The EU Tech Sovereignty Package and Europe’s short window to act
In brief
Europe’s AI sovereignty depends on leverage, not self-sufficiency. The Tech Sovereignty Package is an opportunity to build that leverage, but its ambition must match its objectives. SaferAI proposes nine changes from securing access to frontier AI, agreeing proportionate targets to build-up European compute capacity, mobilizing private investment, or strengthening strategic evaluation capabilities in the EU.
The decisions Europe takes over the next eighteen months will determine whether it shapes how transformative AI develops or merely absorbs the consequences. The AI Office’s Frontier AI Expert Forum concluded that unprecedented near-term intervention is needed if Europe is to retain influence over how this technology unfolds. President von der Leyen’s 2026 State of the Union address called for a massive boost in compute capacity as a condition of Europe’s independence. The Transformative AI Strategy warns that without urgent action beginning this year, the prosperity, sovereignty, and security of all Europeans are at risk. The Strategy, to which SaferAI’s Policy Lead Chloé Touzet contributed to shape, is backed by more than 50 signatories including Nobel laureates, as Daron Acemoglu and Philippe Aghion.
On 3 June 2026, the Commission presented the European Technological Sovereignty Package: two legislative proposals, the Chips Act 2.0 and the Cloud and AI Development Act, alongside the Open Source Strategy and a Strategic Roadmap for Digitalisation and AI in Energy, all framed by the Communication on European Tech Sovereignty, COM(2026) 503.
The Tech Sovereignty package is potentially one of the most critical pieces to shape the future of the EU in the AI international landscape.
Where Europe stands
Under 5% of global AI compute sits within the Union, and Europe is estimated to host roughly 2 GW by the end of 2026, against some 45 GW available globally. In 2025, 59 notable AI models originated in the US and 35 in China, while the EU produced one. On frontier capability, Mistral Large 4 (“Le Chonk”), Europe’s leading model, performs roughly on par with GPT-5.5 and Claude Opus 4.7, released about six months earlier (see also the Artificial Analysis’ Intelligence Index). Late-stage AI investment in Europe ($12 billion) runs at less than a tenth of US levels ($141 billion), and 73% of lead investors in European AI funding rounds are American.
Absent a change of course, the gap will widen as frontier development compounds. The current scale and pace of the EU response does not match the required level of ambition to address it. The Package is an opportunity to correct that trajectory.
Sovereignty is not autarky
In 1576, Jean Bodin defined sovereignty as absolute and perpetual power: supreme authority, indivisible, answerable to no one. Applied to AI, this reads as autarky, build everything, depend on no one. Writing about the economic interdependence of the late twentieth century, Robert Keohane proposed a more useful definition: few states can be entirely free of external authority, and the realistic aim is to use sovereign authority as a bargaining tool to shape the rules everyone operates under.
European sovereignty in AI is, in this view, leverage: building and protecting the assets, relationships and infrastructure that let Europe sustain critical functions, bargain for continued access to frontier capabilities, and influence how the technology develops globally.
We need an ambitious target
Europe is not starting from scratch. The AI Continent Action Plan established AI Factories. The Apply AI Strategycommits €1 billion across ten sectors. InvestAI is to mobilize €20 billion for AI Gigafactories. France pledged €109 billion at the Paris AI Action Summit. The scale elsewhere, however, is different. US companies are reported to be investing roughly $650 billion in AI in 2026 alone. Meanwhile, the Tech Sovereignty Package reportedly calls for €200 billion in compute investment by 2036.
Existing estimates suggest that the EU should target 15% of global compute by 2030, a figure matching the Union’s share of global GDP, putting the real investment requirement at around €1.7 trillion. That is roughly 45 GW of AI data center capacity against 300 GW of projected global capacity. This build-out has to be financed primarily by private capital, with public money reserved for national-security use cases.
Our position is straightforward. We welcome the Package as a coherent and comprehensive approach, and both legislative texts get significant things right. The current drafts of CADA and the Chips Act 2.0, however, do not carry targets ambitious enough to build the leverage the Package’s own objectives describe.
Nine proposals to close the gap
We propose nine strategic changes to CADA and the Chips Act 2.0. They cover four areas: making access to frontier AIan explicit objective, securing that access under adverse scenarios such as export controls, mobilising private investment, and protecting the hardware and software layers where Europe holds or could build a competitive advantage.
SaferAI’s proposals to amend the Tech Sovereignty Package:
- Establish access to frontier AI capability as an explicit objective of CADA.
- Set a target of 15% of global AI compute (~45 GW) hosted in the European Union by 2030, roughly matching the Union’s share of global GDP.
- Reserve sovereign European-owned and hosted compute for national-security and resilience use cases.
- Mobilise existing compute through a European AI cloud public registry.
- Shorten time to power in the Data Center Acceleration Zones (DCAZ).
- De-risk private investment in European compute through a European equity capacity and demand commitments.
- Use the Chips Act to protect Europe’s lithography chokepoint and build a downstream competitive advantage in inference chips and the interface and orchestration layer.
- Create requirements for cloud provider diversification in the Cloud computing sovereignty framework.
- Build an autonomous European capacity to evaluate and adapt foreign open-weight models.
Below, we examine four areas of our proposal package in more detail. The full package, including specific amendments covering all nine recommendations, is available on request.
Access. CADA doesn’t name secure access to frontier AI capability as a goal in itself, even though access is the precondition for almost everything else the Act wants: economic growth, scientific research, or the ability to respond to AI-enabled cyber attacks at machine speed. We propose making it an explicit objective, and Europe has a lever to secure it: the hosting capacity that foreign developers already want here can be exchanged for contractual guarantees on parity of model access and portability.
Visibility. Europe’s existing compute sits fragmented across operators, Member States and sectors, with no way to see what is available or to redirect demand between operators. A public registry of that compute capacity would make supply visible, map where apparently independent capacity depends on the same operator, and let public bodies release capacity to one another during a declared incident.
Funding. New capacity has to be financed overwhelmingly by private capital, and that capital is not currently flowing, for two reasons: investors do not know whether compute built in Europe will be bought, and European equity markets are thin. The Commission’s own Communication puts the EU at 5% of global venture capital funds raised, against 52% in the US. CADA should address both, by creating a European private equity capacity and a demand accelerator that aggregates public and private buyers into a credible order book.
The capacity to check. We propose that CADA create an independent EU-based technical capacity to assess cloud sovereignty levels and to evaluate open-weight models. CADA itself creates the AI Assurance Framework; without European evaluation capacity, meeting its requirements will depend on foreign evaluators. Developing third party independent evaluation capacity would also let Europe evaluate and adapt open-weight models, for example, using confidential-computing infrastructure to run deep audits without providers disclosing their weights. If export controls ever limited access to frontier capabilities, the evaluation capacity is what would let the EU fall back on open-weight models in critical infrastructure and other high-risk sectors with confidence in how they behave.
Next steps
The Tech Sovereignty Package is the right instrument, with a coherent view of the stack. What it lacks are targets proportionate to its own objectives. The legislative process gives Europe roughly eighteen months to fix that. Amended, the Package can build the leverage this piece describes; unamended, it unlikely to change the status quo.
SaferAI will engage with decision makers throughout the legislative process to put these nine sets of amendments on the table.
Note: The full package of proposals with particular amendments are available under request. For more information, contact us at bruno@safer-ai.org and chloe@safer-ai.org.
