgency throughout the report, tinged with the recognition that opportunities missed today may not be recoverable tomorrow.
Ultimately, RAND’s intervention is not a manifesto but a realpolitik-infused roadmap. It avoids grandiose rhetoric and focuses instead on infrastructure, bargaining power, and the logic of leverage. It does not lament Europe’s past failures, nor does it promise that the proposed measures will guarantee leadership. What it does argue — with precision and conviction — is that without a massive and well-coordinated investment in compute, Europe will find itself increasingly dependent, peripheral, and vulnerable in the AI century.