The Care and Feeding of a European AI Unicorn

Cite this Article
Mikolaj Barczentewicz, The Care and Feeding of a European AI Unicorn, Truth on the Market (November 19, 2026), https://truthonthemarket.com/2026/11/19/the-care-and-feeding-of-a-european-ai-unicorn/

Europe knows how to grow billion-dollar startups. It could stand to keep a few more of them.

Mario Draghi’s 2024 report counted 147 startups founded in Europe from 2008 to 2021 that went on to reach a $1 billion valuation. Forty moved their headquarters abroad, most to the United States. The report also demonstrated how much of the money for the next generation of companies goes elsewhere. U.S. companies received 61% of global funding for artificial intelligence (AI) startups, compared with 17% for Chinese companies and 6% for companies in the European Union (EU).

In their March 2026 Digital Fitness Check comments, my colleagues at the International Center for Law & Economics (ICLE) linked the cumulative burden of EU rules to the departure of innovative firms and a 26.1% drop in monthly venture-capital deals after the General Data Protection Regulation (GDPR) took effect. Europe is still losing talent, and the computing power companies need to train the most advanced AI models remains scarce.

That’s the backdrop to a new open letter from European science, technology, and business leaders. They warn that Europe competes “with a fraction of the capital, the compute, and the scaling capacity of its rivals.” Polish Prime Minister Donald Tusk responded by announcing a “Warsaw High-Level Summit on AI” for early 2027. He wants EU governments to get past “another set of declarations” and work together “for technological sovereignty.” He called on Europe to move beyond a “purely defensive regulatory approach” and start “actively creating the market and expanding technological infrastructure.”

What would that mean in practice? I’m with the letter’s authors on adding computing and electricity capacity and making regulatory simplification legally binding. I’d also make it easier for startups to raise money, compete for talent through better stock-option rules, and expand across borders through a “28th regime”—an optional set of EU-wide business rules alongside the 27 national systems. Warsaw’s proposals for an AI research center and a consortium for industrial AI could help, especially if both can work with partners in allied countries.

I’m more skeptical of the proposed heavy subsidies, government-picked “European champions,” and cloud and government purchasing rules that shut out allied suppliers. Those policies can give favored firms an easier ride while making it harder for European companies to get the chips, models, and computing power they need. The test for Warsaw is whether its plans make it easier for companies to build in Europe and compete anywhere.

A Flag Won’t Power a Data Center

Europe has good reason to worry about who can cut off its access to AI. In June, export controls cut users in allied countries off from two leading AI models overnight. Even a European-owned cloud provider would have had to comply with those controls to keep offering the models.

Europe can gain leverage by giving its allies a stake in keeping access open. If they also rely on cloud and AI services hosted in Europe and governed by European law, they have more to lose from cutting Europe off. As Lazar Radic argues, genuine digital sovereignty means building firms that can compete. And as I argued in “Europe’s Sovereignty Stack”:

The durable way to reduce dependence is not to close the market. It is to make dependence more mutual by building additional cloud and AI capacity that is located, operated, and financed in Europe.

Europe needs a strategy that expands computing capacity and keeps allied suppliers accessible. While some specialized AI tasks can run locally at modest cost, Europe still relies on allies for the computing power to train and run the most advanced models, and for access to proprietary models. The EU’s AI Factories and planned AI Gigafactories would add computing capacity, while Chips Act 2 would bring semiconductor permitting rules into line across member states.

The scale of the shortfall shows why those efforts matter. The European Commission’s impact assessment for the Cloud and AI Development Act (CADA) projects that Europe’s data-center capacity shortfall will grow from nearly 3 gigawatts to roughly 19 gigawatts by 2036, with waits of three to 10 years for a grid connection. In June, Bloomberg reported that only two of the five sites in InvestAI’s €20 billion gigafactory program appeared likely to secure funding before 2028

Tusk and the letter’s authors leave room for foreign firms to help build Europe’s AI capacity. The letter ties access to computing power to whether companies are “building and scaling in Europe,” which makes eligibility a question of where firms operate. The Commission also puts sovereignty at the center of its June CADA proposal. Warsaw should clarify what that means in practice. If a company builds and operates a data center in Europe, should it matter who owns it?

The distinction matters under trade law. The World Trade Organization’s General Agreement on Trade in Services (GATS) carves out qualifying government purchases for governmental purposes. The separate Agreement on Government Procurement governs those purchases where it applies. Outside that carveout, ownership restrictions in commercial or regulated cloud markets could invite GATS disputes and retaliation from the very allies Europe relies on for chips, models, and cooperation on export controls.

Leave the Lab Door Open

The first pillar of Warsaw’s plan, “EMBL for AI,” would create a center that brings together researchers, computing power, capital, and talent for “the development of sovereign AI models.” The open letter makes a similar pitch for institutes modeled on the European Laboratory for Learning and Intelligent Systems (ELLIS), with innovations “made in Europe, for Europe.” Both proposals leave room to work with suppliers outside the EU and to pursue projects beyond the most advanced AI models.

The lab that inspired Warsaw’s proposal shows how a European research center can work with partners beyond the EU. The European Molecular Biology Laboratory (EMBL) counts Switzerland, Norway, Israel, and the United Kingdom among its member states and welcomes associate members from outside Europe.

There’s plenty for such a center to do. In a 2021 paper, Caleb Watney and Dirk Auer suggested that the “multiple cutting-edge AI centers” envisioned in the EU’s white paper on AI could host open-source platforms shaped by the needs of small and midsize businesses. They also proposed setting up research teams with a fixed lifespan and a specific job: building datasets anyone could use. That fits the letter’s call to make more data openly available while protecting privacy and security.

Building a European general-purpose model that can compete with the world’s best is a taller order. More computing power in Europe would help, but it can close only part of the gap. Give the center the freedom to use computing capacity and tools from allied countries so it can get on with the research.

An Industrial Champion Still Needs Suppliers

Warsaw’s second pillar, “Airbus for AI,” would bring firms from across Europe together to develop technologies such as robots and factory equipment that can operate on their own. It echoes the open letter’s call for “deliberately building and backing European champions.”

There’s a case for working together. Robotics and factory automation rely heavily on hardware, an area where Europe already has industrial expertise. Many of these specialized AI tasks can run locally at a reasonable cost. Shared testing facilities and common standards for connecting machines and software are worth exploring.

But the Airbus analogy comes with some baggage. Airbus benefited from nearly four decades of government financing for new aircraft, known as launch aid. The World Trade Organization found that the subsidies caused serious prejudice to U.S. interests. And even Airbus buys its engines from outside suppliers. Its best-selling A320neo flies on engines from Pratt & Whitney or CFM International. Industrial AI likewise depends on a fast-changing mix of models, chips, and systems for collecting and processing data. That’s a lot for one consortium to keep up with on its own.

Picking champions also gives firms a reason to compete for government favor and gives the winners less reason to compete in the market. As Alden Abbott puts it, that’s “picking ‘winners and losers’ through subsidies or special government treatment for favored firms.”

An AI Strategy Needs More Than a Purchase Order

The letter calls for earmarking money for AI in the EU’s 2028–2034 budget to “unlock a multiple of that sum in private investment.” Comparing a seven-year EU budget with a single year of U.S. corporate spending on facilities and equipment tells us little about whether the plan would deliver. The test is how much new investment the program would bring in after you subtract what companies would have invested anyway.

Europe’s more urgent task is to make it easier for companies to build and raise money in the first place. That means simplifying regulation and reforming its capital markets.

The letter turns next to public information technology (IT) spending, arguing that “current EU IT budgets should be directed towards European AI providers.” A government contract can help developers secure financing by serving as a reliable first customer. But Europe’s AI industry faces a different problem: it needs more power, chips, computing capacity, and capital to meet demand.

Moreover, if “Buy European” rules exclude firms based on who owns them, public agencies could just end up paying for weaker technology. For most government tasks that require the most advanced models, agencies can manage the risks by keeping control of their own encryption keys, choosing systems that let them move data and applications to another provider, planning that move in advance, and tailoring security to the task.

If governments insist on preferences, they should reserve EU-only supplier rules for core national-security or intelligence work that can proceed without the most advanced models. Cyber defense should stay outside that restriction because it often requires access to those tools. For other purchases, cap how much a supplier’s European status can count toward winning a contract. The CADA proposal’s recitals suggest giving that preference 15 of the 120 points available when agencies score bids.

Twenty-Seven Ways to Keep Things Complicated

The letter’s best ideas would cost the EU budget next to nothing. But an optional EU-wide corporate regime, better stock-option rules, a Capital Markets Union that opens up investment from pension funds and insurers, and binding regulatory simplification all run into the same problem. Member states want to keep control of taxes, corporate law, and capital markets. Regulators and their political backers want to preserve their leeway in enforcement.

A “28th regime” that sends founders back to 27 national rulebooks gives them little reason to stay. Easier incorporation only goes so far. Growing firms also need investors willing to finance expansion, and employees need to be able to defer taxes on stock options until they sell the shares they acquired by exercising them.

The same resistance explains why Brussels keeps settling for half-measures. As the letter warns, “The first Digital Omnibus was a welcome start; it must not be the end.” July’s AI Omnibus mostly pushed back deadlines and left rigid rules in place. The package on privacy and data did even less. The Council’s October negotiating position showed how little tweaks to the text can do when enforcement is disproportionate and driven by ideology.

As I wrote in October, the Council kept language on “legitimate interest” as a legal basis for using data to develop AI, but it added little beyond a provision covering sensitive data that remains after efforts to remove it. The Council also dropped provisions that would have let the Commission set binding standards for pseudonymization—replacing identifying details with codes. Firms that met those standards could have gained a presumption of compliance. The Council instead reinforced the role of the European Data Protection Board. Poland was reportedly among the governments that blocked a weak draft in June. The October version was barely an improvement.

“Binding simplification” means taking on the interests that benefit from the current rules. As my ICLE colleagues explain in their Digital Fitness Check comments, the Commission should propose targeted amendments to resolve conflicts among the AI Act, the GDPR, the Digital Markets Act, and the Data Act.

Europe can afford these reforms. Warsaw’s test is whether governments and regulators are willing to give up some control.