Skip to main content

The Europeans

Menu
An interactive briefing · Europe & the AI supply chain

Anatomy of an answer

The models themselves, the chips they run on, and the buildings that house them.

LAYER 01 / 08 · SOFTWARE

The foundation model

Annotated axonometric diagram of a foundation model: what comes out, attention, the layers, the parameters
FIG. 01  ·  A foundation model, schematic.
IN PLAIN TERMSA foundation model is a very large statistical system trained on much of the public internet. It learns one skill — predicting the next word — so thoroughly that answering questions, translating and coding emerge as side effects. Training one from scratch costs hundreds of millions of euros in computing time, which is why so few organisations do it.
WHO CONTROLS IT — notable models released, by origin
United StatesChinaEuropeOthers
0 of ~40
notable AI models of recent years originated in the EU — chiefly from Mistral (Paris) and Aleph Alpha (Heidelberg).

Europe is present at this layer without being at its centre. Mistral reached a €11.7bn valuation in September 2026 — the fastest ascent in French tech history — and its largest shareholder is now ASML, the Dutch lithography maker from layer 05, which put in €1.3bn. DeepMind was founded in London in 2010 and sold to Google in 2014.

What’s being done: InvestAI, announced in February 2025, is the Commission’s attempt to fix the compute constraint. In practice it works like this. The headline €200bn is a mobilisation target, not a cheque: roughly €50bn of EU money — largely repurposed from existing programmes such as Digital Europe, Horizon Europe and InvestEU — is meant to attract about €150bn in private investment. Inside that, €20bn is ring-fenced for four to five AI gigafactories. Despite the name these are not chip plants but very large computing sites, each planned with around 100,000 advanced AI chips, about four times the size of the EU’s current AI factories. Public grants would cover roughly a third of each site, private investors the rest. The aim is that a European startup or university lab can rent frontier-scale computing power instead of that capacity existing only inside the largest American firms.

« Si l’IA européenne se résume à Mistral, alors nous sommes fichus. »
“If European AI comes down to Mistral alone, then we’re done for. We can’t put all our eggs in one basket — it’s a whole ecosystem that has to be built.”
Roland Lescure, French Minister of the Economy, speaking to reporters in San Francisco, 2 September 2026. Asked by AFP whether Mistral should keep developing frontier models or shift towards supplying computing infrastructure, he declined to say — and warned against resting European AI sovereignty on a single company. Reported by AFP; carried by BFM TV, Le Figaro and Maddyness.

Europe has two model labs of real standing, and one of them now carries most of the continent’s expectations on its own. Their constraint is access to computing power and to capital at the scale their American competitors raise it — both of which sit in layers further down this stack.

LAYER 02 / 08 · HARDWARE

AI accelerators

Annotated axonometric diagram of an AI accelerator package: compute die and high-bandwidth memory stacks
FIG. 02  ·  An AI accelerator package: designed in California, fabricated in Taiwan, memory from Korea.
IN PLAIN TERMSAn "AI accelerator" is a chip built to do one thing extremely fast: the matrix multiplication at the heart of neural networks. GPUs were originally designed for video-game graphics — the same maths, it turned out. Training one frontier model ties up tens of thousands of these chips for months; a single one sells for the price of a car.
WHO CONTROLS IT — AI training-chip market
United States (Nvidia, AMD, Google)Others
0%+
of the accelerators used to train frontier models come from one US company: Nvidia.

Europe designs world-class processors — the Arm architecture inside most phones was created in Cambridge — but no European firm builds frontier training chips. The most serious attempt, Bristol-based Graphcore, was acquired in 2024 by SoftBank of Japan. What's being done: the EU-funded SiPearl (France) is building European processors for supercomputers, and the newest machines of EuroHPC (the EU joint undertaking that funds, owns and operates Europe’s public supercomputers) pair those European processors with Nvidia accelerators for the AI workloads. There is no European alternative at the training frontier today.

Every European lab, university and government buys its training hardware from the same US supplier — so Europe's access to compute is set by American export policy, not European policy.

LAYER 03 / 08 · INFRASTRUCTURE

Data centres

Line illustration of a data hall: two rows of server cabinets receding down a cold aisle
FIG. 03  ·  A hyperscale data hall.
IN PLAIN TERMSAs the saying goes, there is no cloud — just someone else’s computer. That is what a data centre is: a warehouse of other people’s computers, where models are trained and where your prompt is actually answered. Whoever owns the building decides whose models run on it, at what price, and under which country’s law.
WHO CONTROLS IT — Europe's own cloud market, by provider origin
US (AWS, Microsoft, Google)European providersOthers
0%+
of Europe’s cloud market is served by three American companies.

The three American firms are Amazon Web Services, Microsoft Azure and Google Cloud. The European alternatives exist but are an order of magnitude smaller: OVHcloud (France, the largest EU-headquartered provider), Scaleway (France, part of Iliad), IONOS and StackIT (Germany, the latter owned by the Schwarz group behind Lidl), Hetzner (Germany), Open Telekom Cloud (Deutsche Telekom), Aruba (Italy) and UpCloud (Finland). Two distinctions matter here. Cloud providers sell computing; data-centre landlords rent floor space, and Europe’s largest halls are largely operated by American groups such as Equinix and Digital Realty even when the building sits in Frankfurt or Amsterdam. And a European flag on the provider does not by itself change which chips are inside — those come from layer 02 regardless.

In Ireland, data centres consumed 23% of all metered electricity in 2025, up from 5% a decade earlier — the highest national share recorded anywhere.

What’s being done: the AI gigafactories described in layer 01 are the main answer, and the scale gap they would close is worth stating in plain numbers. Europe’s largest public AI machine, JUPITER at Jülich, runs on roughly 24,000 accelerators. Each planned gigafactory is specified at around 100,000. The largest American training clusters already operate in the range of several hundred thousand and are still growing. So the programme would move Europe from roughly a twentieth of frontier scale to something nearer a fifth — a genuine improvement that still leaves European labs training on less hardware than their competitors, and it depends on private investors funding about two-thirds of each site.

Most European data, including public-sector data, is processed on infrastructure operated by American companies. The planned gigafactories would narrow the gap in computing capacity without closing it.

Next in the series · Part 2Making the chips

How chips are manufactured, and the one machine in the process that Europe controls.

Coming soon.