In September 2025, a single wire of €1.3 billion left Veldhoven for Paris and quietly rewired the balance of European technology. ASML — the Dutch firm that builds the lithography machines without which no leading-edge chip can exist — became the largest shareholder in Mistral AI, a laboratory founded two years earlier by three researchers who had walked out of Meta and DeepMind. The round valued Mistral at €11.7 billion. Within a week, the word circulating in Brussels, in Bercy, and along the corridors of the Élysée was the same word, and it was not profitable, not scalable, and not even European. It was sovereign.
I have spent the past decade auditing things that people would prefer remained unaudited. In 2017, as a twenty-one-year-old cryptography PhD candidate at UCL, I read fifteen ICO whitepapers line by line and found tokenomics engineered for speculation and dressed in the language of utility; the series that came out of it, "The Soul of Code," reached fifty thousand readers and taught me the lesson I have never been able to unlearn — that the most dangerous object in this industry is a claim repeated by everyone and checked by no one. So when I saw the word sovereign attached to an eleven-billion-euro valuation, my instinct was not admiration. It was a narrower question. Sovereign over what, verified by whom, and — the part the market always forgets — for how long?
There is a version of this story that Europe tells about itself, and it is a good story. After two decades of watching American hyperscalers harvest its graduates and Chinese laboratories harvest its data, the continent finally has a champion of its own: fluent in French, generous with open weights, headquartered in a converted Parisian factory, and blessed by a president who mentions it in the same breath as Airbus. I want to believe that story. But belief is precisely the commodity this industry overproduces, and from the chaos of 2017, we forged a compass that points away from enthusiasm and toward evidence.
The merits deserve to be stated first, because they are real and they are technical. Mistral's founding trio — Arthur Mensch out of DeepMind, Guillaume Lample and Timothée Lacroix out of Meta's AI research group — shipped Mistral 7B in September 2023, a model that embarrassed several competitors twice its size. Mixtral 8x7B followed in December, applying a sparse mixture-of-experts architecture that activated only a fraction of parameters per token, delivering frontier-adjacent quality at a fraction of the inference cost. That was not marketing; that was engineering, and it is why the developer community adopted the weights so quickly. Then came the deliberate two-tier turn: open weights for the small and mid-range models as a distribution strategy, closed weights for the frontier tier — Mistral Large, Codestral, Pixtral — sold through APIs and enterprise contracts. It is a sophisticated architecture of leverage, and I say that with genuine respect.
The funding ladder tells you what the market thought of it. A €385 million round in December 2023. Six hundred million more in June 2024 at a €5.8 billion valuation, with Samsung, IBM, Nvidia, and the American venture funds crowding in. Then the ASML round, €1.7 billion led by a €1.3 billion ticket from the Dutch lithography monopoly, at €11.7 billion. Alongside the capital came the infrastructure: Mistral Compute with Nvidia, anchored by a cluster of eighteen thousand GB200 systems in Essonne, south of Paris, a facility whose power draw alone would have made it a national asset in any decade prior. And alongside the infrastructure came the state: defence procurement through the DGA, deployments across the French administration, a partnership with Agence France-Presse, and Le Chat positioned, explicitly, as the domestic alternative to the American assistants.
None of this is scandalous on its own. States have always backed strategic industries, and Airbus is the proof that the model can produce something that outlives its architects. What makes the current moment worth examining is the political clock hanging over it. Emmanuel Macron was elected in 2017 and re-elected in 2022; the constitution bars him from a third consecutive term. The 2027 presidential election will be the first in a decade in which his name is not on the ballot, and the AI portfolio he has personally curated — the Paris AI Action Summit of February 2025, the "third way" rhetoric between American hyperscalers and Chinese state models, France's lobbying inside Brussels for a lighter-touch implementation of the EU AI Act — is a portfolio of personal diplomacy as much as national policy. Procurement preferences, regulatory shelter, and international promotion all flow from an office that will change hands. I have written before that true ownership is non-negotiable, and the same discipline applies here: a company whose moat is a phone number in the Élysée does not have a moat. It has a landlord.
So let us do what nobody in the press cycle did, and take the word apart. Sovereignty in artificial intelligence is not one thing; it is four, stacked like geological strata, and Europe's position in each of them is radically different.
The first layer is silicon, and here Europe is genuinely strong — ASML's monopoly on extreme ultraviolet lithography is the single most concentrated chokepoint in the global technology stack, and the European Chips Act has begun to draw fabrication back onto the continent. The second layer is weights: who owns the parameters, and who may copy them. Mistral's open releases give Europe a partial claim here, but the frontier tier remains closed, and the most capable European model is still a proprietary asset held by a private company with a Dutch lithography firm as its largest shareholder. The third layer is data, and it is contested to the point of litigation, with the AI Act's transparency obligations colliding with copyright holders across the continent. The fourth layer is inference verification — the ability to prove that a given output genuinely came from the model and configuration it claims to have come from. That fourth layer is empty, and it is the layer that determines whether any of the other three can be trusted.
Why does the empty layer matter more than the full ones? Because an AI system is not a product you buy; it is a process you delegate to. When I founded The Trustless Circle in 2020 — a Discord community that grew to ten thousand non-technical members — I manually verified more than two hundred protocols against open-source standards and built a Trust Score dashboard that cut our community's incident rate by roughly eighty percent. The lesson from that work was not that users need to understand smart contracts. They do not, and they never will. The lesson was that users need the ability to check, or to check someone who checked — and that the moment verification becomes a matter of faith, the entire structure above it becomes a rumour with a flag pinned to it. A model you cannot inspect is a probability distribution with a press office.
So what does real verification look like, honestly assessed? Start with the cheapest mechanism. A model provider can publish a cryptographic commitment to its weight manifest — a hash, signed, timestamped, immutable. This is nearly free to produce and almost nobody does it properly. It proves which file you received; it says nothing about what the file does, and it collapses entirely if the provider silently swaps the weights behind an unchanged API endpoint.
Next, confidential computing. Nvidia's H100 and Blackwell parts support confidential execution modes that produce hardware-signed attestation quotes, allowing a client to verify that a specific model ran inside an isolated enclave on genuine silicon. This is a real improvement, and it moves the trust anchor from the model provider to the chip vendor — a vendor, I note, whose largest European shareholder is also Mistral's largest European shareholder. Better is not the same as neutral.
Then there is zero-knowledge machine learning, where the proof-generation overhead for anything beyond a small model still runs orders of magnitude above the cost of the inference itself. For settlement of high-value disputes, that trade is rational; for a chat turn, it is hopeless. Optimistic verification — publishing an output, opening a challenge window, and paying staked watchers to dispute — is cheap when nobody disputes, which is precisely its failure mode. And content provenance standards such as C2PA, useful as they are, prove where an image came from, not which weights wrote the sentence beneath it. Those two claims are confused constantly and are not equivalent.
This is the gap my team set out to close with the Human-Centric AI Ledger, an initiative I launched this year with two million dollars in grants from ethical technology funds. The design is deliberately unglamorous: commit the weights manifest, the inference configuration, and the input; produce an attestation binding them to an output; allow independent auditors to reproduce the computation; stake and slash on dispute; and keep the whole record on a public chain so that no single vendor can rewrite history. The architecture is hybrid by necessity — hardware attestation on the hot path for latency, zero-knowledge proofs on the dispute path for finality — because the honest answer to "can we verify AI?" is yes, at a cost, and the interesting question is who is willing to pay it. In the meantime, the agent rails are arriving faster than the proofs: ERC-8004's identity, reputation, and validation registries for autonomous agents, proposed in August 2025, and payment primitives that let machines settle with machines. These are the scaffolding of a verification economy, and most of the capital flowing into them this year is flowing in ahead of any working proof.
There is a cost curve underneath all of this that the bull market prefers not to discuss. Cheap data availability made on-chain attestation economically viable; before blob space, writing a provenance record was an indulgence. But blob demand is finite and the supply of attestations is not, and my expectation — unchanged since the Dencun upgrade — is that saturation arrives within two years and the price of writing a proof of provenance roughly doubles when it does. Build your unit economics for that world, not this one.
And then, finally, the political layer, which brings us back to the wire from Veldhoven. A private AI champion has no continuous price discovery on its political risk. Its valuation updates quarterly, in private, when a fund decides what a state contract is worth. Token markets, for all their pathologies, quote their judgments every second of every day. Which means the most honest public signal about the durability of Europe's AI strategy may currently be sitting in the prices of AI-adjacent tokens and decentralized compute networks — instruments that are themselves, I should say plainly, mostly unverifiable. That is an uncomfortable place to find price discovery. It is also the one we have.
Now the contrarian turn, and I will aim it at my own side of the aisle. The comfortable version of this essay ends with the state as villain and the decentralized alternative as remedy. I do not believe it. In this bull market, a project with one hundred million dollars raised and a whitepaper about AI agents is doing precisely what the 2017 ICOs did, in better typography: manufacturing a fragmentation narrative, selling the fix, and shipping nothing that can be checked. When I audit these networks, I ask one question, and it is the same question I would ask Mistral. Show me an output. Show me the weights that produced it. Show me the proof that binds them, and show me who is staked to lose if the proof is false. If the answer is a PDF, a partnership announcement, or a promise that verification is on the roadmap for the next mainnet upgrade, then it is marketing — and the loudest voices in this cycle are solvent because nobody has demanded the proof yet.
What would change my mind? Three things, none of them unreasonable. Verifiable inference at production latency for a cost multiple in the single digits rather than the thousands. Published evaluation harnesses that a third party can reproduce without the model provider's cooperation. And agents carrying attestation histories — real ones, with slashing behind them — rather than reputation scores that can be minted from a fresh wallet on a Tuesday. The same scepticism applies to the projects grafting inference onto settlement layers that were never built to carry it; you do not make a Rolls-Royce more useful by loading cargo onto the bonnet, and you do not strengthen a monetary network by asking it to think. The deepest risk to Europe, though, is not that it picks the wrong champion. It is that it mistakes procurement for capability and regulation for safety, and ends up with a well-funded, well-regulated, unverifiable monoculture — which is a worse outcome for its citizens than a competitive market full of imperfect and inspectable models.
The 2027 election is the stress test, and it will arrive before most of these proofs do. Between now and then, the useful question is not which candidate will protect the national champion, or whether a new administration will redirect a procurement budget, or which minister signs which partnership. The useful question is who can hand a hospital, a defence ministry, or a regulator a verification receipt for a decision that a machine made on their behalf — a receipt that survives the vendor, the government, and the election that replaces it. Trust is not a metric; it is a memory we share. But memory needs a ledger, and a ledger needs verification. When the next president takes the office, will Europe discover that it bought sovereignty — or merely rented it, at eleven billion euros, from a landlord who is already packing?
