Thinking from infinite possibility
November 11, 2026
AI Startups & Opportunities
The Sovereign Cloud Playbook: The Optimal EU Policy to Build Champions
November 10, 2026
ENSI Research Division · Built on a library of 150 primary documents across 15 angles (EU legislation and impact assessments, competition-authority decisions, government legal opinions, central-bank and EIB evidence, national compute strategies from nine countries, think-tank opinion on both sides of the sovereignty debate, and the official record of the Schwarz Group’s STACKIT build-out).
The argument, before the list
Every serious European document on cloud now begins with the same number and ends with the same shrug. The number: European providers’ share of their own cloud market fell from 29% in 2017 to about 15% in 2022 and “has remained stagnant since then”, while three American hyperscalers control more than 70% (CADA proposal, p.3). The shrug belongs to Mario Draghi, who wrote in 2024 that “it is too late for the EU to try and develop systematic challengers to the major US cloud providers: the investment needs involved are too large” — and then, in the same sentence-chain, that the Union should nonetheless “ensure that it has a competitive domestic industry that can meet the demand for ‘sovereign cloud’ solutions” (Draghi report, p.36).
That pair of sentences is the whole policy problem, and most of the European debate has misread it. Draghi is right on the first half and under-ambitious on the second. It is too late to build a European AWS that competes for every workload on Earth — big-five hyperscaler capex will exceed $600bn in 2026 alone (MUFG, p.1), more than the entire stock of US venture funds is worth in euros. But “meeting the demand for sovereign cloud” is not a niche. It is the tier of public administration, defence-adjacent industry, health, critical infrastructure, regulated finance and — increasingly — the AI agents that will act inside all of them. That tier is large enough to sustain two or three operators at genuine hyperscale, and it is the only tier where Europe can legitimately shape demand rather than simply bid for it.
The reframe this report argues for is simple: the goal is not autarky and not a champion by decree. The goal is a contestable European cloud and AI market in which two to three EU-controlled operators reach hyperscale inside a clearly defined sovereign tier, and every operator — American included — competes on open terms everywhere else. Sovereignty is built where Europe has leverage: jurisdiction, operations, keys, software, interoperability and the order book. It is not built at the chip layer, where 96% of AI chips sold in the EU in 2024 came from US vendors (Chips Act 2.0 assessment, p.22), and where no plausible policy changes that before the mid-2030s.
The test case running through this playbook is the Schwarz Group’s STACKIT — the most interesting European experiment in years precisely because it did not start as a policy project. A retailer with €185.6bn of revenue built a cloud to run its own 600,000 workplaces, Lidl and Kaufland, then opened it to the market, won a place among four providers of the EU institutions’ own sovereign cloud at the Commission’s “SEAL-3” level, became the infrastructure layer of the German federal AI platform, and announced an €11bn, 200 MW data centre at Lübbenau with room for up to 100,000 GPUs (see Report 1 for the company in detail). STACKIT is what Europe gets without a coherent policy: real infrastructure, real certifications, a captive anchor customer — and, by credible press accounts, external cloud revenue still measured in tens of millions of euros rather than billions. Every instrument below is tested against one question: would it turn a STACKIT, an OVHcloud, a Scaleway or an IONOS into a sovereign-tier operator that can stand on its own order book within a decade?
The honest answer today is no — not because Europe lacks money, but because it lacks the four things that actually build infrastructure champions everywhere they have been built: a guaranteed order book, a single market for certification, priority access to power, and a capital stack matched to an asset that looks more like a power plant than a start-up. Those are policy choices. None of them requires out-spending Silicon Valley.
Summary of the main points
Define the goal correctly. Not a European AWS, not autarky: a contestable market plus two to three EU-controlled operators at hyperscale in a legally defined sovereign tier, with control of jurisdiction, operations, keys and software — not chips.
Demand is the scarcest input. Every successful case (Airbus, Japan’s SAKURA internet, France’s SecNumCloud tier, the US JWCC) paired supply support with a guaranteed buyer. Every failure (Cloudwatt, Numergy, much of Gaia-X) subsidised supply into a market that never bought.
Ten gaps explain Europe’s stagnation: fragmented, price-only procurement; no certify-once label; the EUCS deadlock; grid queues of 7–13 years; a capital stack built for software, not infrastructure; gigafactory funding that is thin and scattered; a cloud IPCEI one-seventh the size of the chips IPCEI that missed the scaling players; unremedied licensing and egress lock-in; no sovereign offtake; and no measured targets.
Eleven instruments, ranked. First a sovereign-tier offtake guarantee; then a certify-once EU sovereign label with a FedRAMP-style presumption of adequacy; a security-of-supply award criterion; grid priority in sovereign compute zones; a matched capital stack; gigafactory reform; conditionalities; competition enforcement; a domestic-compute tax credit; a capped economic-security capex subsidy; and a mission office with kill-or-scale discipline.
The public money required is moderate and mostly recoverable. ENSI’s labelled estimate is roughly €70–90bn of public commitment over 2027–2034, of which well over half is purchases, guarantees and equity rather than grants.
The institutional answer is a coalition, not a 27-way consensus. The Commission sets the label and the market rules; the EIB carries the capital stack; a France–Germany–Netherlands–Nordic core commits the first order book; mid-sized states such as the Czech Republic join early as anchor buyers and site hosts.
What Europe should actually try to achieve
Policy fails most often not in the instrument but in the target. Europe has spent a decade oscillating between two wrong goals: autarky (”a fully European stack from silicon to software”) and neutrality (”any cloud is fine if the data sits in Frankfurt”). The first is unachievable; the second has been tested by courts, legal opinions and one very public sanctions episode, and found wanting.
The right target has five parts.
A defined sovereign tier. Not all workloads need sovereignty. The French experience shows the cost of over-reach: the Cour des comptes puts SecNumCloud-grade hosting at 25–40% above standard cloud, with only about ten qualified providers whose services “do not have the depth of the hyperscalers” (French Court of Audit, p.10–11). The tier must be narrow enough to be affordable and broad enough to sustain operators at scale: highly critical public workloads (which CADA already wants on sovereign cloud by 2035 — CADA proposal, p.78–79), health and social-security data, defence-adjacent industry, critical-entity operations under NIS2, regulated finance under DORA’s concentration-risk lens, and public-sector AI agents with delegated authority.
Two to three EU-controlled operators at hyperscale inside that tier. One is fragile, five is Gaia-X. The Airbus lesson is consolidation behind a single integrated product (Airbus study, p.2, p.49); the Cloudwatt–Numergy lesson is that splitting an effort between two state-sponsored rivals and excluding the provider the market had already chosen destroys both (French Senate 2019, p.136, p.222). Two to three gives competition without dilution.
Control of the layers Europe can actually control. Jurisdiction and ownership of the operator; operations by EU staff; customer-held keys; open, portable software (Kubernetes, open-weight models, open APIs); and the order book. The chips will be American, and the EU’s own Cloud Sovereignty Framework already reflects that by grading supply-chain sovereignty in levels rather than demanding purity (Cloud Sovereignty Framework, p.3, p.6).
A contestable market everywhere else. For the 80%-plus of workloads outside the sovereign tier, the aim is switching, not exclusion: fewer than 1% of UK customers switch provider a year and Microsoft’s licensing partially forecloses rivals, the UK’s competition authority found (UK CMA decision, p.15–17). Contestability is how a sovereign-tier operator grows out of its tier.
Capacity at the scale demand requires. The CADA impact assessment projects EU data-centre demand of 37 GW by 2030 against 28 GW of central-scenario supply — a gap of about 9 GW, widening to 19 GW by 2036 (CADA impact assessment, p.44–49). Sovereign operators should own a meaningful slice of the new build, not just the leftovers.
The sovereignty split — and why both camps are half right
The European debate is polarised between two schools, and anyone designing policy must understand who is paying for which argument.
The ownership-and-immunity school — France’s SecNumCloud, the CISPE association of European providers, the EuroStack movement, the French Senate, Eurosmart — defines sovereignty by who controls the operator. SecNumCloud’s section 19.6 is the hardest version: EU headquarters, non-EU shareholders capped at 24% individually and 39% collectively, no veto and no board majority (SecNumCloud rules, p.50–51). Its strongest evidence is legal. A legal opinion written for Germany’s Federal Interior Ministry concludes that the CLOUD Act and FISA Section 702 reach EU-located data wherever a US parent has “control”, read broadly (German legal opinion, p.3–5). A memo commissioned by the Dutch government on the AWS European Sovereign Cloud found that US access to metadata and sanctions-driven suspension were “improbable” but “not inconceivable”, and the Dutch minister drew the conclusion that full sovereignty is therefore not achieved (Dutch legal review of AWS, p.1–2; Dutch minister’s letter, p.1–3). And under oath before the French Senate, Microsoft France’s representative said he could not guarantee that French citizens’ data would never reach foreign authorities without French consent (French Senate inquiry, p.23).
The risk-based school — ECIPE, the Atlantic Council, Bruegel, ICLE, Clifford Chance, the Dutch and Nordic governments — defines sovereignty by demonstrated controls: encryption, customer-held keys, EU data residency, contractual exit, concentration-risk management. Its strongest evidence is economic. ECIPE finds that of 8,725 public cloud contracts in nine member states about 97% already go to domestic providers, 100% where sovereignty language is used, while single-bidder tenders rise from 28% to 46% under sovereignty framing (ECIPE on tenders, p.2–3); it also prices the cost of CADA’s highest sovereignty levels at a central €115bn over 2027–2036, about 70% of it re-architecting (ECIPE on CADA, p.2–3). Bruegel argues that capital is not the binding constraint — the EU has only 2 GW of AI compute, about 5% of the world total, because of permitting and grid access — and that strict local-content criteria risk a negative supply response (Bruegel on AI compute, p.1).
The funding must be named. ECIPE’s 2026 single-market paper was supported by Microsoft; its 2025 “Breaking barriers” study was commissioned by the Open Cloud Coalition; Implement Consulting’s “Clearing the cloud” was written with Google. On the other side, CISPE, Eurosmart, Bitkom and the EuroStack letter are advocates for European providers. Neither camp is disinterested; both supply real evidence.
The synthesis is not a compromise; it is a division of labour. Ownership-and-immunity rules are right for a narrow sovereign tier, where the risk is not data theft but switch-off — the Greenberg Traurig memo shows that even the best US-run sovereign design cannot remove sanctions exposure, and that is a risk no encryption scheme mitigates. Risk-based rules are right for everything else, where exclusion would raise costs, shrink bidder pools and, as ECIPE’s single-bidder data warns, hand rents to whichever domestic incumbent wins. The Commission’s own Cloud Sovereignty Framework is already a hybrid — ownership-type objectives carry about 35% of the score, legal immunity only 10% (Cloud Sovereignty Framework, p.6). The playbook below hardens the hybrid into two explicit tiers, and then builds the demand, capital and power to fill the top one.
The diagnosis: ten things Europe is missing
Europe’s cloud stagnation is usually explained by one cause — scale, or capital, or regulation. The evidence points to ten distinct gaps, each with its own fix. They are ordered by how directly they block a sovereign-tier operator from reaching scale.
Gap 1 — There is no order book. Public demand is fragmented and bought on price.
EU public procurement is worth about €2.5 trillion a year, roughly 15% of EU GDP (EP briefing on procurement, p.1–2). It is the largest pool of addressable demand in the European economy, and it is spent in a way that structurally favours scale incumbents: in eight member states, more than 80% of awards are made on lowest price alone (ECA Special Report 28/2023, cited in the same briefing, p.5). The buyers do not act as one. The UK spends at least £14bn a year on digital, £9bn through central frameworks, yet has only 15 commercial staff covering 19 strategic suppliers and no pipeline of departmental demand to negotiate with (UK NAO, p.6, p.10). Dutch central government runs 1,588 cloud services, does not know the type of 26% of them, and had no risk assessment for 67% of its material public-cloud services; the Court of Audit found ministries “insufficiently use government-wide expertise and buying power” (Dutch Court of Audit, p.6–7). The EU institutions themselves spent only €87m on public cloud in 2025 (Commission answer to MEPs, p.1). The result: a European provider must win the same customer twenty-seven times, ministry by ministry, on price, against an incumbent whose marginal cost is set by a global fleet.
Gap 2 — There is no “certify once, sell everywhere”.
In the United States, a FedRAMP authorisation carries a statutory “presumption of adequacy”: every federal agency must treat it as sufficient for its own authorisation at or below that impact level, and agencies have reused authorisations hundreds of times across more than 300 offerings (FedRAMP memo, p.2, p.6). One security assessment opens the whole federal market. Europe has the opposite: Germany’s BSI C5, France’s SecNumCloud with more than 700 unit requirements (SecNumCloud overview, p.3–5), Spain’s ENS, the Netherlands’ own policy, and an EU certification scheme (EUCS) that has been deadlocked for years over whether to include immunity requirements at all (cep, p.1). The March 2024 EUCS draft dropped the immunity annex and replaced it with an attestation that merely discloses which jurisdictions a provider is subject to (R Street, p.1). France’s push for an immunity level “has so far remained isolated within the EU” (French Court of Audit, p.10). For a European provider, every national certification is a fixed cost; for a hyperscaler with thousands of compliance staff, it is a rounding error. Fragmented certification is a hidden subsidy to scale.
Gap 3 — Power is allocated by queue, not by strategy.
The binding physical constraint is not land or capital but the grid connection. The IEA records connection waits of up to 10 years in the Netherlands and 7 in Germany, 5–7 in the UK, against a US average of 1–3 years; Dublin is paused until 2030 (IEA, p.94). Ember finds average waits of 7–10 years, and up to 13, in the Frankfurt–London–Amsterdam–Paris–Dublin hubs, and that countries with half the connection time attract about 20% more data-centre growth by 2030 (Ember, p.18). Meanwhile, the queue is clogged with hoarded capacity: European data centres use only 44% of their nominal power rating, and Irish data centres draw about 34% of contracted capacity (same, p.18). Today, CADA admits, permits often take 48 months or more against its target of under 18 (AI Continent plan, p.11). A sovereign operator that wants to build 200 MW joins the same first-come queue as a speculative developer reselling to a US hyperscaler.
Gap 4 — The capital stack is built for software, not for power-plant-like assets.
Europe’s capital debate is framed around venture capital, and the numbers there are genuinely bad: the EU raises about 5% of global VC against the US’s 52% (Draghi report, p.31); US VC funds total about €930bn against about €150bn in the EU, and more than half of EU-located funds’ investments go outside the EU (ECB 2026, p.64–66); Atomico counts a $375bn growth-capital gap (Atomico 2025, p.20). But VC is the wrong instrument for a cloud build. Of the roughly $2.9 trillion of global data-centre spend for 2025–2028, about half comes from hyperscaler cash flow and half externally — $800bn of it private credit, with private equity, sovereign funds, corporate and securitised debt making up the rest (MUFG, p.1–2). A European sovereign operator needs project finance, long-dated infrastructure equity, and investment-grade offtake contracts to lend against. Europe has the savings — about €10 trillion of retail money sits in bank deposits (Savings and Investments Union, p.3) — but European pension funds put about 0.009% of assets into VC against 0.028% in the US (Atomico 2025, p.43), and there is no standard, bankable asset class called “sovereign European compute” for them to buy.
Gap 5 — The flagship AI instrument is thin and scattered.
The AI Gigafactories are the EU’s most visible bet: InvestAI aims to mobilise €200bn, including a €20bn facility for up to five gigafactories of more than 100,000 advanced processors each (AI Continent plan, p.4, p.8–9). But the legal design caps the EU’s contribution at 17% of the computing capex, which member states must at least match; the private consortium carries the rest plus all operating costs (gigafactory regulation, p.8). A third of capex from the public side is not trivial — but the call attracted 76 expressions of interest across 16 member states and 60 sites (Commission news, 30 June 2025; not in library as PDF), and a submission-system failure in November 2026 forced every bidder to resubmit. The German contenders — Deutsche Telekom, Schwarz Digits, IONOS with Hochtief — are now bidding separately after a joint consortium failed (Elektronikpraxis). Compare Airbus: launch aid covered 100% of A300 and about 75% of A320 development costs, inside a single consortium, over two decades (Airbus study, p.25, p.60). Europe is offering a minority share, to many rivals, for a single generation of hardware.
Gap 6 — The cloud IPCEI is small and missed the operators that are actually scaling.
IPCEI-CIS, the Important Project of Common European Interest on next-generation cloud, commits up to €1.2bn of state aid to unlock €1.4bn of private money across 19 projects in seven member states (IPCEI-CIS announcement, p.1). The microelectronics IPCEI is about €8bn public for €13bn private — roughly seven times larger (IPCEI progress report, p.15). And the Schwarz Group, which went on to build Europe’s most ambitious new sovereign data-centre programme, does not appear among the direct participants (IPCEI-CIS decision, p.13–15). The instrument funds open-source middleware research, which is useful, but it is not designed to finance operators or capacity, and its selection did not track who was building.
Gap 7 — Lock-in remains largely unremedied.
The UK’s competition authority found three adverse effects on competition in cloud: concentration with significant unilateral market power for AWS and Microsoft, technical and commercial barriers to switching (notably egress fees), and Microsoft licensing practices that partially foreclose rivals — to the point that a rival’s input price can exceed Microsoft’s own retail price (UK CMA decision, p.16–17). If prices are just 5% above competitive levels, UK customers overpay about £500m a year (p.17). The CISPE complaint estimates the extra cost of running Microsoft software on non-Microsoft clouds at about 28% (CISPE complaint, p.3–4). After the CMA’s decision, AWS and Microsoft cut some egress fees in March 2026, but uptake has been “limited so far” (CMA update 2026, p.3–4). In the EU, Digital Markets Act investigations into AWS and Azure as possible gatekeepers produced preliminary findings in June 2026 (web context). A sovereign-tier operator cannot grow out of its tier if every customer it courts is locked by licence terms elsewhere.
Gap 8 — Nobody guarantees to buy sovereign capacity.
Supply subsidies without demand commitments are how Europe lost its first sovereign-cloud war. France’s state-sponsored Cloudwatt and Numergy, launched in 2009–2010, failed by 2016 for lack of market uptake at a public cost of about €56m — while OVH, excluded from the scheme, became Europe’s cloud leader without subsidies (French Senate 2019, p.136, p.222). The French Commission on AI drew the lesson in 2024 and proposed a “public guarantee of usage” for private compute installed in France (French AI Commission, p.11). France finally delivered it in July 2026: a commitment to buy €100m of compute from the gigafactory selected on French soil, explicitly to de-risk the private investment (French €100m commitment, p.1). It is the only such offtake in Europe — and it is national, small, and tied to one site.
Gap 9 — There are no measured targets, so nothing is managed.
The European Court of Auditors found that the EU’s AI ambitions were “not effectively coordinated” between the Commission and member states, that targets were never updated and not specific about results, and that EU AI investment did not keep pace with global leaders (EU auditors on AI, p.4–5). Its batteries review judged that policy effective overall but lacking quantified, time-bound targets (EU auditors on batteries, p.4–5). The first data-centre reporting cycle under the Energy Efficiency Directive drew returns from only about 36% of facilities in scope, with some member states reporting none (Commission data-centre energy report, p.6–9). One year after Draghi, implementation trackers disagree wildly — Institut Montaigne counts 30% legal implementation (Institut Montaigne, p.7–11), the JEDI tracker 14% of the top twenty (JEDI Draghi tracker, p.1). CADA is the first instrument to carry numeric goals; it still lacks an owner who reports against them.
Gap 10 — The chips and the models are someone else’s, and the software layer is under-built.
Ninety-six per cent of AI chips sold in the EU in 2024 came from US vendors, and all leading-edge foundry capacity sits in Taiwan and South Korea (Chips Act 2.0 assessment, p.22). Chips Act 1.0 triggered about €80bn of announced investment but no leading-edge AI logic (p.11). US access policy can change within months: the January 2025 AI Diffusion rule placed Poland, Portugal, Greece and the Baltic states outside the allied tier and capped compute exports to them, before being rescinded in May 2025 (CSIS on export rules, p.4; RAND on export rules, p.3–4). The model layer is no better: the published model catalogue of Europe’s most-certified sovereign AI serving platform lists only American and Chinese open-weight models (STACKIT AI models). Europe therefore must build sovereignty where it can — operations, jurisdiction, keys, orchestration, the agent runtime — and hedge the chip dependency through diversity of supply and allied agreements rather than pretend to remove it. Draghi’s broader warning applies: close to 30% of European unicorns founded 2008–2021 moved their headquarters abroad (Draghi report, p.8, p.31). The talent that builds these layers follows the order book.
What the ten gaps add up to
Read together, the gaps describe a market in which a European operator faces higher fixed costs (Gap 2), slower and dearer power (Gap 3), costlier and less suitable capital (Gap 4, Gap 5), fragmented and price-only buyers (Gap 1, Gap 8), and locked-in customers (Gap 7) — while the support it gets is small, scattered and unmeasured (Gaps 5, 6, 9). None of this is about European providers being technically incapable: STACKIT holds BSI C5 across its service catalogue, including AI Model Serving, and reached SEAL-3 in the Commission’s own tender. It is about the market structure they are asked to win in. That is why the remedy starts with demand.
How the playbook is ranked
The eleven instruments below are ranked by leverage per public euro on the binding constraint. Demand comes first because every successful case in this research pairs supply support with a guaranteed buyer, and every failure lacks one. Market-wide rules (certification, award criteria, grid allocation) come next because they cost little and change the economics for every operator at once. Capital instruments follow, because capital is plentiful in Europe but mis-shaped. Direct subsidies come last, not because they are useless — Japan’s work — but because without the first six they buy capacity nobody uses. Each instrument is set out as a seven-part brief: In short · Why it ranks here · Evidence and world practice · Design details and numbers · Risks and what failed elsewhere · The agentic engine · First moves and owner.
Instrument 1 — A European Sovereign Tier Offtake Guarantee
In short. A multi-year, multi-vendor purchase commitment for sovereign-tier compute and cloud, aggregated across willing member states and the EU institutions, priced at market and awarded only to operators that pass the sovereign label (Instrument 2).
Why it ranks first. A bankable order book converts every other instrument into financeable capacity. Lenders lend against contracted revenue; a sovereign operator with five-year public offtake can raise project finance at infrastructure rates instead of equity at venture rates.
Evidence and world practice. The US Joint Warfighting Cloud Capability is an up-to-$9bn multi-vendor vehicle, made mandatory for Defense enterprise cloud by a 2023 memo (press reporting; the contract itself is not public). France’s €100m offtake from its gigafactory is explicitly designed “to de-risk the capital investments” (French €100m commitment, p.1). Japan admitted SAKURA internet to its Government Cloud on a conditional, time-bound basis, alongside a capex subsidy (Sakura report, p.2). CADA already allows joint procurement by the Commission on behalf of member states and creates a EuroCloud Federation for public capacity (CADA proposal, p.66). Bruegel shows EU joint procurement works where scale and monopsony matter — vaccines, gas, ammunition (Bruegel on joint procurement, p.1).
Design details and numbers. Five-to-seven-year framework, at least three awardees (no single-vendor JEDI-style risk), volume floors rather than fixed prices, step-in rights and portability clauses. Scale (ENSI estimate, labelled): if the sovereign tier is 10–15% of public cloud and AI spend, and EU public digital spend runs at the order of €80–100bn a year, the addressable sovereign tier is roughly €8–15bn a year; an aggregated guarantee covering 25–35% of that — €3–5bn a year — is enough to underwrite two to three operators’ first hyperscale campuses. This is a purchase, not a subsidy; the public sector receives the service.
Risks and what failed elsewhere. JWCC went entirely to incumbents; the UK’s aggregation produced discounts for existing suppliers rather than new entrants (UK NAO, p.10). The design must therefore restrict the vehicle to label-qualified operators and publish utilisation. Over-specification raises prices — the 25–40% SecNumCloud premium (French Court of Audit, p.10–11) must fall with volume or the tier must shrink.
The agentic engine. A demand-forecasting agent aggregates ministries’ workload pipelines into the forecast the NAO found missing; a utilisation agent reports take-up monthly (Report 3).
First moves and owner. Commission (DG DIGIT and DG CNECT) as joint-procurement agent; a founding coalition of France, Germany, the Netherlands and two Nordic states commits volumes in the first 12 months; the EIB recognises the contracts as eligible collateral.
Instrument 2 — One EU Sovereign Cloud Label with a Presumption of Adequacy
In short. Turn the Commission’s Cloud Sovereignty Framework into a legally recognised two-level label — “Sovereign” (ownership and immunity, for the critical tier) and “Trusted” (risk-based controls, for everything else) — with a statutory presumption that any authority in any member state must accept it.
Why it ranks second. It converts a 27-fold compliance cost into a single one and turns one sale into many. It costs almost nothing in public money.
Evidence and world practice. FedRAMP’s presumption of adequacy is the template: authorisations were reused hundreds of times across more than 300 offerings (FedRAMP memo, p.2, p.6). The Commission’s framework already exists, with eight objectives and SEAL-0 to SEAL-4 assurance levels used as minimum gates per objective (Cloud Sovereignty Framework, p.2–3). SecNumCloud’s section 19.6 supplies the hard ownership test (SecNumCloud rules, p.50–51); Gaia-X requires EU headquarters only at its top label level (Gaia-X rules, p.13).
Design details and numbers. “Sovereign” = SEAL-3 or higher on strategic, legal, data and operational objectives, with the SecNumCloud-style ownership caps (24%/39%) as the legal floor; “Trusted” = SEAL-2 with customer-held keys and audited exit. Mutual recognition with C5 and SecNumCloud on day one, through a mapping annex. The Eurosmart proposal to amend the Cybersecurity Act’s Article 54 so that a “High+” EUCS level can require immunity offers a legislative route (Eurosmart, p.1–2); CADA’s harmonised sovereign-cloud definition offers a faster one.
Risks and what failed elsewhere. The EUCS deadlock shows the political risk of trying to make one level fit all; two levels end the deadlock by letting risk-based states use “Trusted” for most workloads. US trade retaliation is a real risk — ECIPE warns of up to $12bn of retaliatory tariffs (ECIPE 2023, p.2; industry-aligned source) — which argues for a narrow sovereign tier, not none. GAO found FedRAMP authorisations hard to cost (web context): keep the label’s audit burden proportionate for smaller providers.
The agentic engine. Continuous-compliance agents monitor control evidence and flag drift between audits, cutting recertification cost (Report 3).
First moves and owner. Commission with ENISA, ANSSI and BSI; adopt via CADA’s implementing acts; the joint ANSSI–BSI statement on sovereignty criteria of November 2025 (web context) is the political anchor.
Instrument 3 — A Security-of-Supply Award Criterion That Replaces Price-Only
In short. In the revision of the public procurement directives, require a security-of-supply and sovereignty criterion for designated critical-tier ICT purchases, weighted enough to matter.
Why it ranks third. Price-only buying is the mechanism by which public demand flows to scale incumbents. Fixing the award rule shifts demand across all 27 member states at zero fiscal cost.
Evidence and world practice. Over 80% of awards in eight member states are lowest-price (EP briefing on procurement, p.5). Chips Act 2.0 already proposes a security-of-supply criterion “alongside price” for semiconductors in critical infrastructure (Chips Act 2.0, p.4–5) — the template for cloud. The French Senate asks for an EU-level European preference modelled on the Buy American Act (French Senate inquiry, p.21–23). CADA proposes that “European added value” may weigh up to 15 of 120 award points, and that member states aim to award at least 25% of cloud and AI procurement to innovative SMEs (CADA proposal, p.34, p.66).
Design details and numbers. Is 15 of 120 enough? For “Trusted”-tier purchases, yes — 12.5% is a meaningful tie-breaker that keeps price and performance primary. For “Sovereign”-tier purchases, no: there the label should be an eligibility condition, not a score. The recommendation is a two-speed rule: label as a gate in the sovereign tier; a 15–20% weighting for supply security in the trusted tier; and a duty to publish the justification whenever a non-labelled provider wins a critical-tier contract, as CISPE proposes (CISPE manifesto, p.2–4).
Risks and what failed elsewhere. ECIPE’s single-bidder finding (28% rising to 46% under sovereignty framing — ECIPE on tenders, p.2–3, Microsoft-supported) is the warning: a gate without a sufficient supply base creates local monopolies. Pair the gate with Instruments 1 and 5 so that at least three label-qualified operators exist in every lot, and allow cross-border bidding so smaller states are not captured by one national provider — ECIPE’s warning that smaller member states would become dependent on French and German providers (ECIPE on CADA, p.2–3) is best met by a single European market for the label, not by abandoning it.
The agentic engine. Tender-analysis agents check every critical-tier award against the rule and publish the justification register automatically (Report 3).
First moves and owner. Commission (DG GROW) in the procurement-directive revision; Council and Parliament; CADA as the early vehicle.
Instrument 4 — Sovereign Compute Zones with Grid Priority and Use-It-or-Lose-It Queues
In short. Designate a limited number of large sites (500 MW to 1 GW+) in uncongested grids as Sovereign Compute Zones, with pre-cleared permits, guaranteed connection dates, and priority for label-qualified capacity — while reforming connection queues everywhere to end hoarding.
Why it ranks fourth. A seven-to-thirteen-year wait outweighs any subsidy. Power allocation is a state asset; Europe is currently allocating it by queue order.
Evidence and world practice. The UK’s AI Growth Zones are sites of at least 500 MW, with at least one above 1 GW, with planning and power priority (UK compute roadmap, p.12, p.23). The US AI Action Plan uses categorical permitting exclusions and federal land (US AI Action Plan, p.17–18). China steers new capacity to hub nodes where power is cheap and bars projects where utilisation is below 50% (China’s compute plan, p.3–4). Ember finds flexible or phased connection agreements could cut queue time to about one year (Ember, p.5), and ENTSO-E recommends milestones, deposits and use-it-or-lose-it rules against speculative queue-holding, with up to 3.8 GW of realistic data-centre flexibility by 2030 (ENTSO-E, p.17, p.25). CADA proposes Data Centre Acceleration Zones and an 18-month permit ceiling (CADA proposal, p.78–79). Bruegel, from the risk-based camp, agrees that permitting and grids, not capital, are the binding constraint (Bruegel on AI compute, p.1).
Design details and numbers. Target 4–6 Sovereign Compute Zones by 2030 with a combined 3–4 GW of reserved connection capacity — roughly a third to a half of the CADA impact assessment’s 2030 supply gap of about 9 GW (CADA impact assessment, p.44–49). Prefer the regions Ember identifies as growing fastest outside the FLAP-D hubs — the Nordics, Iberia, Italy, France, eastern Germany (Lübbenau is a brownfield power-plant site of exactly this type). Require flexible connection, waste-heat use and EED reporting as conditions.
Risks and what failed elsewhere. Ireland’s grid shows concentration risk: data centres used almost 80% of Dublin’s electricity in 2023 (Ember, p.10) and rise to about 30% of national demand by 2032 (EirGrid, p.61). Zones must sit where new generation is coming, not where demand already strains the network.
The agentic engine. Grid-twin agents simulate connection scenarios and flexibility offers, compressing studies that now take months (Report 3).
First moves and owner. Member states (energy ministries and TSOs) designate zones; ENTSO-E and ACER set queue rules; the Commission links zone status to CADA strategic-project status.
Instrument 5 — A Capital Stack Matched to the Asset
In short. Build a standard, bankable asset class for sovereign compute — EIB guarantees and first-loss tranches under InvestAI, a European Sovereign Compute Infrastructure Fund open to pension and insurance money, and growth equity for the software and AI layers through the European Tech Champions Initiative.
Why it ranks fifth. Europe does not lack savings; it lacks a vehicle that packages contracted sovereign capacity into something long-term investors can hold.
Evidence and world practice. About half of global data-centre spend through 2028 is externally financed, with private credit the largest component (MUFG, p.1–2). The EU has about €10 trillion in retail bank deposits, of which a flow of about €350bn a year could be redirected into market investment (Savings and Investments Union, p.3). The EIB Group already holds 30% of the EU venture-debt market and 24% of VC (EIB Investment Report, p.8–15); the EIF backs 40–50% of VC-backed EU start-ups in a typical year (EIB scale-up study, p.8–10). ETCI 1.0 committed €3.55bn as a cornerstone for growth funds of at least €1bn, but only about 10% had reached scale-ups by end-2024 (ETCI evaluation, p.2–4). Atomico estimates that matching US pension allocation to VC would add about $210bn over a decade (Atomico 2025, p.43).
Design details and numbers. Three layers. (a) Infrastructure debt: EIB/InvestAI guarantees covering 30–50% of senior project loans for label-qualified campuses with public offtake, plus a 10–15% first-loss tranche funded from the European Competitiveness Fund’s €51.49bn Digital Leadership window (Competitiveness Fund proposal, p.10). (b) Infrastructure equity: a European Sovereign Compute Infrastructure Fund, target €20–30bn, with an EIB/national-promotional-bank cornerstone of about a quarter and the rest from pension funds and insurers, enabled by SIU reforms and a solvency treatment for contracted digital infrastructure. (c) Growth equity for the platform, AI and agent layers through ETCI 2.0 (reported at about €15bn target; press) and the Scaleup Europe Fund. What can a STACKIT-type operator attract? Its campuses can attract infrastructure capital of several billion euros each once offtake exists — the Telekom–Schwarz talks reportedly involved Brookfield as financier (heise). Its software and AI layer is a venture-and-growth story — Europe’s AI start-ups raised over $21bn in 2025 but invest about 12 times less than the US at late stage (Prosus and Dealroom, p.3, p.8).
Risks and what failed elsewhere. Guarantees without offtake simply transfer risk to the taxpayer; SAKURA internet reported a loss on subsidised GPU capacity in 2026 (web context). Rapid hardware depreciation (Epoch finds AI supercomputer hardware cost doubling yearly — Epoch AI) argues for financing the shell and power on long tenors and the chips on short, contract-matched tenors.
The agentic engine. Portfolio-risk agents track utilisation, contract cover and hardware residual value across the fund in real time (Report 3).
First moves and owner. EIB Group; Commission (DG ECFIN, DG FISMA for SIU and solvency treatment); national promotional banks (KfW, Bpifrance, CDP, Invest-NL, NRB in Czechia) as co-cornerstones.
Instrument 6 — Gigafactory Reform: Fewer, Bigger, Faster, Conditional
In short. Concentrate the gigafactory programme on three to five awards with deeper public participation for winners that combine site, power, offtake and an EU-controlled operator — and convert the rest of the envelope into competitive regional grants that reward permitting speed.
Why it ranks sixth. It is the largest existing pot of AI-infrastructure money; reforming it costs nothing new.
Evidence and world practice. The regulation’s 17%-plus-match structure (gigafactory regulation, p.8) and its EU-control requirement for the coordinator (p.4) are sound. What is weak is dispersion: 76 expressions of interest, a 100,000-processor bar already overtaken by private clusters (xAI’s Colossus reached 200,000 chips in 2025 — Epoch AI), and German bidders fragmented. Bruegel proposes turning gigafactory subsidies into competitive regional grants that reward fast permitting and attracting allied hyperscalers’ chips into EU data centres (Bruegel on AI compute, p.1).
Design details and numbers. Allow the regulation’s alternative form — a guaranteed purchase of compute time worth up to 17% of capex — to become the default, so public money buys capacity rather than grants it. Score bids on secured power date, offtake already signed, sovereign label, and conditionalities (Instrument 7). Encourage national champions to bid jointly rather than separately.
Risks and what failed elsewhere. Concentration risks political backlash from states that lose; the regional-grant window answers that. Over-building training capacity when demand shifts to inference and agents is a real risk (Report 3): require that at least half of each gigafactory’s capacity be offered for inference and public-sector use.
The agentic engine. Evaluation agents score bids consistently against published criteria and make the scoring auditable (Report 3).
First moves and owner. EuroHPC JU and the Commission, in the current resubmission round (large-site deadline 3 December 2026).
Instrument 7 — Conditionalities: Public Money Buys Public Goals
In short. Attach a standard set of conditions to every public euro in the sovereign tier — offtake, guarantees, gigafactory grants, zone status: open APIs and portability, fair-access pricing for SMEs and researchers, reinvestment commitments, interoperability through the EuroCloud Federation, and change-of-control protection.
Why it ranks seventh. Without conditions, support for two or three operators simply creates two or three new incumbents. Conditions keep the sovereign tier contestable and prevent the “sovereignty paradox” ECIPE documents, where domestic winners resell hyperscaler infrastructure.
Evidence and world practice. Mazzucato and Rodrik set out a taxonomy of conditions attached to grants, calls, tax breaks and partnerships — access pricing, reinvestment, openness — so that public money delivers public goals (Mazzucato and Rodrik, p.1). The new economics of industrial policy finds that what matters most is not the instrument but how the public–private relationship is managed over time, the “embedded autonomy” that separates Korea’s successes from failures elsewhere (Juhász, Lane and Rodrik, p.2, p.33). The Commission’s framework already scores change-of-control assurances under its strategic objective (Cloud Sovereignty Framework, p.2–3). The legal opinion for the German Interior Ministry adds a condition specific to Europe’s own champions: an EU provider with a US branch or US customers may itself fall under US jurisdiction (German legal opinion, p.5–6) — so sovereign-tier entities must be ring-fenced from any US-facing business.
Design details and numbers. A standard conditionality annex: (1) open, documented APIs and free bulk export; (2) at least 10–15% of subsidised capacity offered at published access prices to SMEs, universities and public research, mirroring CISPE’s “Sovereign Cloud Credits” (CISPE manifesto, p.2–4); (3) a clawback if control passes outside the EU within ten years; (4) a ring-fenced sovereign entity; (5) EED reporting and waste-heat use.
Risks and what failed elsewhere. Gaia-X shows what over-engineered governance does: EuroStack lists its failure modes as competing national agendas, red tape, lack of scalability and lack of common vision (EuroStack report, p.15). Keep the annex short and identical across instruments.
The agentic engine. Conditionality-monitoring agents check API availability, access-price publication and capacity shares continuously (Report 3).
First moves and owner. Commission (DG COMP and DG CNECT) as a standard State-aid and CADA annex.
Instrument 8 — Enforce Contestability: Egress, Licensing, Gatekeepers
In short. Use the Digital Markets Act, the Data Act’s switching provisions and Article 102 to remove egress charges, licensing penalties and bundling that keep customers from moving to any alternative — European or otherwise.
Why it ranks eighth. It is the precondition for a sovereign-tier operator to grow beyond its tier, and it benefits customers everywhere.
Evidence and world practice. The CMA’s three adverse effects (UK CMA decision, p.16–17), the 28% licensing penalty alleged by CISPE (CISPE complaint, p.3–4), and the “limited” uptake of 2026’s voluntary egress cuts (CMA update 2026, p.3–4). In France, the big three captured 80% of the growth in public cloud spending (French competition authority, p.6). IE University’s Center for the Governance of Change calls software licensing the incumbents’ “post-regulatory” lever and recommends procurement support for emerging European providers (IE University, p.3–4). Even the Open Cloud Coalition–commissioned ECIPE study estimates large GDP gains from removing barriers to cloud choice (ECIPE 2025, p.2, p.8).
Design details and numbers. Designate AWS and Azure under the DMA (preliminary findings June 2026, web context); impose licence-price parity regardless of the cloud on which software runs; zero egress for switching and multi-cloud; publish switching rates annually.
Risks and what failed elsewhere. The 2024 CISPE–Microsoft settlement covered European providers but not AWS or Google (EuroStack report, p.66) — settlements fragment remedies. Prefer general obligations over bilateral deals.
The agentic engine. Switching agents that migrate workloads and estimate exit costs make the legal right to switch a practical one (Report 3).
First moves and owner. Commission (DG COMP); national competition authorities; coordination with the CMA.
Instrument 9 — A Domestic Compute Tax Credit for Users
In short. A tax credit for firms — especially SMEs and AI start-ups — that lease compute and AI services from label-qualified capacity located in the EU.
Why it ranks ninth. It moves private demand, not just public, while leaving firms free to choose among qualified operators.
Evidence and world practice. The French Commission on AI proposed exactly this, alongside its usage guarantee (French AI Commission, p.11). India subsidises users per GPU-hour on privately operated, auctioned capacity — ₹65 per GPU-hour, with 38,000 GPUs onboarded against a 10,000 target (IndiaAI factsheet, p.1–4). Canada’s AI Compute Access Fund (up to C$300m) subsidises SME compute (web context). AI adoption is low in Europe: only 13.5% of EU businesses and 12.6% of SMEs use AI (Apply AI Strategy, p.3–4).
Design details and numbers. A 20–30% credit on qualifying spend, capped per firm, for five years; cost at the order of €1–1.5bn a year EU-wide (ENSI estimate, labelled), administered nationally under a common EU framework so it does not fragment the market.
Risks and what failed elsewhere. India’s portal is reportedly underused and its subsidy concentrated in a few model builders (web context). Tie the credit to adoption outcomes, not capacity onboarding.
The agentic engine. Eligibility agents verify qualifying spend directly from billing data (Report 3).
First moves and owner. Member states (finance ministries), with a Commission framework under the Apply AI Strategy.
Instrument 10 — A Capped Economic-Security Capex Subsidy
In short. Allow member states, under a CADA strategic-project designation, to co-fund up to a capped share (20–30%) of the capex of label-qualified sovereign campuses — only where offtake and grid dates are already secured.
Why it ranks tenth. Direct capex support works, but only on top of demand, rules and capital. In isolation it repeats Cloudwatt.
Evidence and world practice. Japan designated cloud programmes “specified critical products” under its Economic Security Promotion Act and certified SAKURA internet’s ¥100bn GPU-cloud plan, half covered by subsidy — while admitting SAKURA to its Government Cloud (Sakura investor deck, p.3, p.6; Sakura report, p.2). Airbus’s launch aid ran higher still (Airbus study, p.60). Mecklenburg-Vorpommern’s up-to-€120m aid for Schwarz’s Dummerstorf site shows regional aid is already being used, case by case (ZDFheute).
Design details and numbers. Cap at 20–30% of shell, power and network capex (below Japan’s 50%, because the EU stacks offtake and guarantees on top); higher caps only in cohesion regions; conditionality annex mandatory.
Risks and what failed elsewhere. SAKURA’s 2026 GPU loss (web context) shows subsidised capacity still needs paying customers; subsidising GPUs that depreciate in three to four years is the weakest use of public money — fund the long-lived assets.
The agentic engine. Not central; monitoring as in Instrument 7.
First moves and owner. Member states under State-aid rules; DG COMP to adopt a CADA-linked framework.
Instrument 11 — A Sovereign Cloud Mission Office with Kill-or-Scale Discipline, and KPIs That Someone Owns
In short. A small, empowered mission office — ARPA-style in method, not in research focus — that owns the targets, runs the offtake portfolio, scores the operators and kills what does not work.
Why it ranks last but binds the rest. Every ECA review cited here criticises missing targets and coordination. A mission office is how the other ten instruments become one programme.
Evidence and world practice. The ARPA bundle that transfers is programme-manager problem definition, kill-or-scale portfolio discipline, speed and ex-post accountability; the binding constraint in Europe is “demand architecture” — SPRIND has no buyer-mapping function (Harvard on ARPA in Europe, p.10–12). The Dutch benchmark of DARPA, ARPA-E, ARIA and SPRIND finds the model works only with transition partners and early customers (Dutch ARPA benchmark, p.4–7). The Luiss report proposes an EU industrial-policy agency to scrutinise spending (Luiss report, p.14–15).
Design details and numbers. Fewer than 100 staff; budget for programme work of about €100–150m a year; publishes a quarterly scorecard against the KPIs below; empowered to recommend withdrawal of offtake or zone status.
The agentic engine. The mission office is the natural first customer for the full agentic engine of Report 3: scanning, forecasting, compliance and portfolio agents in one sovereign deployment.
First moves and owner. Commission, hosted jointly by DG CNECT and DG GROW with an EIB liaison; directors appointed on fixed terms.
How much public money is justified
The figures below are an ENSI estimate, built bottom-up from the instruments above for 2027–2034. They are labelled as such and should be read as an order of magnitude, not a budget line.
Sovereign-tier offtake (Instrument 1): €25–35bn over eight years, ramping to €3–5bn a year — purchases of services the public sector needs anyway.
Guarantees and first-loss tranches (Instrument 5a): €10–13bn of budget provisioning, backing roughly €30–40bn of senior project debt.
Infrastructure fund cornerstone (Instrument 5b): €6–8bn of public equity in a €20–30bn fund.
Gigafactory envelope (Instrument 6): about €8bn — the existing EuroHPC contribution of up to €4.12bn (gigafactory regulation, p.3, p.6) plus member-state matching, redirected rather than added.
Capex subsidy (Instrument 10): €12–15bn of national State aid, capped per project.
Compute tax credit (Instrument 9): €8–10bn of tax expenditure.
Mission office (Instrument 11): about €1bn.
Total: roughly €70–90bn, of which about €45–55bn is purchases, guarantees and equity — recoverable or returning value — and about €25–35bn is true grant or tax expenditure.
For scale: the EuroStack proposal asks for €300bn over ten years (EuroStack report, p.9); the proposed European Competitiveness Fund’s Digital Leadership window alone holds €51.49bn for 2028–2034 (Competitiveness Fund proposal, p.10); the Digital Decade national roadmaps already list €205.1bn of public measures, more than half of them expiring in 2026 as the Recovery Facility closes (Digital Decade 2025, p.8, p.10). The envelope needed is moderate. What is missing is not the money but its shape.
World best practice, one line per country
United States — certify once (FedRAMP’s presumption of adequacy), concentrate demand (JWCC up to $9bn), deregulate the build (permitting exclusions, federal land); worked for scale, but every dollar went to incumbents (FedRAMP memo; US AI Action Plan).
United Kingdom — a three-layer compute portfolio in which sovereign compute is “almost certainly the smallest component”, up to £2bn for public research compute, and AI Growth Zones of 500 MW+ with power priority; strong on sites, weak on a domestic champion (UK AI action plan, p.8–9; UK compute roadmap, p.7, p.12).
Japan — the most complete champion package: economic-security designation, a 50% capex subsidy for SAKURA internet, and a conditional Government Cloud slot; the subsidised capacity still struggled for customers (Sakura investor deck).
France — SecNumCloud immunity for sensitive data, cloud-first with justification above €1m, and the €100m gigafactory offtake; protected demand works, but at a 25–40% cost premium and with a narrowed scope since 2023 (French cloud doctrine, p.9; French Court of Audit, p.9–11).
Canada — C$2bn split between private data centres, public supercomputing and an SME access fund, after The Dais found Canada last in the G7 on AI compute; it defines sovereignty by ownership but has no answer to supply-chain disruption (The Dais, p.6).
India — the state does not build: it runs lowest-bid reverse auctions on private clouds and subsidises users per GPU-hour; fast to scale capacity, slow to scale use (India GPU update, p.3–6).
China — central planning of compute geography (”East Data, West Computing”), with new builds barred where utilisation is under 50%; effective siting discipline at the cost of market signals (China’s compute plan, p.3–4).
South Korea — the state buys GPUs and places them with domestic cloud champions (NAVER Cloud, NHN, Kakao); a direct champion model, untested at scale (web context).
UAE and Saudi Arabia — sovereign capital plus cheap energy traded for US security conditions (Stargate UAE, G42, HUMAIN); fast, but sovereignty conditional on Washington’s licences (web context).
The pattern is unambiguous: the states that built domestic capacity paired supply support with demand they controlled; the states that only built sites or research machines ended up hosting someone else’s cloud. State-backed AI supercomputing clusters in Europe and Central Asia grew from 3 in 2018 to 44 in 2025, mainly through EuroHPC (AI Index 2026, p.333) — Europe builds public research machines well. Its gap is commercial sovereign cloud at scale.
How the EU should act: who does what
The Commission sets the rules of the market: the two-level label and presumption of adequacy (Instrument 2), the award criterion (3), the conditionality annex (7), DMA enforcement (8), and acts as joint-procurement agent (1). CADA is the vehicle; it must be passed with its numeric targets intact.
The EIB Group carries the capital stack (5): guarantees, first-loss tranches, the fund cornerstone, and ETCI 2.0 for the software and AI layers.
EuroHPC JU reforms the gigafactory awards (6) and ties them to offtake and inference access.
A founding coalition — France, Germany, the Netherlands and the Nordics — commits the first order book. France brings the strongest legal tier; Germany brings the largest administration, the federal AI platform and the Deutsche Verwaltungscloud standards (German administration cloud strategy, p.7–8); the Netherlands brings the most rigorous legal analysis of US exposure; the Nordics bring uncongested, clean grids.
Mid-sized states such as the Czech Republic should not wait for a 27-way consensus. Their leverage is as early anchor buyers and site hosts: commit a share of national critical-tier workloads to the joint offtake, offer a Sovereign Compute Zone on a strong grid node, adopt the label by mutual recognition before it is mandatory, and route the national promotional bank into the infrastructure fund. A state of ten million cannot build a hyperscaler; it can make itself one of the first and most reliable customers of two or three European ones, and host part of their capacity.
The operators — STACKIT, OVHcloud, Scaleway, IONOS, Telekom’s T Cloud and others — must accept the conditions: consolidate bids, ring-fence sovereign entities, open APIs, publish external revenue and utilisation.
The first twelve months
Months 1–3: Council conclusions endorsing the two-tier definition of sovereignty and the founding offtake coalition; Commission publishes the conditionality annex; EuroHPC adds offtake and power-date criteria to the current gigafactory scoring.
Months 3–6: Label specification adopted with C5 and SecNumCloud mapping; first joint-procurement notice for sovereign-tier compute; member states designate candidate Sovereign Compute Zones; ACER and ENTSO-E consult on use-it-or-lose-it queue rules.
Months 6–9: EIB launches the guarantee and first-loss window; fund cornerstone commitments from at least four national promotional banks; first DMA gatekeeper designations for cloud.
Months 9–12: First offtake contracts signed with at least three label-qualified operators; mission office operational with its first public scorecard; tax-credit framework agreed; the procurement-directive revision tabled with the security-of-supply criterion.
The KPIs that should be published every quarter
European providers’ share of the EU cloud market (baseline about 15% — CADA proposal, p.3); target 25% by 2032.
Number of EU-controlled operators with more than 1 GW of contracted capacity; target three by 2032.
Label-qualified sovereign capacity in operation (GW) and its utilisation rate.
Share of critical-tier public workloads on label-qualified services; target 100% by 2035, as CADA intends.
Median grid-connection time in designated zones; target under 18 months.
Annual switching rate among cloud customers (baseline under 1% in the UK).
Volume and utilisation of the joint offtake; private capital mobilised per public euro.
External (non-captive) revenue of sovereign-tier operators — the single most honest test of whether a champion is real.
Close: what to do first
If Europe does only one thing, it should commit to buy. A guaranteed, aggregated, label-gated order book for the sovereign tier is the instrument that makes everything else bankable, and it is the one every successful precedent has had and every failure lacked. If it does three things, it should add the certify-once label and grid priority — two rule changes that cost almost nothing and remove the largest fixed costs a European operator faces.
The Schwarz Group’s experiment shows the opportunity and the limit at once. A private European company has already announced up to €16.6bn of data-centre investment, reached the top sovereignty level awarded in the Commission’s own tender, and won federal anchor contracts — without an order book, without a single market for its certifications, and joining the same grid queue as everyone else. Europe does not need to invent a champion. It needs to stop making the ones it has fight with one hand tied behind their backs — and to hold them, in return, to conditions that keep the market open. The question is no longer whether Europe can afford a sovereign cloud. It is whether it can afford to keep asking its own builders to win a market it refuses to organise.
