Sovereign AI: How Europe Wins by Refusing to Run the Same Race

European sovereign AI will not be built by matching US or Chinese spending. Europe controls roughly 5% of global compute versus 80% for the United States — a head-on race is lost before it starts. Europe wins differently — by leveraging its asymmetric strengths (lithography, low-carbon electricity, talent, industrial data, trust) rather than mimicking the excess of its rivals.

Why Europe should not try to match American scale

There is a difference between sumo and judo. Sumo is won by mass: being bigger, pushing harder. That is precisely the race playing out between Washington and Beijing — whoever amasses the most compute and capital. Judo, by contrast, does not pit strength against strength: it uses the opponent's momentum to unbalance them. David Yoffie and Mary Kwak of Harvard built a theory of business strategy around this in Judo Strategy: movement, balance, leverage. Europe's mistake is asking how to enter the sumo ring. The right question is: why would it want to?

The arithmetic settles the debate. An MIT scenario report estimates that Europe controls only 5% of global compute versus 80% for the United States, and that Europe's €200 billion InvestAI commitment is dwarfed by a single year of American spending. European industry pays roughly twice the American electricity rate. And the four announced AI “gigafactories” will not be operational before 2027–2028, on public-procurement timescales.

Put differently: if Europe enters the sumo ring, it loses before the fight begins. Less compute, less capital, more expensive electricity, slower institutions. Every euro spent imitating American scale is a euro spent losing more slowly. The honest conclusion is not surrender but redirection: find the dimensions where the giants' mass becomes a handicap.

The Chinese lesson: the logic, not the tactic

In January 2025, Nvidia lost $589 billion in market capitalisation in a single session — the largest single-day loss in stock-market history. The cause was not a war or a recession, but simply a cheaper Chinese AI model. DeepSeek, denied access to top-tier Nvidia chips by US export controls, trained on deliberately throttled H800s and turned constraint into advantage: a Mixture-of-Experts architecture, multi-head latent attention, FP8 mixed-precision training. The scarcity meant to strangle it produced the kind of efficiency that resource-flush American labs had no reason to pursue.

The myth must be corrected: the famous “$5.6 million” covered only a single training run at notional rental rates. SemiAnalysis estimates the real cost at around $1.3 billion in server investment. The lesson is therefore not that “a cash-strapped startup beat the giants”. It is sharper: a challenger, with a fraction of the resources, won by refusing its opponents' terms — efficiency over scale, openness over enclosure, speed over capital.

Europe must copy this strategic logic and reject the questionable tactic. The logic is sound: choose the dimension where the adversary is weakest, turn its mass against it, move faster than it can react, and commoditise the layer it monetises. The tactic to discard is free-riding — training your own models on a rival's costly outputs: beyond the legal risk, it is a one-off exploit that invites exactly the retaliation it provokes. Durable sovereignty is built on assets a rival cannot revoke.

What are Europe's asymmetric AI advantages?

Europe possesses more of these dimensions than the declinist narrative admits. They fall into two groups: assets it already holds, and games it can still choose to play.

Its asymmetric advantages span six areas. In lithography, ASML holds ~100% of the EUV market and ~94% of global lithography — the chokepoint of every AI chip ever made, with no GPU possible without these machines. In low-carbon electricity, France runs on a mix that is ~95% low-carbon at ~20 gCO₂/kWh, offering stable, clean, affordable power for compute. In AI talent, Europe produces ~30% more AI professionals per capita than the United States. In industrial data, companies such as Siemens, Schneider, Bosch, Airbus and ASML hold the proprietary datasets on which the next AI paradigm will run. In the agent layer, Mistral and a cohort of fast-moving teams are well positioned. And in trust and regulation, the AI Act and GDPR make provable compliance the product, not the friction.

The chokepoint nobody mentions

Every compute race between the United States and China runs through one European company. ASML, in the Netherlands, holds a near-monopoly on EUV lithography: 100% of the EUV machine market and roughly 94% of global lithography, according to market analyses (Tech Market Briefs). These are the machines that etch every advanced chip on Earth — no Nvidia GPU, no Apple silicon, no AI revolution without them. A position built over three decades, more than €10 billion in R&D, and more than 16,000 active patents. The implication overturns the declinist narrative: Europe already holds the most defensible position in the entire AI stack — its base.

The cleanest, cheapest electricity on the continent

Compute is, at bottom, electricity disguised as silicon. France ran on a mix that was close to 95% low-carbon in 2025, with a carbon intensity of roughly 19.6 gCO₂/kWh — against a European average of around 175 gCO₂/kWh, according to RTE (RTE 2025 Analyses and Data). A data-centre load is a very stable, dispatchable baseload — exactly the profile nuclear serves best. Europe cannot outpace American compute, but France and the Nordic countries can offer what the saturated American grid struggles to provide: abundant, clean, cheap electricity for training and inference. An advantage measured in decades, not quarters.

The deepest talent pool

Europe trains more AI talent than either rival: roughly 30% more professionals per capita than the United States, and nearly three times as many as China. Its weakness is not the pipeline but retention — it exports its best minds. Yet the current is reversing: hardening US immigration policy is pushing researchers back across the Atlantic, and the “Choose Europe” initiative (€500 million) is trying to keep them. The move is not to produce more talent, but to stop giving it away.

The data the giants don't have

The frontier is shifting from text to physical AI — robotics, agents, industrial systems. American labs, built on consumer-internet data, do not hold the factory-floor, machine-tool, power-grid, aerospace, and precision-manufacturing data on which the next paradigm will run. Europe does: Siemens, Schneider, Bosch, Airbus, ASML, automotive, pharmaceuticals. This is uncontested water — competing where the giants have no foothold rather than where they are entrenched. It is also the core of our sovereign thesis: whoever holds the data holds the asset.

Privacy as a primitive of sovereignty

Sovereignty that stops at hardware and energy remains incomplete. The missing layer is cryptographic. For a bank, a hospital, a defence ministry or a government agency, data residency and provable compliance are not friction: they are the product. This is the Brussels Effect turned into a commercial strategy — the jurisdiction that writes the rules can sell the only stack that natively satisfies them.

Fully Homomorphic Encryption (FHE) takes this logic to its limit: it allows computation on encrypted data without ever decrypting it. Privacy ceases to be a contractual promise and becomes a mathematical guarantee. It is a primitive of sovereignty, not a feature. The global leader in the field, Zama, is a French unicorn we back. From this building block was born Zaïfer, a Zama × PyratzLabs joint venture dedicated to confidential, compliant on-chain finance (as reported by DL News) — demonstrating that privacy can become the infrastructure of regulated markets, not their obstacle.

This is precisely the terrain we intend to finance next: a regulated fintech carrying confidential tokenised funds, at the intersection of FHE, our asset-management activity, and an anticipated UCITS licence. Sovereignty is not a slogan; it is an ownership structure — and privacy is now its technical foundation.

Owning the agent layer and becoming the third pole of trust

The value of AI is migrating from the model to the agent — systems that plan and execute multi-step work rather than respond to a single prompt. This layer rewards small, fast teams, not capital-heavy incumbents: it is the only place where a European start-up can ship every week while a large lab is still deliberating. The goal is not to clone OpenAI, but to own the orchestration and application layer that sits above any model. That is the operator logic we apply across our portfolio, from Kiln to MrChief.ai.

The most underrated advantage remains trust. A large share of the world wants AI that depends neither on Washington nor on Beijing. Europe's rule-of-law posture — often mocked as a handicap — is exactly the brand a neutral buyer would pay for: governments, regulated industries, states unwilling to hard-wire their dependence on a superpower. Sovereignty as a service, anchored in a single market of roughly 450 million people and more than twenty languages.

Two older ideas must anchor the discipline. Gary Hamel and C.K. Prahalad's “strategic intent”: resource-poor challengers win by setting ambitions disproportionate to their current means. And Alexander Gerschenkron's “advantages of backwardness”: late industrialisers leapfrog stages by adopting the frontier directly. Read correctly, Europe's deficits are the preconditions for a leap. One caveat the data imposes: Europe does not lack capital — it lacks channels, with considerable household savings sitting idle. Mobilising this patient capital towards its own asymmetric bets is the financial half of the same strategy. That is precisely the role of a Paris-listed investment company anchored in Europe.

Frequently asked questions

What is sovereign AI in Europe?

Sovereign AI refers to artificial intelligence whose underlying technology — chips, energy, data, models and infrastructure — is owned and governed from within Europe, rather than leased from a foreign superpower. It rests on control of the value chain and on verifiable privacy and compliance guarantees.

Why can't Europe catch up with the United States and China in the AI race?

Because the head-on race is played on scale: compute and capital. Europe controls roughly 5% of global compute versus 80% for the United States, pays higher electricity prices, and advances on slower institutional timelines. Mimicking that excess amounts to losing more slowly.

What are Europe's asymmetric AI advantages?

Lithography (ASML's near-monopoly on EUV), low-carbon affordable electricity (French nuclear), one of the world's deepest talent pools, proprietary industrial data, the agent layer, and regulatory trust (AI Act, GDPR).

Why is privacy (FHE) a sovereignty issue?

Fully Homomorphic Encryption allows computation on encrypted data without decrypting it. For regulated sectors, this mathematical guarantee turns compliance into a commercial advantage and enables the building of sovereign infrastructure where data never leaves its owner's control.

How does Pyratz Corp. fit into this thesis?

By deploying capital and operators into European-anchored frontier-tech companies — including Zama (FHE) and the Zaïfer joint venture — and by funding a layer of confidential, compliant finance from a Paris-listed company.

To follow this thesis and our progress, consult our investor relations and read our founding article Asymmetric Innovation.

This does not constitute investment advice. Pyratz Corp. (MLPTZ, ISIN FR0013371507) is listed on Euronext Access Paris; trading has resumed on Euronext Access Paris following the re-listing.