The Strait of AI

On the evening of 12 June 2026, at 5:21 p.m. Eastern Time, a single government directive did something the sovereignty debate had only ever discussed in the conditional tense. It switched off a frontier AI model — not in one country, but everywhere at once. The order reached any foreign national, inside or outside the United States, and to comply with something drawn that broadly, the model’s maker had no real option but to disable it for everybody.

The episode will be picked over on its own terms for weeks. Was the security concern real or overblown? Was the response proportionate? Fair questions, and not the ones this essay is about.

The news itself was never the interesting part. What mattered was what it made visible. For one evening an abstraction acquired a timestamp: the “kill switch” that policy people invoke on panels stopped being a hypothesis and became a logged event with a time attached to it. And once you have watched the valve close, you cannot unsee the pipe.

This site is an attempt to map the pipe.

Why this matters at all

It is worth being clear about what travels through these bottlenecks, because “AI” is still too easily filed under productivity tool, chatbot, research toy. That filing is already out of date.

The thing that is actually happening is that AI is turning into the nervous system of economies — not a feature bolted onto existing processes, but the tissue through which decisions move and advantage accumulates. Three registers make the point.

The first is scientific. AlphaFold did not nudge protein-structure prediction along; it folded decades of plausible work into a couple of years and handed the result to every lab on earth. That pattern is now repeating across materials science, drug discovery, climate modelling, pure mathematics. A lab with fluent access to frontier models is not simply faster than one without — it is working on a different and larger set of problems, because it can see further into the space of what might be solved. The lab without that access isn’t lagging. It is, increasingly, spending its effort in the wrong places.

The second is efficiency, and efficiency compounds. A firm that pushes AI deep into its operations — logistics, legal, engineering, what it knows about its customers — does more than trim costs. It starts a loop: better decisions throw off better data, which trains better models, which make better decisions. The gap that opens between such a firm and its competitor is not a one-off saving but a widening structural distance. The same holds between states. A government that can read its regulatory filings, parse its satellite imagery and model its policy choices at machine speed is simply playing a different game from one that cannot, and for a long while the difference is invisible from the outside. Then it isn’t.

The third is the one that unsettles me most, and it is qualitative rather than quantitative. At some point — and a lot of large organisations are passing it right now — AI stops being a tool people reach for and becomes the medium the organisation thinks in. Memory, coordination, strategy, the reflex of responding to events: all of it starts to run through AI infrastructure. Past that point, access to frontier capability is no longer an edge. It is a precondition for functioning.

Which is why the valve is worth taking seriously. It is not a tap on a luxury. It is, more and more, a tap on the central nervous system.

The new oil had its straits

Calling AI “the new oil” has become a reflex — intelligence as the substrate on which competitiveness and power now rest. The line is half right, and the missing half is the half that counts.

Oil never made anyone powerful simply by existing in the ground. It made them powerful because of where it had to flow. The twentieth century turned less on who owned the wells than on who sat astride the passages the tankers were forced through — Hormuz, Malacca, Suez. A strait is not a wall. It is a place where a great deal has to move through very few hands, which gives whoever holds it a quiet, standing kind of leverage that goes unnoticed until the day it is used.

So the question is not whether AI is the new oil. It is: where are its straits, and who is standing in them?

Not all of the stack can be held

Here is the distinction that makes the rest of the map readable, and it is the one most discussions skate past.

Jensen Huang describes modern AI as a five-layer cake — energy, chips, infrastructure, models, applications — each layer sitting on the one below, all the way down to the power station. It is a useful picture. For the purposes of sovereignty, though, those five layers fall into two quite different kinds.

The model layer leaks. Capability diffuses almost as fast as it appears. Weights get distilled, quantised, fine-tuned, occasionally leaked outright; a clever method published on a Monday is running on other people’s hardware by Friday. DeepSeek is the standing proof — boxed in on chips, it routed around the constraint with sheer efficiency and then gave the result away. You can put a valve on a model, but what is inside seeps through the walls. Controls at this layer tend to rebound on the people who impose them: they hurry along the very independent, open alternatives they were meant to delay.

The layers underneath do not leak. You cannot distil a lithography machine. You cannot torrent a fabrication plant, a gigawatt of firm power, or a hyperscale data centre. These things are physical, expensive, slow, and stuck in particular places — built over years and continents, not weekends. This is where the real straits are. This is the part of “the new oil” that genuinely behaves like oil.

The costliest mistake a country can make here is to be afraid of the wrong layer — to pour its anxiety into the capability it could rebuild, while quietly renting the infrastructure it cannot.

What flows through the strait, you can replace. The strait itself, you have to own.

The bottlenecks, one by one

The straits of AI are not a single chokepoint but a stack of them, and one power happens to stand at most. Each deserves its own piece; what follows is just enough to show the shape of the system.

Energy. Huang puts it at the foundation, and he is right. Intelligence produced in real time needs power produced in real time, and the ceiling on how much intelligence a country can generate is, in the end, how much firm, affordable, low-carbon electricity it can park next to a data centre. This is the most physical of the bottlenecks and also the most tractable — the one place where Europe is in better shape than it tends to realise.

Silicon. The oldest and hardest of them. One Dutch company holds a near-total monopoly on the extreme-ultraviolet lithography machines without which no leading-edge chip gets made. One Taiwanese foundry makes essentially all the advanced accelerators the field runs on. Neither fact can be undone by a vote in Brussels; both are the residue of forty years of compounding specialisation. And the geography of it — ASML in Veldhoven, TSMC on an island 180 kilometres off the Chinese coast — is a first-order geopolitical fact in its own right.

Compute. Even with power and chips in hand, someone has to build the factories that turn them into usable intelligence. The four big US hyperscalers plan to spend north of $700 billion on AI infrastructure in 2026 alone. Europe’s entire planned AI investment, public and private, spread over several years, comes to a fraction of one of those years. The shortfall is not only money; it is the specialised trades — the electricians, pipefitters and network engineers — needed to pour that capital into concrete and copper at speed, and those people are scarce everywhere, Huang among many noting it.

Talent. The frontier is built by a few thousand people worldwide who sit at the crossing point of research depth, engineering skill and hands-on infrastructure experience. Europe trains them in respectable numbers and then watches a good share leave. The pull is partly money — and here the gap is not the tidy 30 or 40 per cent it is sometimes described as, but something closer to a separate economy. A senior researcher at a top US lab can earn a total package in the high six or seven figures, the bulk of it equity, against a salary in the low-to-mid hundreds of thousands at a strong European lab. No European institution can credibly match that, and most do not try. But the deeper pull is not the pay at all; it is that the large compute clusters needed to do serious work physically sit where those labs are. Talent follows compute the way compute follows power, and the bottleneck feeds itself — the places with the infrastructure attract the people who go on to build more of it.

Capital. OpenAI is valued near a trillion dollars; Anthropic is not far behind. Mistral, Europe’s strongest lab, raised at something like €20 billion — a real achievement, and a different order of magnitude. The frontier is not a one-time purchase. It demands the ability to keep ploughing revenue back into the next training run, which needs the next tranche of compute, which needs the next round of capital, and the firm that can sustain that loop pulls steadily away from the one that cannot. The same compounding that builds corporate moats builds national ones.

Culture and language. The quietest bottleneck, and the one most likely to be waved past in a debate fixated on hardware. The models the world is starting to think with were trained mostly on English text, by mostly American teams, carrying a particular set of assumptions about knowledge, authority and worth. Peer-reviewed work has found that today’s leading models sit closer, in their values, to English-speaking Protestant societies than to the populations of the hundred-plus other countries tested against them. For a French or Arabic or Swahili speaker, the model is not neutral plumbing; it is plumbing built to someone else’s specification, with someone else’s defaults baked in. This one is answered not by regulation but by building — multilingual data, home-grown research, models calibrated to a culture rather than borrowed from another.

What Europe actually has

Laid out like that, the list reads as a catalogue of European shortfalls. It is not only that. Europe comes to this moment holding some real and durable cards that the breathless US-versus-China coverage tends to skip.

It has research of genuine depth. DeepMind is British. The transformer — the mechanism under more or less every frontier model — owes a great deal to people trained in European universities. INRIA, the Max Planck institutes, ETH Zurich and a long tail of others keep producing the foundational work the whole industry then builds on. That is not soft power; it is the headwaters of the stack.

It has a frontier lab. Mistral, founded in Paris in 2023 by people who had left DeepMind and Meta, is the one European outfit building models that hold their own at the frontier on any honest benchmark. Its open-weight strategy — releasing strong models under permissive licences — is not a runner-up’s consolation. It is a coherent wager that openness builds an ecosystem, the ecosystem builds adoption, and adoption builds the data flywheel that keeps a lab competitive. It also happens to fit European instincts and European law better than the closed approach does.

It has a regulatory framework with reach. The AI Act is genuinely contested — celebrated as the first serious attempt to govern the technology, attacked for loading costs onto the very labs Europe needs to thrive. Both can be true. Worth remembering, though, that GDPR drew exactly the same “anti-competitive overreach” complaints and then quietly became the global default, the bar anyone selling into a world market had to clear wherever they were based. The AI Act could travel the same way. Whether it does is a question of implementation, and that is still being written.

It has a real commitment to open source — longer and more consistent, across the EU institutions, the national research funders and the developer communities thick on the ground in France, Germany and the Nordics, than anywhere else of comparable size. In AI, open weights are the thing that keeps the model layer contestable instead of captured. Europe’s reflex here is correct, even when nobody bothers to phrase it strategically.

And it has the energy. France draws roughly 70 per cent of its electricity from nuclear, exports the surplus across the continent, and is building more. In a world where the binding constraint on AI is firm, cheap, low-carbon power, that is not some legacy asset to be managed gently into retirement. It is a forward advantage of the first rank — the single bottleneck in Huang’s stack where Europe is not standing downstream of someone else.

None of this is grounds for complacency. It is grounds for building from a stronger footing than the alarm of 12 June would suggest — provided the building goes into the durable layers and not the leaky one.

Two things that could move the map

Two developments deserve a closer eye than the daily model-release noise — not because their outcomes are settled, but because either could redraw the chart, and Europe sits unusually well on both.

The first is the turn toward the physical world. Every frontier model today is, underneath, a very accomplished predictor of text. Yann LeCun — Turing laureate, founder of Meta’s FAIR lab, and the most prominent dissenter from the language-model consensus — has argued for years that this is a dead end for anything resembling real understanding of physical reality. His alternative is the world model: a system that learns how the world behaves by engaging with it rather than by reading about it. At the end of 2025 he left Meta to build exactly that, and pointedly chose Paris over the Bay Area to do it. His company, AMI Labs, raised just over a billion dollars in a March 2026 seed round — the largest in European history — with NVIDIA, Toyota Ventures and Bezos Expeditions among the backers. His own account of the geography is the part worth keeping: Silicon Valley, he said, is so hypnotised by generative models that the work has to be done somewhere else, in Paris. Whether world models turn out to be the unlock or run into walls of their own, the bet is serious, the money is real, and its centre of gravity is European.

The second is robotics, and here Europe has an inheritance most discussions overlook. The application layer of AI is not only screens and text; embodied intelligence — machines that act in the physical world — may be where the next decade’s transformation lands hardest. And large-scale car manufacturing is the most demanding form of precise, quality-controlled physical production humanity has ever organised. Stellantis, Renault, Volkswagen, BMW, Mercedes, Bosch: that is exactly the industrial base, the engineering culture and the global supply network that physical AI needs in order to be built and deployed at scale. As the combustion engine fades, the real question for European carmaking is not only how to survive the shift to electric, but whether the manufacturing competence it has banked over a century can be turned toward making and fielding robots. Answer that well and you have not just rescued an old industry; you have stood up a new one on its foundations.

Neither is a sure thing. World models may not arrive on LeCun’s timetable; the car industry may pivot or may stall. But both are genuine asymmetric bets — places where the map simply looks different for Europe than the standard race framing allows, and where the durable-layer logic holds: manufacturing know-how, physical plant and energy are not things anyone can download.

The temptation to confuse the costume for the clothes

The danger for Europe is not that it fails to learn the lesson of 12 June. It is that it learns it in the most comfortable and least useful form.

The comfortable version is procurement. Sign the sovereign-cloud deal. Anoint the national champion. Require the public sector to use the European model. None of that is worthless — preference and procurement do shape markets. But sovereignty bought at the application layer while the infrastructure underneath stays rented is not sovereignty. It is a flag flown over someone else’s pipe, which is precisely the arrangement 12 June laid bare: a continent feeling sovereign at the visible layer while wholly dependent at the durable ones.

You do not become sovereign by choosing whose model to rent. You become sovereign by owning the power, the silicon and the compute that any model has to run on.

The implication is uncomfortable, which is probably why it gets avoided. Real autonomy means spending on things that are slow, costly, unglamorous and useless for a press release this quarter — data centres, power, chip-design capacity even where the fabrication stays abroad for now, and a research environment good enough to keep the people it trains. These are commitments measured in decades, fighting for room in every budget against things that show results before the next election.

The question 12 June actually put on the table was never “which model should we use.” It was: how many decades of dependency are we willing to accept, and at which layers?

A map, not a verdict

Standing at a strait and occasionally testing the gate is what powers do. The urge to keep a hand on a strategic resource is statecraft, not scandal, and nothing here is meant as an indictment of whoever currently holds most of the bottlenecks.

What it is meant to be is a frame — a sharper way of putting the sovereignty question than most of the versions now in circulation. The bottlenecks are real. They are layered. They are not equally durable, and the right response to each is different from the right response to the others, in ways that matter a great deal for where scarce money and scarcer political attention should go.

This is a first sketch of the chart, not the finished thing. The pieces that follow will take the bottlenecks one at a time — the technical reality of each, its geography, the shape of Europe’s exposure, and the most credible ways to ease it.

If you work where these decisions get made, or you have a correction, a figure, or a vantage point that belongs on this map, the address is at the foot of the page. A chart gets better the more people have sailed the water it describes.