Europe Doesn't Have a Model Problem. It Has an Engine Problem.

Europe's AI debate keeps returning to models, and keeps reaching for the same answer: build our own, fund a champion, catch up at the frontier. I read it the opposite way. Europe does not have a model problem. It has an engine problem, and more European models paid for with more European public money is aimed at the wrong thing.

The dependence is not theoretical, and it has already been demonstrated. In June 2026, a US export control directive barred foreign nationals from Anthropic's newly released Claude Fable 5 and Mythos 5. To comply, Anthropic pulled both models for every customer in the world. They had been live for days. The controls were lifted at the end of that month and access was restored, which makes the demonstration no less complete: a directive from one government removed a capability from every customer on earth, and nobody outside that government had a say in it.

Across my Chief AI Officer network the reflex was immediate: Europe needs its own frontier models, and fast. That reflex is wrong, and it is wrong for the same reason we keep falling behind in the first place.

We never had a talent problem

Start where the panic starts, with talent, because it is the assumption underneath everything else. Europe does not lack the ability to build good models. ETH Zurich, EPFL and Switzerland's national supercomputing center released Apertus, a fully open and capable model, weights and training data included. Mistral builds strong models in Paris. Britain grew DeepMind, one of the best AI labs on earth, in London, and then watched Google buy it because no European company saw what was coming.

That is the tell. The talent was never missing. The buyer was.

So when people say European models trail because they are simply not as good, they have the causality backwards. Quality is not the input. It is the output.

Distribution is the cause, not the reward

Frontier models do not get good in a laboratory and then go looking for users. They get good because of their users.

Distribution is not the prize for building a great model. It is how a great model gets built. Put a system in front of hundreds of millions of people and every interaction becomes signal: every query, every correction, every thumbs down, every jailbreak attempt sharpens the next version. A model with a hundred million users learns at a rate a model in a quiet lab cannot touch.

There are two ways to win those users, and Europe missed both. You can own a platform, the way Microsoft and Google push models through Windows, Android, search and the cloud. Or you can invent a category and seize it, the way OpenAI took ChatGPT from nothing to a hundred million users in two months, the fastest-growing consumer product the world had seen at the time, with no inherited pipe and the nerve to ship. Either route buys the usage that forges quality. We had the labs, never built the platform, and never made the category-defining move while the window was open.

I am not the first to say this. Venkat Venkatraman has argued for years that digital competition is won not by the best technology but by whoever orchestrates the ecosystem and controls how it reaches the market. His work, from The Digital Matrix to Fusion Strategy, keeps returning to the same point: a capability with no ecosystem to carry it is inert. That is Europe's position exactly. We keep building the capability and conceding the ecosystem.

A European model with no distribution never enters that loop. It is not behind because our scientists are weaker. It is behind because it never gets stress-tested at the scale that makes a model great.

Why the talent leaves

Capital is the other half of the engine, and it does two jobs. It buys the compute the loop runs on, and it pays the people who run it.

This is where the talent question finally resolves. We treat Europe losing its best AI minds as a riddle of culture or ambition. It is neither. Talent follows the flywheel. Where there is capital and distribution there is revenue, and where there is revenue there are salaries and equity a European lab cannot match, plus the most interesting problems in the field. The brain drain is not the disease. It is a symptom of an engine that spins somewhere else.

The sequence runs one way. Capital funds the race, distribution wins the share, share makes the revenue, revenue pays the people, the people improve the models. Pull out capital or distribution and the wheel never turns, however good the model you start with.

The number that should end the debate

Here is the figure I keep returning to. In the first quarter of 2026, four American firms spent roughly 130 billion dollars on AI infrastructure. Across that same year the five largest were guiding toward 660 to 690 billion. Europe's flagship answer, the InvestAI plan announced in February 2025, is 200 billion euros spread over years. They spend in a single quarter most of what Europe has planned in total.

The reflexive European conclusion is that the EU must therefore write the big check and match the spend. The same numbers say that is hopeless. If American capex runs toward 700 billion a year, no European state budget will ever match it. The InvestAI envelope is not too small. It is the wrong instrument. You cannot out-subsidize a market with a ministry.

Now the honest distinction, because this is where the argument is won or lost. There is one layer where public money genuinely belongs: compute. Sovereign infrastructure, the EuroHPC supercomputers, the Alps machine in Lugano that trained Apertus, is exactly the kind of long-horizon bet states are built for, and Europe should do far more of it. But the labs and the products that run on top are a different layer, and there the state has never been the funder. In the United States the government did not fund OpenAI or Anthropic. Microsoft put well over ten billion into OpenAI. Amazon and Google put billions into Anthropic. Venture did the rest. Washington's role was structural: the deepest capital markets in the world, a system that rewards risk, and industrial policy aimed at chips, not at the labs.

So Europe does not have an AI funding problem. It has a private-capital problem and a risk-appetite problem, and it keeps demanding from the state the one thing America never used.

Build the market, not the lab

If the state wants to move the needle on the labs, it cannot do it with a checkbook. It has to build the conditions that let private capital do what it did in America. A real Capital Markets Union in place of twenty-seven fragmented ones. Pension and insurance capital allowed to take the risk it is currently kept from. Exits deep and liquid enough to be worth chasing. Do that, and the money already sitting idle across Europe starts funding labs the way it does in California.

But capital is only the half the state can unlock, and it is worth being honest that a Capital Markets Union does not hand you the other half. Money funds the lab. It does not deliver the hundred million users that make the lab's model great. Europe gets that reach in only two ways, the same two it missed before. It backs a European company with the scale to own a platform and push a model by default, instead of letting that company be acquired the moment it matters. Or it stops trying to win the horizontal model race and seizes the verticals where Europe still has distribution and proprietary data: regulated finance, industrial systems, healthcare, the places a focused European player can reach users and data a general US model cannot. Capital lights the engine. Distribution is the road, and Europe still has to choose which road it will actually build.

This is also the honest answer to the dependence that started the conversation. A sovereign model you cannot fund to the frontier or distribute to real users does not buy you independence. It buys you a flag. The shutdown is a genuine warning, but the cure is not a European model nobody uses. It is a European player with the engine to be a real alternative, and that engine runs on private capital and distribution, not on a subsidy.

I will be fair to the other side. China has been gaining ground with a heavy state model, so public capital is not the only path that works. But it is not the path Europe says it wants. We talk about matching America, then reach for the one instrument America never touched. Pick a lane.

The road, not the car

None of this means Europe should stop building models. It means we should stop believing the model is the bottleneck, and stop believing the state is the engine.

We keep trying to win the race by engineering a faster car. The Americans built the road, with private money, and let a thousand drivers onto it. The car was never the hard part.

Europe does not have a model problem. It has an engine problem. And you do not fix an engine problem with a faster car or a bigger subsidy.

So the real question is not whether Europe can build frontier models. We can, and we have. It is whether Europe is finally willing to build the market that would fund them, and choose the road that would carry them, the Capital Markets Union that has been on the table for over a decade and never delivered. I have my doubts. I would genuinely like to be argued out of them.