Europe’s AI dependency problem is an operating-model problem, not only a compute gap

Europe’s AI debate is moving from competitiveness to operational dependence. In a Vienna speech reported by Reuters, European Central Bank President Christine Lagarde warned that imported AI could become leverage across borders, healthcare, banking, transport and public administration if access or commercial terms changed.

Lagarde’s prescription has three parts: build more European computing capacity, develop models that are “good enough” for most tasks and run them on European infrastructure, and adopt AI fast enough to capture productivity benefits. She said Europe’s data-centre capacity gap could grow more than sixfold within a decade and cited a possible productivity-level gain of up to 4% over ten years if adoption is rapid.

Those are scenario claims, not forecasts that every organisation can bank. Reuters’ report does not reproduce the underlying capacity model, assumptions behind the sixfold gap or the productivity methodology. The speech is a strategic intervention by a central-bank president, not an engineering plan or causal evaluation.

Map dependency by workflow

Compute location is only one layer. A European-hosted application can still depend on overseas model weights, orchestration software, identity services, safety filters, developer tooling or specialist talent. Conversely, a service supplied from abroad may have strong export, interoperability and continuity provisions.

Boards should therefore ask where a critical workflow could fail if a provider changes price, access, licence, export policy or product direction. The inventory should include the model, data store, embedding and retrieval layers, evaluation assets, logs, tool credentials and the people who can operate an alternative.

The result should be a portability test, not a flag on an architecture diagram. Can the organisation export prompts, configurations, evaluation cases and audit evidence in usable form? Can it move a representative workload to another approved model within a defined recovery time? Which quality losses are tolerable, and which language, safety or regulatory requirements break?

Capacity without capability can disappoint

Building infrastructure can expand strategic choice, but it does not automatically produce competitive services or adoption. The scarce complements may include power, network connections, finance, high-quality data, model engineering, domain evaluation and change capability inside user organisations. Training plans should be connected to workloads that regional infrastructure is intended to support.

Lagarde also argued that Europe already bears part of the cost: US technology firms borrow in European markets, and European pension funds hold US technology shares. That macro-financial framing is important, but it does not show that a specific data-centre programme will improve resilience or productivity. Investment decisions still need demand, energy, location and workforce evidence.

The counterargument is straightforward: forced localisation can raise costs, slow access to the best tools and fragment standards. Resilience should therefore be proportional. Critical public and regulated workflows may warrant tested alternatives and local control; low-risk, easily replaceable uses may not.

The practical decision is to separate sovereign capability from symbolic location. Use the Skills Atlas to identify operating and evaluation gaps, then run exit exercises against real workflows. European compute is useful when organisations can deploy it, measure it and switch to it. Without those complements, a larger regional footprint may still leave the decisive capability elsewhere.

A minimum evidence package

Before scaling the change, the responsible team should preserve the exact source, model or policy version, the affected workflow, baseline, decision owner and review date. It should state what would count as success, what would count as a material failure and who can stop the use. Results should separate technical performance from adoption, business outcome and distribution across affected groups. Where evidence is incomplete, the scope should remain bounded and reversible. This discipline does not decide the policy or product question in advance. It makes the next decision auditable and allows a later reviewer to distinguish new evidence from a changed assumption. The organisation should also retain an accessible human route for challenge whenever the system materially affects work, opportunity or rights.