EU shifts from regulating AI to mandating its adoption across member states

A legislative shift of considerable consequence passed largely without public debate last June. The Cloud and AI Development Act, published as the centrepiece of the European Commission's Tech Sovereignty package, does not merely regulate artificial intelligence. It compels member states to build the infrastructure and institutional capacity that will make large-scale deployment inevitable.

Titles II and III of the regulation read differently from almost any previous EU technology law. The obligations are not addressed to the companies that train models or operate platforms. They are addressed to national governments. Each member state must now produce a national cloud and AI strategy, establish at least one Centre for AI, and designate a data centre acceleration zone. Bottlenecks in data access are to be removed. The Commission and capitals together must ensure that designated frontier projects receive the computational resources they require.

Positive obligations replace precautionary logic

Gabriela Zanfir-Fortuna, vice president for global privacy at the Future of Privacy Forum, characterises the act as arguably the EU's first comprehensive tech regulation to impose positive obligations to enable innovation and the widespread adoption of AI. The phrase marks a departure. For two decades, the Union's digital acquis, from the eCommerce Directive through the GDPR to the Digital Services Act and the AI Act, has been built on a precautionary logic: identify risks, assign duties to private actors, empower regulators to intervene. CADA inverts that architecture. The state becomes the enabler; the market becomes the beneficiary.

The political framing has centred on sovereignty. Debate in the Council and Parliament has focused on hyperscaler dependence, assurance levels for cloud providers, and whether Europe can finally construct a credible domestic compute stack. Those are legitimate questions. But they obscure a deeper transformation. Widespread AI adoption is no longer treated as a development that governments must prepare for. It has become an objective of supranational policy in its own right.

The reverse centaur and the distribution of agency

Cory Doctorow's recent book, The Reverse Centaur's Guide to Life After AI, supplies a vocabulary for what that objective entails. Doctorow distinguishes two configurations. A centaur is a human directing a machine, the chess player who uses an engine to amplify judgement. A reverse centaur is a human whose dexterity and insight are harnessed by a system they do not control. The distinction is not technological. It is organisational. It asks who sets the targets, who bears the risk, and who captures the surplus.

Doctorow illustrates the dynamic with a concrete example from journalism. A media group does not necessarily deploy AI to make ten journalists twenty percent more effective. It discovers that three journalists augmented by language models can approximate the output of the original ten. The three who remain are not liberated from routine tasks. The output quotas of the seven who left are reassigned to them, alongside the new duty of verifying the model's work and absorbing liability when it errs. The machine does not serve the worker. The worker serves the machine.

Productivity claims and the evidence gap

The Commission's competitiveness agenda proceeds from an assumption that the technology industry has repeated for years: mass adoption is inevitable, and the only question is whether Europe leads or follows. That assumption rests on projected productivity gains that have not yet materialised at macroeconomic scale. Aggregate statistics from the major economies show no clear break in total factor productivity since the current wave of generative models began circulating. Microeconomic studies report task-level speedups, but the translation into economy-wide growth remains contested among economists at the OECD and the IMF.

Doctorow describes the inevitability narrative as a form of vulgar Thatcherism: there is no alternative, so the infrastructure must be built. The European variant adds a sovereignty gloss, there is no alternative, so we had better build the data centre acceleration zones. But the logical structure is identical. Policy treats a contested commercial trajectory as a settled fact, then reorganises public resources to accommodate it.

The macroeconomic blind spot

Even if the optimistic scenario materialises, models improve, European firms become more productive, a domestic AI industry emerges and dependence on American hyperscalers diminishes, a structural problem remains. The capital flowing into AI is driven by a straightforward proposition: the technology promises to displace or replace a substantial fraction of the labour force. For an individual firm, labour is a cost to be minimised. For an economy, labour is also the primary source of household income, consumption demand and tax revenue.

That tension does not appear in the impact assessments accompanying CADA. The regulation's recitals emphasise strategic autonomy, industrial leadership and the reduction of external dependencies. They do not model the fiscal consequences of a sustained decline in the labour share of value added. They do not address how social protection systems financed on payroll contributions would function if employment contracts are replaced by API calls. The AI Act, which entered into force in 2024, asks whether systems are safe, transparent and accountable. The competitiveness agenda asks how to accelerate uptake. Neither asks who the gains accrue to, nor who bears the adjustment costs.

Sovereignty as a scaffold for adoption

The sovereignty argument is not a rhetorical flourish. Europe's compute deficit is real. The Union accounts for roughly ten percent of global data centre capacity while generating eighteen percent of GDP. Dependence on US providers creates vulnerability to extraterritorial law, supply chain disruption and strategic coercion. Building domestic capacity is a legitimate industrial policy goal. But CADA bundles that goal with a mandate for accelerated adoption that is logically distinct. A sovereign cloud infrastructure could exist without a legal requirement that every ministry, hospital and university integrate frontier models into their workflows. The regulation conflates the two.

Member states now face a compressed timetable. National strategies are due within twelve months of the act's entry into force. Centres for AI must be operational within eighteen months. Acceleration zones must be designated and granted streamlined permitting. The Commission will monitor compliance through a new governance board reporting to the Digital Decade policy programme. Failure to meet milestones triggers the standard infringement procedure under Article 258 TFEU.

Implementation timeline and the first test

The first concrete test arrives in early 2027, when national strategies are submitted. Several capitals have signalled resistance. France and Germany have advanced their own cloud initiatives, NumSpot and the Open Telekom Cloud, and may seek to align CADA requirements with existing projects rather than create new structures. Italy and Spain have raised concerns about energy availability for acceleration zones, given current grid constraints. Poland and the Baltic states, by contrast, have welcomed the compute mandate as a route to digital convergence.

The Parliament's report is scheduled for adoption before the 2029 European elections. Its conclusions will shape whether the next Commission treats the reverse centaur problem as a regulatory gap or as an externality the market must resolve. For now, the acceleration zones are being designated, the centres are being staffed, and the infrastructure is being procured. The question of who the machine serves has been deferred.