There is a particular kind of strategic clarity that only arrives once you stop pretending you can win the race everyone else is running. Europe is behind in large language models, and no honest observer thinks a European lab is about to out-train OpenAI or the Chinese state on raw scale — the capital is not there and will not be. The reflex is to read this as decline. The more interesting reading is that being unable to win a race forces a better question: is this the only race, and is its finish line even in the right place? Europe’s most thoughtful AI bets are three different answers to that question — sovereignty, architecture, and the choke point — and each is stronger than the gloom suggests, and weaker than the boosters claim.
Current Conditions
Move One: Sovereignty as Architecture
The first move reframes the goal from capability to control. The question stops being “whose model scores highest” and becomes “whose model can you actually depend on — legally, physically, politically — when it matters.” A model whose weights, training data and compute sit under European jurisdiction is one that no foreign government can switch off by directive, no rival can revoke access to in a dispute, and no opaque terms-of-service can quietly repurpose. Mistral’s open-weight strategy embodies this: a model you can download, run on your own hardware, and audit is sovereign in a way an API call to a foreign cloud never is. Aleph Alpha’s focus on government and regulated industry is the same instinct — explainability and data control as the product, not raw benchmark scores.
This is a real advantage, and readers of this letter will recognise why it resonates here: it is the same principle as holding your own keys rather than trusting a custodian. But honesty requires naming its limit. Sovereignty over the model is thin if the hardware it runs on, and in some cases the very weights, originate elsewhere. You do not fully control a system whose silicon you cannot manufacture and whose architecture you license. Sovereignty is a genuine strategic axis — trust, jurisdiction, the right to switch nothing off but your own decisions — but on its own it is a legal and political moat, not a technical one. It buys independence of governance, not independence of capability.
Move Two: Changing the Road
The second move is the one that should interest anyone who suspects the current paradigm is not the final one — and it is the boldest thing happening in European AI. In early 2026 Yann LeCun, one of the architects of modern deep learning and a Turing Award winner, left Meta and founded AMI Labs in Paris. It raised just over a billion dollars at a 3.5-billion valuation — the largest seed round in European history — on a thesis that is openly heretical: that large language models are an evolutionary dead end on the road to real intelligence, and that the future belongs to world models.
The technical heart is an architecture LeCun proposed in 2022, JEPA — Joint Embedding Predictive Architecture. Where an LLM predicts the next token in a sequence, learning the surface statistics of language, JEPA learns in an abstract representation space — it tries to capture how the world behaves, not how its description reads. LeCun’s argument, compressed: the world is unpredictable, a system that tries to predict every detail of the future will fail, and you cannot reach human-level intelligence by scaling up a machine that fundamentally does not understand the reality it operates in. If that is right, then the enormous LLM lead held by the US and China is a lead down a road that does not arrive — and the rational move for a latecomer is not to sprint down the same road, but to start walking a different one early.
You cannot reach human-level intelligence by scaling up a system that fundamentally does not understand the world. — Yann LeCun, founder, AMI Labs
This is the through-line of everything this letter cares about: the suspicion that the dominant instance of a thing is not the same as the category itself, and that a paradigm can be both wildly successful and ultimately a local maximum. A world-model bet is a wager that the destination lies off the current map entirely. But here the discipline must be absolute, because this is exactly the kind of claim that flatters a latecomer too easily. The architecture bet is unproven — emphatically so. AMI’s own CEO concedes it could take years to move from theory to a usable product, that this is fundamental research and not an applied startup with revenue in six months. He even predicts, with a grin, that “world models” will become the next funding buzzword — every company slapping the label on to raise money. World models might be the next paradigm. They might also be a brilliant idea that does not cash out for a decade, or ever. “We changed the track” is genuine strategy when the new track demonstrably leads somewhere; it is a comforting story when it merely lets you stop competing on the track you were losing. Today, which one this is remains honestly unknown.
One more caveat the boosters skip: AMI is only half a European story. It is headquartered in Paris — a real coup for the French AI scene, personally blessed by the president — but LeCun himself remains based in New York, with offices in Montreal, Singapore and Zürich, and its backers include Nvidia, Bezos, Samsung. The head is transatlantic; the flag is French. That is still a meaningful win for Europe, but it is not the clean sovereign triumph the headline implies.
Move Three: The Choke Point
The third move is the one that needs no faith at all, because it already works — and it is the strongest card Europe holds. ASML, a Dutch company, is the sole maker of the extreme-ultraviolet lithography machines required to fabricate the most advanced chips in the world. Not the best maker — the only one. Every leading-edge processor that trains and runs every frontier model, American or Chinese, is etched by a machine from one company in the Netherlands. It is the textbook choke point: the single irreplaceable node in a global supply chain, and therefore the single most powerful position in the entire AI stack. Europe’s strongest play in artificial intelligence is not a model at all. It is the one machine the whole field, adversaries included, cannot build without. (I took that machine apart on its own terms in The One Machine. And Who Builds It. — the bottleneck inside the AI bottleneck.)
And there is a quiet verification of the strategy’s coherence: ASML has become the largest shareholder in Mistral, investing well over a billion to do it. The hardware choke point and the sovereign-model hope are deliberately interlocking — Europe’s one unarguable advantage in the stack reaching down to anchor its model-layer hope. On paper, that is exactly the kind of vertical coherence the strategy needs.
The Catch in the Choke Point
That last point deserves the issue’s sharpest light, because it is where the strongest card reveals its hidden weakness. A choke point is power only as long as you control the valve. ASML cannot sell its most advanced machines freely to China — not because the Netherlands decided so alone, but because US export rules reach across and govern what the Dutch company may export. The single most powerful position in the global AI stack is already operated, in part, under another government’s veto. Hold the bottleneck and you are simultaneously powerful and hostage: indispensable to everyone, and not fully free to act.
This is the same lesson that recurs whenever a critical capability concentrates in one place — the holder of the choke point becomes the prize everyone else negotiates over, and discovers their leverage is also a leash. It is exactly what happened when the dollar’s settlement monopoly was turned into a sanctions weapon: the very act of wielding the choke point taught everyone else to route around it, as I argued in When the Dollar Becomes a Weapon. Sovereignty and choke-point dominance, the issue’s first and third moves, turn out to constrain each other: the more irreplaceable your node, the more others will insist on a say in how you use it.
None of which makes the strategy wrong. It makes it real — a genuine set of bets with genuine failure conditions, which is worth far more than a triumphal narrative. Europe has chosen, sensibly, not to lose a race it could not win, and to compete instead on control, on a contrarian architecture, and on an irreplaceable machine. Whether that is foresight or sophisticated consolation depends on outcomes not yet in: whether sovereignty deepens into capability, whether the other road arrives anywhere, and whether the choke point stays a sword rather than becoming a leash. The strategy is sound. The result is unwritten. Both of those are true at once, and saying so is the whole point.
What to Actually Take From This
The European AI debate is usually conducted in two useless registers — reflexive declinism and PR triumphalism — and the truth is a structured set of bets that deserve to be judged individually, on their merits and their failure modes.
Changing the track can beat running faster — if the track leads somewhere. The smartest response to an unwinnable scaling race is to question whether the race ends where everyone assumes. The world-model bet is exactly that move, made by someone with the credibility to mean it. But the bet is unproven by its founders’ own admission, and “we chose a different paradigm” is strategy only if the paradigm delivers — otherwise it is the most sophisticated form of cope there is.
Sovereignty is control, not capability — do not confuse them. Owning the jurisdiction your model runs in is a real and underrated advantage, the same logic as holding your own keys. But it is a legal and political moat, thin while the silicon and sometimes the weights come from elsewhere. Sovereignty buys independence of governance. It does not, by itself, buy independence of capability, and treating the first as the second is the strategy’s most seductive error.
A choke point is a sword that is also a leash. ASML is Europe’s strongest card and the single most irreplaceable node in the AI stack — and it already operates under a foreign government’s export veto. The lesson generalises: whoever holds the indispensable bottleneck becomes powerful and constrained in the same motion, the prize everyone negotiates over. Real leverage, real leash. Knowing it is both is what separates a strategist from a cheerleader.
Instrument Check — Worth Your Attention
Study — LeCun on world models and JEPA, the primary-source interviews (2026). Go to the source for the architecture bet rather than the hype around it. LeCun’s own framing — why next-token prediction is, in his view, a dead end, and what JEPA is meant to do instead — is the clearest statement of the contrarian thesis. Read it sceptically: it is the most credible case against the current paradigm and also a billion-dollar pitch from the man who needs it to be true.
Read — Chip War, Chris Miller. The definitive account of why semiconductor manufacturing is the real terrain of technological power, and why a single Dutch company’s lithography machines became the most strategically contested hardware on the planet. Essential for understanding why Europe’s strongest AI card sits in fabrication, not models — and why that card is played on a board others partly control.
Follow — the sovereignty thread: When the Dollar Becomes a Weapon. This issue’s logic — control over capability, holding the node others depend on — is the same structure that runs through this letter’s work on sound money and self-custody. The principle is constant across domains: who can switch your system off is a more important question than how fast it runs.
Flight Log — Dispatch From Altitude
There is an old truth in aviation that every pilot learns and every airline lives by: you do not win by having the fastest aircraft. For decades the instinct was that speed was the prize — and the supreme expression of that instinct, Concorde, was a technical marvel that crossed the Atlantic in half the time and lost money for its entire commercial life. The race for raw speed was the obvious race, and it turned out to be the wrong one. The carriers that endured did not win on Mach number. They won on the things underneath it — networks, efficiency, the control of the slots and hubs everyone else had to route through.
That last point is the one worth sitting with, because it is the choke point in disguise. The most valuable asset in commercial aviation is often not an aircraft at all — it is a slot portfolio at a saturated airport like Heathrow, or a landing right into a market no one else can serve. You can own the finest fleet in the world, and if you cannot get a take-off slot at the airport everyone needs, your speed is worthless. The airlines that understood this competed not on the glamorous axis of velocity but on the unglamorous, decisive axis of access — the bottleneck that the whole network has to pass through. Hold the slot and the fast aircraft must come to you.
But every slot-holder knows the second half of the lesson too, and it is the leash inside the sword. A slot is granted, regulated, and revocable. It is power held at someone else’s pleasure — the airport authority’s, the regulator’s, the bilateral treaty’s. The carrier that controls the indispensable access point is powerful precisely to the degree that others depend on it, and constrained for exactly the same reason: the more critical your slot, the more everyone else insists on rules about how you use it. Indispensability and freedom pull against each other. The choke point is real leverage and a real leash, in the same breath — on the ramp as in the chip supply chain.
So the lesson the cockpit offers Europe’s AI strategy is one it already knows in its own industry. Do not race for the fastest aircraft when you will lose on budget — compete on the network, the efficiency, the access others cannot replicate. Bet, where you can, on the route the rest of the field has not yet realised it should fly. And hold the choke point with both eyes open: knowing it is the strongest position on the board, and knowing that the strongest position is also the one everyone else is most determined to put rules around. The aircraft is not the prize. The slot is. And the slot is never quite as freely yours as it looks from the flight deck.