What is sovereign AI? The European version explained

The result is a term that sounds like consensus and conceals profound strategic disagreement. To take it seriously, you need to know which version is on the table.

Definition one: data residency

The oldest and most legally grounded version of sovereign AI is essentially a GDPR argument with a generative twist. Under this reading, sovereignty means that the data used to train, fine-tune, or run inference on an AI model is stored and processed inside a defined jurisdiction, typically the European Economic Area or a specific member state. The model itself can be foreign-built and foreign-owned; what matters is that personal and sensitive data never touches infrastructure outside the zone.

This is the definition baked most firmly into the EU Cloud Code of Conduct, the industry self-regulatory framework administered by SCOPE Europe and audited against GDPR requirements. The Code of Conduct covers data location, access controls, and portability obligations, and its AI-related provisions treat sovereignty primarily as a data-handling question. For public-sector procurement officers in Germany, France, or the Netherlands, this is often the only definition that maps onto existing compliance frameworks, which explains its institutional durability.

Its weakness is obvious: a model trained entirely on American or Chinese data, by American or Chinese engineers, on American or Chinese hardware, can still satisfy data residency requirements if the inference endpoint sits inside an EU data centre. Critics, including the Joint Research Centre at the European Commission, have noted in their AI Watch reporting that data residency alone does not address dependency risks in the model layer or the compute layer. It is, in that reading, sovereignty for lawyers rather than sovereignty for engineers.

Definition two: model residency

The second definition shifts the locus of control from data to the model itself. Under this view, sovereign AI requires that a foundation model be developed, trained, and owned by entities inside the target jurisdiction, with weights that are accessible to that jurisdiction without a foreign intermediary's permission. Licensing a frontier model from an American hyperscaler and running it on domestic infrastructure does not count; building your own does.

This is the definition that Mistral AI in Paris and Aleph Alpha in Heidelberg have most consistently championed in their public policy positions, and for obvious commercial reasons: both companies produce European large language models and both benefit directly from procurement rules that treat model origin as a sovereignty criterion. Mistral has argued explicitly in European Parliament briefings and in public commentary that the ability to audit, modify, and redistribute model weights is a precondition for genuine AI independence. Aleph Alpha's founders have made comparable arguments about the need for European nations to control the "cognitive infrastructure" underpinning public administration.

The model residency definition gained significant political traction during the negotiations over the EU AI Act, where several member state delegations pushed for provisions that would favour domestically developed systems in high-risk public-sector applications. Those provisions were diluted in the final text, but the underlying logic persists in national AI strategies from France's national AI strategy updates to Germany's AI action plan published by the Federal Ministry for Economic Affairs and Climate Action.

Definition three: talent residency

Less frequently named but arguably more structurally important is the talent residency definition, which holds that sovereign AI requires a self-sustaining domestic workforce capable of building, maintaining, and improving frontier systems independently. Under this view, a European company that retains its top researchers in Europe, trains successors through European universities, and does not depend on talent pipelines from outside the continent is genuinely sovereign, even if some of its infrastructure is foreign-supplied.

This definition is associated most strongly with academic and research institutions. The European Laboratory for Learning and Intelligent Systems, known as ELLIS, has consistently argued that talent is the binding constraint on European AI independence and that policy should prioritise fellowships, salary competitiveness, and research infrastructure over data-centre nationalism. The Institute for Data and AI Policy, operating in the Brussels policy ecosystem, has similarly emphasised that hardware and data advantages erode quickly without the human capital to exploit them.

Talent residency is the definition most uncomfortable for European governments to act on, because the interventions it demands, such as substantially raising researcher salaries in public institutions, competing with American and British tech salaries on compensation, and reforming immigration rules for non-EU AI specialists, are expensive, politically fraught, or both. It is also the definition that gets talked about least in procurement documents, which is precisely why critics argue European sovereignty policy is being built on sand.

Definition four: compute residency

The fourth definition has risen fastest in political prominence since 2023 and focuses on the physical infrastructure layer: the chips, the data centres, and the interconnects on which AI systems run. Under this view, sovereign AI requires that European actors control sufficient compute capacity to train and run frontier models without dependence on foreign-controlled cloud providers or foreign-manufactured semiconductors.

Gaia-X, the Franco-German cloud infrastructure initiative, was conceived partly around this logic, aiming to create a federated European data and compute ecosystem governed by European rules. Its policy publications have consistently framed European cloud and compute infrastructure as a sovereignty prerequisite, not merely a commercial preference. The European High Performance Computing Joint Undertaking, EuroHPC JU, reflects the same instinct in the public sector, investing in supercomputing facilities across member states specifically to reduce dependence on non-European compute for research and, increasingly, for AI training workloads.

The compute residency definition gained further momentum from the global chip shortage of 2021 and 2022 and from the United States' export controls on advanced semiconductors, which reminded European policymakers that access to the hardware underpinning AI is itself a geopolitical variable. ASML in the Netherlands, Europe's only manufacturer of extreme ultraviolet lithography machines and a critical node in global chip supply, features prominently in this conversation, though the company sits upstream of finished AI accelerators rather than producing them directly.

## By The Numbers

The scale of the challenge across all four dimensions is clarified by the numbers. European compute investment, model development funding, researcher retention rates, and data infrastructure spending each tell part of the story of where the continent actually stands relative to its sovereignty ambitions, and the gaps remain significant despite several years of policy attention.

Which definition is winning?

Honestly, the model residency definition is winning the political argument right now, primarily because it has the most organised commercial lobby behind it and because it maps most cleanly onto the procurement and industrial policy instincts of European governments. When France's government invests in Mistral or when Germany's public entities prefer Aleph Alpha's solutions in pilot programmes, they are operationalising model residency, regardless of what their official policy documents say about data protection or compute infrastructure.

That is not necessarily the right answer. A sovereign model built on foreign compute, trained partly on foreign data, and maintained by a team that could relocate to San Francisco next quarter is sovereign in name only. The Joint Research Centre's AI Watch analyses have repeatedly flagged that single-layer definitions of sovereignty create false reassurance, and their scepticism deserves to be taken seriously by policymakers who are currently writing cheques on the basis of model provenance alone.

The most rigorous version of European sovereign AI requires progress on all four dimensions simultaneously, and the hard truth is that Europe is behind on three of them. Compute capacity is improving but not yet sufficient for frontier training runs. Talent retention remains a genuine crisis, with salary differentials between European research institutions and American tech companies still measured in multiples rather than percentages. Data governance frameworks are mature but increasingly inadequate for the scale and complexity of modern AI training.

Model residency is the one dimension where European companies have made genuinely competitive progress, which is why it dominates the conversation. But letting the strongest lobby define the terms of a national security and industrial policy question is a habit that tends to produce outcomes optimised for the lobby rather than the nation.