TL;DR

The European Union’s supercomputing body has ordered a €387.8m ($449.4m) AI machine from Bull, the French state-owned manufacturer, for installation in Finland by the second half of 2027. It is the largest contract Bull has ever won, months after the French government extracted the business from the troubled Atos for as much as €404m.

What is being built

LUMI-AI is the sixth machine commissioned under the AI Factories programme, and will sit beside the existing LUMI system at Finland’s supercomputing centre. The bill is split down the middle between EuroHPC and a consortium of Finland, Poland, Denmark, Norway, Estonia and the Czech Republic.

The hardware choices matter as much as the price. Compute comes from AMD rather than Nvidia, with IBM supplying storage and Nokia the networking. The machine is intended for training and large-scale inference alongside conventional scientific simulation.

It joins a network that has been assembled steadily rather than announced dramatically: a €8.2bn budget covering 2021 to 2027 has delivered a dozen supercomputers, Germany’s Jupiter among them, anchoring 19 AI factory centres, with as many as seven more sites now being lined up.

The contrast with Britain

This is worth reading against our recent report on Nscale securing a £33bn compute contract with Anthropic. Both are sovereign-capacity stories, but the structures differ fundamentally. Europe is buying publicly funded machines from a state-owned European manufacturer and allocating capacity to a consortium of member governments. Britain’s route runs through private operators serving American labs, on American silicon.

Neither is self-evidently better. The European model is slower and answers to committees; the British one attracts private capital and moves faster. But they answer different questions. EuroHPC is asking who owns the machine. The UK approach mostly asks who builds the building — the argument now running through Scotland, Wales and the Home Counties.

Looking forward

Whether public compute produces anything is still unproven. OVHcloud has finished an initial foundation-model training run on Jupiter, and Domyn, the Italian startup at the head of the EUROPA consortium that the Commission backs, has won access to train a 400-billion-parameter model of its own. Neither has published a model or specifications, leaving Mistral as Europe’s only established frontier contender. Capacity is being procured well ahead of anything that demonstrably uses it.