← All posts / Industry

Twelve Racks a Month: Inside Bull's Doubled Angers Factory, Europe's Only Supercomputer Plant

Bull, France's state-owned supercomputer maker, has reopened its expanded Angers factory after an €80 million rebuild — doubling output to 12 racks a month, assembling Alice Recoque and LUMI-AI in parallel, and scaling toward 24 racks as Europe races to close its AI compute gap with the US and China.

Twelve Racks a Month: Inside Bull's Doubled Angers Factory, Europe's Only Supercomputer Plant

On October 1, 2026, Bull reopened the doors of an expanded factory in Angers, in western France, and in doing so quietly redrew the industrial map of European AI. The plant — the only factory in Europe dedicated to building supercomputers — has doubled its production capacity from six racks a month to twelve, following an €80 million expansion, with headroom to reach 24 racks a month next year if demand keeps climbing.

The news came via a Reuters exclusive from Angers, where chief executive Emmanuel Le Roux said Bull has received more orders this year than ever before. The context is a continent playing catch-up: the European Union has committed €7 billion ($7.87 billion) over 2021–2027 to a pan-European network of supercomputers, an explicit attempt to close a compute gap with the United States and China that policymakers increasingly describe as a strategic vulnerability.

What actually changed on the factory floor

The doubling is not a projected target — it has already happened. Output went from six racks a month before the expansion to twelve now, and executives say the plant can push to 24 racks a month in 2027 on demand it expects to keep growing. For a factory whose output pace had been a binding constraint on European system deliveries, that is the difference between queuing flagship machines serially and building them in parallel.

The two flagship orders illustrate the point. France’s machine — a supercomputer named after French computer scientist Alice Recoque, ordered for the CEA, the country’s nuclear and alternative energies commission — accounts for 94 racks and is due for delivery in stages from the end of this year at a cost of €554 million to the French government and the EU. It is being assembled in parallel with LUMI-AI, Finland’s €388 million system destined for Kajaani and the biggest contract in Bull’s history. Executives were blunt about the physics of the old layout: the two orders could not have been built alongside each other before the expansion.

Bull’s competitive record explains why the queue exists at all. The company has won 15 of the 18 tenders issued by EuroHPC, the EU body that co-funds Europe’s supercomputing network with member states, competing against the likes of Hewlett Packard Enterprise. It built JUPITER, Europe’s first exascale machine — the fastest class of supercomputer in the world — which is now running at the Jülich Supercomputing Centre in Germany. And the demand is no longer purely public-sector: Airbus inaugurated Bull-built supercomputers in Toulouse and Hamburg this year under a five-year deal that media reports had pegged at close to €100 million and which tripled the planemaker’s simulation capacity (Bull told Reuters that figure was not correct but declined to disclose the real one). On the AI side, cloud operator OVHcloud and European Commission-backed Domyn are both attempting to train frontier-class models on JUPITER, alongside other companies Bull declined to name.

Sovereignty, from 1980s embargoes to a Foxconn partnership

The factory’s expansion is the industrial endpoint of a story that began with nuclear deterrence. France completed its purchase of Bull from debt-laden IT group Atos on March 31, 2026, valuing the company at up to €404 million. CEO Emmanuel Le Roux told Reuters the original motive was strategic: in the 1980s, after the United States stopped shipping supercomputers to the CEA, France turned to Bull to build the machines used to simulate nuclear tests. The lesson — that access to top-tier compute can be revoked at a geopolitical whim — is now the operating assumption behind Europe’s entire AI infrastructure push.

That assumption shows up in the bill of materials. The Angers factory assembles systems using Nvidia, AMD, and Intel processors, or European parts, depending on customer demands — but the European share of components in the machines being built has climbed to 70%, up from 20–30% five years ago. Interconnects, cooling, and integration increasingly come from the continent even when the accelerator silicon does not.

The globalization layer has not disappeared, though; it has been re-routed through Europe. Under a partnership signed in June 2026, Bull and Taiwan’s Foxconn will manufacture Nvidia’s NVL systems at a plant in the Czech Republic, with final assembly taking place in Angers. The arrangement lets Bull participate in the Nvidia-centric buildouts that dominate commercial AI demand while keeping the final integration — and the industrial know-how that comes with it — on French soil.

Why a rack factory matters in the AI race

It is tempting to read “12 racks a month” as a small number next to the gigawatt-scale datacenter campuses that American hyperscalers announce. It is not the same race. Europe’s bottleneck is not raw floor space but the ability to design, integrate, cool, and deliver tightly coupled supercomputing systems — the kind that top benchmark lists, train sovereign models, and simulate everything from aircraft to warheads. BullSequana architectures with warm-water direct liquid cooling, custom interconnects, and system-level integration are precisely the craft that cannot be imported overnight, and Angers is where that craft lives.

The doubling also changes the tempo of Europe’s AI Factories program. EuroHPC has been signing contracts for next-generation AI-optimised systems at a steady clip — LUMI-AI was the sixth such procurement — and each machine needs racks physically assembled, tested, and shipped. A factory running at double capacity shortens the lag between a signed contract and delivered compute, which is the lag that determines whether European researchers and companies train on domestic machines or rent abroad.

There are open questions. Whether demand actually materialises at the 24-rack level depends on EuroHPC’s next funding cycle and on national programs holding their budgets. The €7 billion committed through 2027 buys a network, but not parity with US private infrastructure spending, and executives acknowledge demand “it expects to keep growing” rather than guarantees it. And the 70% European component share still leaves the most valuable piece — the GPU — in American hands in most configurations.

But as a signal of direction, the Angers expansion is unambiguous. Europe has decided that supercomputer manufacturing capacity is strategic industrial territory, that it is willing to nationalise a champion to secure it, and that it will scale the physical plant — €80 million at a time — to keep pace with its own ambitions. Twelve racks a month is the new baseline. The continent’s AI compute plans now have a factory floor to match.