The Tech Sovereignty Package: Europe’s Chance to Build Strength

On 3 June the European Commission adopted its Tech Sovereignty Package: the Cloud and AI Development Act (CADA), a second Chips Act, an open source strategy, and an energy-sector roadmap. Read together, it reflects a sharper sense of where Europe's strength lies. Rather than chasing a fixed share of global chip production or a spending race with the hyperscalers, it works with the leverage Europe actually holds: indispensability in parts of the chip supply chain, the room to unlock private investment in compute, and the ability to shape demand through procurement and possibly private sector contracts. That is the right instinct, and the Commission deserves credit for it.

Three things still stand between that instinct and Europe actually putting its strengths to use, one in each of the main instruments. In CADA, the sovereignty requirements should not end up foreclosing access to frontier-capability models for the high-stakes uses, where trailing the frontier would directly cost Europe its competitiveness. In the Chips Act, the piece still missing is an EU-level coordination mechanism that pools leverage to deter the deliberate restriction of chip access. And related to the open source strategy, public investment should be redirected from data centre build-outs to research that makes open-weight models safer and more robust.

Cloud and AI Development Act

CADA carries two crucial ideas: Data Center Acceleration Zones and the Cloud Service Sovereignty Framework. After covering these commendable features, we will discuss the Cloud and AI Leadership Initiatives.

Before discussing specific measures the overall ambition level is worth reflecting on. The Communication estimates that expanding European data centre capacity will need around €200 billion by 2036, most of it private. In stark contrast, the five largest US hyperscalers spend more than that in roughly four months, with about three-quarters going to AI infrastructure. A substantially larger investment would be necessary to harness AI for EU competitiveness.

Data Centre Acceleration Zones

For the acceleration zones: each Member State must designate pre-selected sites where the planning and environmental permitting (one of the key bottlenecks to speedy construction), is handled once and up front for the whole zone rather than done project by project. A developer inside a zone inherits a pre-cleared baseline permit, a single contact point and faster handling, with a preference for brownfield sites. This resembles the idea of Special Compute Zones model, and we think tackling the key permitting bottlenecks will make a real difference to how fast compute (computational capacity) gets built in Europe.

However, the current proposal misses a key piece of the puzzle: it excludes grid connection from the acceleration zones. The fast track covers the building permit but stops short of connecting to the electricity grid, which remains outside the regime and carries no deadline. That connection often leads to delays of several years and up to a decade in the busiest hubs, so a developer could clear permitting in a year and still wait years for power. Bringing grid connection inside the acceleration regime would prevent EU data centre construction from facing similar delays. Beyond grid connections, another change worth pushing for is a shorter permitting deadline with a default approval if the process does not conclude within the preset timeframe (Art. 13(5) sets a 12-month cap but no consequence for breaching it) as that reduces risks for developers even further and might attract more private capital.

Cloud Service Sovereignty Framework

The cloud service sovereignty framework scores cloud services on how insulated they are from non-EU control and law, across four assurance levels. Public bodies match each activity to a level, and for sensitive uses may buy only the higher ones; critical-sector private operators can run the same assessment, and the Commission may require it of them. This gives buyers a shared language for how much sovereignty a task needs and protects a procurement lane for fully sovereign providers, both worth having.

However, the proposed framework could greatly benefit from including a key variable: model capability. The higher sovereignty levels are easier to reach with self-hosted open-weight models, whose capabilities trail the frontier by several months. For many use cases, like summarising documents, drafting routine correspondence or translating administrative text, that gap does not matter. However, for competitive and high-stakes use-cases like R&D and defence frontier capabilities will be necessary to ensure European competitiveness. Too stringent sovereignty requirements risk foreclosing those capabilities altogether, locking entire public sectors out of the best tools for the uses where they matter. Therefore, we recommend:

• Making the capability one of the factors weighed in the Art. 29 risk assessment, alongside the data-sensitivity criteria it already covers
• Exempting mission-critical cases (e.g., in defence) from the mandatory higher tiers (Art. 30)
• and securing reliable access to frontier-level models, so the trade-off is made deliberately.

Another recommendation would be to speed up the assessment process of providers in the higher tiers. Currently, each provider’s completed audit is evaluated by its home Member State and then subject to review by the other 26. An opt-in central recognition layer, housed in the Commission, that evaluates a provider's audit once for Union-wide effect would speed up the process significantly. Member States could rely on this by default, or supplement with stricter checks where they have specific cause, leaving national control intact.

Cloud and AI Leadership Initiatives

CADA's public-investment arm, the Leadership Initiatives, would likely channel redundant additional public investments along with private sources into data-centre build-outs. Commercially viable infrastructure development is already being financed by private balance sheets and credit markets at a scale public money cannot match. The five largest hyperscalers alone plan around $660–690 billion of capital spending in 2026, roughly three-quarters of that AI infrastructure, and Morgan Stanley puts global data-centre capex at some $2.9 trillion through 2028, with credit markets expected to fill most of the funding gap. Against that, EU co-financing is marginal. That money would be better spent on fixing a market failure: making the open source ecosystem more resilient, which we discuss later.

Open source strategy

The open source strategy is mostly about software and infrastructure, and on that ground it is the part of the package we would credit most warmly. It reforms public procurement so open source can compete, where tenders today are built around proprietary incumbents and shut out SMEs, as pointed out in the package’s communication. It also shifts EU funding away from scattered one-off grants toward the ongoing maintenance of the critical components and shared dependencies that much of Europe's digital infrastructure quietly runs on. This is the clearest case in the package of public money doing what only public money will. These components are used everywhere, often for free, which means almost everyone benefits from their upkeep but no single company has much reason to pay for it. The work of keeping them secure and maintained therefore tends to fall to a handful of volunteers and stays chronically underfunded, even though a failure in one of them can cascade across the systems that depend on it. The xz-utils backdoor showed what is at stake: a widely used compression library, maintained by a single overstretched developer, was nearly turned into a route into millions of systems before the attempt was caught, and largely by luck. Funding this maintenance is exactly the kind of public good the market will not supply on its own.

As a comparison, funding focused on training new open weight models is a worse use of budget, since such models are still expensive, overtaken within months, and capable open models are already coming out of private and academic labs. The strategy should instead fund the tools and infrastructure that sit around models and make them dependable in practice: the systems for testing and evaluating them, the safeguards that keep them from being misused, and the shared infrastructure that lets others build on them safely. Unlike a model, which is overtaken within months, this work keeps its value for years.

Once weights are open, their safety properties cannot be patched, revoked or monitored: safety fine-tuning can be removed cheaply, white-box access enables far stronger jailbreaks, and models cannot be recalled. The open source strategy asks for "appropriate safeguards" but neither defines nor funds them. Given that many public institutions might start using open weight models due to the CADA sovereignty criteria, a constructive step would be to fund the resilience that would make open models safer to use, through tamper-resistance, machine unlearning and robust evaluation. Projects in this area are arguably more promising funding recipients than publicly-funded compute build-outs.

Chips Act 2.0

The new Chips Act is a much improved proposal than its predecessor. It drops the old goal of raising Europe's share of global chip production to 20% by 2030, a target the EU's auditors expected to miss by almost half, and reorients around reducing strategic dependencies and making Europe indispensable in the segments it already leads: lithography equipment, precision optics, high-power lasers and pre-competitive research infrastructure. It also builds capacity to understand and track the supply chain, through a dependency map and a "digital twin" that can flag risks early.

But it guards against the old threat, not the new one. Its crisis tools, the priority-rated orders and common purchasing it can activate in an emergency, are built for a pandemic-style shortage: a fire at a fab, a drought, a ship stuck in a canal. They do little against the sharper risk today, that chip access is cut off deliberately, as an instrument of geopolitical leverage. This threat is no longer hypothetical. China has curbed exports of gallium and other inputs to apply pressure; the United States' AI Diffusion Framework sorted even allied nations into tiers with different access to advanced chips, and the proposed GAIN Act would put US buyers first in the queue for domestically made chips. A strategy designed for accidents cannot deter a rival who turns off access on purpose. The Commission is meeting the challenge halfway, but pushing harder on the following would help:

• mandatory routine reporting on exposure, concentration and substitutability: the Commission can compel disclosure only reactively, once a risk is flagged (Art. 38) or a crisis declared (Art. 41), whereas the leverage map it needs must be built in calm times, well before any shock, when reporting is purely voluntary (Art. 33(5), Art. 35(2))
• dependency mapping in both directions, so the Commission sees where others rely on European chokepoints and not just where Europe is vulnerable (Art. 33(1) maps the Union's "strengths and weaknesses", but its concrete factors look only at Europe's own dependencies and bottlenecks, 33(1)(b) and (d); there is no mandate to map where others depend on European chokepoints)
• a strategic-asset designation that extends to research, not only production: the new recognition for strategic projects and first-of-a-kind initiatives is a real step, but it reaches only actual manufacturing projects (Art. 14(3); first-of-a-kind is defined in production terms, Art. 2(10)), leaving research infrastructure, one of Europe's strengths, without the same protective status
• and the analytical capacity to act on any of this: the proposal staffs the effort with around a dozen people at the Commission, against the 140-plus the US put behind its chip programme, with Member State teams thinner still

Europe controls real chokepoints but has no way to wield them collectively. Export-control decisions sit with 27 Member States individually, which lets a major power isolate and pressure one government at a time: a single Member State has absorbed most of the cost of US controls on China, its leading equipment supplier's China sales restricted while some US firms kept selling. Industry wants this fixed: in one experts workshop, European chip firms strongly favoured a unified EU regime over today's fragmentation, and expected it to roughly double Europe's negotiating leverage. The EU's one collective tool, the Anti-Coercion Instrument, is too slow for a crisis that moves in weeks and may not even apply when restrictions are framed as national security. Among the pieces still missing:

• decide export restrictions jointly, for example by giving the European Semiconductor Board a say before any Member State curbs exports of strategic assets, so that a concession one country is pressured into making becomes collective leverage the Commission can trade for reciprocal guarantees, such as security of supply.
• options short of the unusable last resort of cutting off advanced chip manufacturing equipment: narrowing access to pre-competitive research, tightening IP licensing, controlling consumables such as photoresists and etching gases, or withholding maintenance on installed equipment
• a pre-funded pot to cover firms' losses when a collective measure is used, so no single company or country carries the bill for Europe's defence alone.

Taken together, the new Chips Act is a clear improvement: it swaps an impossible-to-achieve production target for a focus on Europe's strengths where the Union can become indispensable, and builds the dependency mapping to see the supply chain vulnerabilities clearly. It is key to ensure the EU can act on the vulnerabilities it finds; mapping dependencies in both directions, extending strategic-asset status to research, and the staff to do the analysis, as well as a way to stand up for itself, wielding Europe's chokepoints collectively through joint export decisions, a wider range of response options, and a way to share the cost in case of disruptions, so the map becomes real bargaining power.

Energy roadmap

The fourth component, a roadmap for digitalisation and AI in energy, is a soft instrument working through a voluntary commitments, with a data-centre sustainability label as its one binding piece. Its flagship, AI.grids, a pan-European model for optimising grid efficiency and reliability, is a sound idea for public spending and, as shared infrastructure, is much more likely to qualify as a public good than a subsidised private asset. What it cannot do is close the grid-connection bottleneck flagged under CADA, since it makes the grid smarter without speeding up access to it. Bringing grid connection inside CADA's acceleration regime is what would close that loop.

What we would prioritise

Most of what we would change is sharpening rather than redirecting: bring grid connection inside CADA's acceleration regime, so permitting speed is not undone by connection delays; weigh the capability gap openly wherever procurement steers toward sovereign providers, with carve-outs for R&D and defence; move public money from computing infrastructure subsidies and short-lived training runs toward the public goods the market will not supply, such as the maintenance of critical open source dependencies and the hardening of open-weight models; and give the Chips Act a coordination layer for deliberate restrictions on access, not only accidental shortages.

These are refinements to a package whose foundations are sound. The Commission’s diagnosis is increasingly sharp and several of its proposed instruments point to Europe’s strengths. Arq will keep working on these questions, building on our research into compute infrastructure, open-weight model safety, and Europe's wider AI preparedness.