Everything About AI
NextFin News - Europe has committed roughly €200 billion to build artificial-intelligence infrastructure under its AI Continent Action Plan, a sum that would sound decisive in almost any other industry. In 2025 alone, the United States attracted about $286 billion in private AI investment. Europe attracted about $21 billion. That is the tension at the heart of the region's sovereignty push: a plan sized like a superpower, backed by capital flows that are not.
The question put to Suzanne Ashman, managing partner at the Sovereign AI Fund, and Johan Bergqvist, chief financial officer of Mistral AI, at a technology conference in London this week was whether Europe can build AI sovereignty at scale. The short answer is yes — but only at selected layers of the stack, and only if Brussels abandons the dream of full-stack autonomy. Full sovereignty over chips, cloud, models, and applications is economically prohibitive for Europe; managed sovereignty over compute location, public-sector procurement, and open-weight models is within reach. Getting the layers right, rather than pursuing sovereignty in general, is the strategic challenge of 2026, as policy analyst Swarnim Shrivastava has argued.
The Situation: A €200 Billion Mobilization Against a Moving Frontier
The European Commission's AI Continent Action Plan, unveiled in 2025 and built around the InvestAI initiative, aims to mobilize approximately €200 billion of investment. The structure matters: it is not a €200 billion budget appropriation. It is a €20 billion European fund designed to co-finance roughly €200 billion of private capital, supporting up to five AI gigafactories, 19 smaller AI factories, and 13 antennas linked to existing EuroHPC supercomputers. By mid-2025, the Commission had received 76 expressions of interest, spanning 60 sites across 16 member countries. The operational target is to triple European data-centre capacity within five to seven years.
The ambition is not rhetorical. Europe has real champions. Mistral AI, the Paris-based frontier lab, secured $830 million in debt financing in March 2026 from a seven-bank consortium led by France's Bpifrance to purchase 13,800 Nvidia Grace Blackwell GB300 GPUs for a new 44-megawatt data centre near Paris. ASML committed €1.3 billion for roughly an 11 percent stake in the company, which last raised equity at an €11.7 billion post-money valuation. A separate joint venture backed by Bpifrance, UAE fund MGX, Nvidia, and Mistral is building a 1.4-gigawatt AI campus near Paris, with construction expected in the second half of 2026 and operations by 2028. Germany and Canada launched a Sovereign Technology Alliance in February 2026 to coordinate on sovereign AI capabilities.
Yet the arithmetic of the race keeps changing. While Europe assembles its financing package, the frontier has already moved to a scale that makes the target recede. The gap is not just about money; it is about installed capacity, operating control, and the concentration of the stack's most valuable layers outside Europe.
The Analysis
The Stack Is Not One Thing — and Neither Is Sovereignty
The first mistake in the European debate is treating AI sovereignty as a binary condition: either Europe controls the whole stack or it controls nothing. In practice, sovereign AI is a portfolio of decisions across layers. At one end sits full-stack sovereignty — domestic chips, domestic cloud, domestic models, domestic applications. It is politically persuasive and capital-prohibitive. A nation that tries to own every level is not buying independence; it is reinventing a wheel built elsewhere, at a cost of hundreds of billions of dollars and years behind the curve.
More limited forms are both cheaper and more defensible. Compute sovereignty means ensuring that sensitive workloads run on infrastructure governed by European jurisdiction, even if the foundation models are global. Application sovereignty means adapting existing models to European languages, legal systems, and public services rather than matching frontier scale. Sovereign AI as a service means renting isolated cloud regions from global providers with contractual guarantees on data location and operational control. Europe's actual strategy is a hybrid of all four, which is both its strength and its vulnerability: it does not need to win every layer, but it must win the layers that matter.
The Capital Gap: Structural, Not Cyclical
The headline number deserves scrutiny. Stanford University's AI Index reports that U.S. private AI investment alone reached about $286 billion in 2025, against roughly $21 billion across Europe — a gap of more than 13 to 1. Europe's entire multi-year public mobilization is therefore comparable to less than a year of American private AI spending. This is a structural gap, not a cyclical fluctuation. It reflects the depth of U.S. capital markets, the concentration of the world's largest technology companies, and the fact that AI infrastructure is now a scale business where the leader's cost of capital becomes a competitive weapon. A company that can borrow $830 million against a GPU purchase — as Mistral did — is already an outlier in Europe. In the United States, that transaction would be routine.
The implication is uncomfortable: Europe cannot outspend the United States, and it should not try. It must out-prioritize it — concentrating capital on the layers where sovereignty actually changes outcomes rather than dispersing it across the whole stack.
The Compute Reality: 2 Gigawatts Against 35
Money eventually becomes silicon and electricity, and there the gap is starker. Research from the Bruegel think tank estimates that the European Union had about 2 gigawatts of operational AI compute capacity in 2026, roughly 5 percent of global installed capacity. The United States had about 35 gigawatts, or 78 percent. China had about 5 gigawatts, or 11 percent. Even as global AI compute capacity is projected to grow more than eight-fold by 2031, Bruegel finds that the relative distribution is not projected to change fundamentally — meaning Europe's share stays near the low single digits even if it builds aggressively. An independent industry observatory corroborates the picture: Europe holds under 5 percent of global AI compute, while U.S. hyperscalers command roughly 70 to 72 percent of the regional cloud market.
Tripling data-centre capacity in five to seven years would lift Europe toward the low teens of global share — meaningful progress, not parity. And capacity is not the same as capability: the most advanced AI accelerators remain subject to export controls and supplier concentration, and the newest generation of chips is typically allocated first to the hyperscalers that pre-ordered them.
The Cloud Paradox: Europe Builds, Americans Operate
Here is the uncomfortable fact that no sovereignty plan has fully solved: building a data centre in Europe does not mean Europe controls it. The cloud market structure makes this concrete. The three U.S. hyperscalers — Amazon, Microsoft, and Google — together accounted for approximately 70 percent of the roughly $60 billion European cloud market in 2024, rising to more than $70 billion in 2025. European-headquartered providers such as SAP, Deutsche Telekom, OVHcloud, and Orange share most of the remaining 30 percent. Locating AI infrastructure in the EU is a necessary but insufficient condition to preserve sovereignty, Bruegel notes, because the model layer may still be cut off even when the infrastructure is local.
This is where the proposed Cloud and AI Development Act, or CADA, enters — and where the debate gets contested. Adopted as a Commission proposal in June 2026, CADA establishes four cumulative sovereignty levels for public procurement, based on local-content requirements for hardware and software components. The logic is that public money should buy sovereign infrastructure. But Bruegel warns the criteria risk the reverse: they would make it more costly to build infrastructure in the EU, potentially slowing the shift of AI compute load from the United States to Europe and leaving the bloc with less sovereignty, not more. The think tank's alternative is to target AI service outputs — the tokens consumed in the EU — by requiring that compute services for European customers be located in the EU, rather than policing the nationality of the hardware inside the racks.
The distinction matters because it defines what victory looks like. If sovereignty means the metal sits in Europe, the plan can succeed. If it means Europe controls the compute, the models, and the terms of access, the plan succeeds only if it addresses the operating layer — the software, the orchestration, and the contractual control that determines who can shut what off. CADA is the mechanism; whether it lowers or raises the cost of sovereignty is still being decided.
The Energy Constraint Nobody Is Pricing In
There is a second-order constraint that turns the capacity target into a grid problem. Ember, an energy think tank, projects that European data-centre electricity demand will rise from 96 terawatt-hours in 2024 to 168 TWh in 2030 and 236 TWh in 2035 — an increase of nearly 150 percent over the decade. In Sweden, Denmark, and Norway, demand is set to triple by 2030. Data centres are becoming one of the fastest-growing sources of power demand in Europe, ahead of electric vehicles and comparable to electrified industry.
Tripling compute capacity in five to seven years therefore requires not just capital and chips but grid connections, transmission buildout, and generation that Europe has historically permitted slowly. A gigafactory without a grid connection is a very expensive warehouse. This is the layer where European industrial policy has genuine leverage — and where delays can silently convert a capacity target into a missed one.
The Counter-Thesis: Why Managed Sovereignty Is Enough
The strongest case against pessimism is that Europe never needed full-stack parity to be sovereign. Sovereignty, on this reading, is about resilience and bargaining power, not autarky. Europe already has regulatory power — its AI Act gives it the ability to set terms for the world's largest AI companies. It has a concentrated, high-value public-sector customer base. It has world-class research institutions and a credible open-weight model ecosystem. And it has the geopolitical alignment to secure chip supplies from the United States and Asia that a genuinely adversarial power would not have.
On this view, the €200 billion plan is not an attempt to beat the United States at its own game. It is an insurance policy: enough domestic capacity to keep public administration, defense, and critical infrastructure running if access to foreign models or cloud is disrupted; enough local champions to give European buyers an alternative at the negotiating table; enough open-weight capability to prevent a single vendor from owning the model layer.
This argument is coherent and, in many respects, correct. But it has a flaw: insurance only works if it pays out when you need it. A 2-gigawatt compute base that remains dependent on U.S. accelerators, U.S. cloud orchestration, and U.S.-led financing does not provide meaningful redundancy in a genuine disruption. The counter-thesis holds only if Europe actually closes the operating-control gap that CADA is designed to address — and only if the regulation lowers rather than raises the cost of building in Europe. Without that, managed sovereignty risks becoming managed dependence with a European flag.
Conclusion: What to Watch
The verdict splits by time horizon. In the short term — the next 12 to 18 months — expect announcements, procurement preferences, and the first wave of AI-factory conversions. Sentiment will improve; the narrative will shift from "Europe is absent" to "Europe is building." That is real, but it is not yet sovereignty.
In the medium term — three to five years — the test is delivery. The base case is that Europe reaches the low teens of global AI compute share, retains its open-weight model champions, and carves out a protected public-sector market for sovereign vendors. The downside case is that grid delays, permitting, and capital costs push the capacity target out while the frontier moves again, leaving Europe with local data centres running foreign models on foreign chips. The upside case is that CADA's rules shift operating control without pricing European builders out of the market, that European procurement consolidates around a small number of sovereign providers, and that the energy buildout arrives on schedule.
In the long term, the structural question is whether Europe can convert regulatory power and public procurement into durable industrial capability — the same question it has faced in cloud, in semiconductors, and in platforms, with mixed results. The AI race is not a sprint Europe can win; it is a dependency-management exercise it can either handle well or badly.
The falsifying signal is specific and quantifiable: if, by 2028, European AI compute capacity remains below 10 percent of the global base while U.S. hyperscaler share of the European cloud market stays above 60 percent, then the sovereignty plan has built infrastructure without control — and the thesis that Europe can achieve meaningful sovereignty at scale is wrong. Watch the capacity tracking from Bruegel, the cloud-share updates from Synergy Research Group, and the CADA implementation timeline. Those three numbers will tell the real story long before the political speeches do.
"The AI Campus will be a transformational infrastructure for France — built in France, to fuel France in the era of AI. It will revolutionize science, education, and industry," Nvidia chief executive Jensen Huang said in announcing the 1.4-gigawatt Paris campus.
Europe's AI sovereignty will not be won in the data centres it builds. It will be won in the contracts it signs, the grids it connects, and the layers it chooses not to fight for.
Explore more exclusive insights at nextfin.ai.
