Build5 publishers3 min readPublished Updated
Bull doubles its Angers rack output to 12 a month to build two supercomputers in parallel
Bull spent 80 million euros doubling its Angers factory's output from six supercomputer racks a month to 12. The larger line lets it build France's 94-rack Alice Recoque and Finland's LUMI-AI at once, something executives said the old factory could not do.
The Engineer · Build desk
What happened
- Alice Recoque, ordered for France's nuclear research agency, is due in stages from the end of 2026 at a reported cost of 554 million euros to the French government and the EU.
- France closed its purchase of Bull's advanced computing business from Atos on March 31, at an enterprise value of up to 404 million euros including contingent payments, according to Atos.
- Reuters describes Angers as Europe's only factory dedicated to building supercomputers.
- Le Roux told Reuters that Bull won 15 of 18 EuroHPC tenders, a win rate that rests on his own account and has not been independently audited.
- Airbus switched on Bull-built supercomputers in Toulouse and Hamburg this year under a five-year deal worth close to 100 million euros, tripling its simulation capacity.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Commercial buyers share one integration line with two national flagship builds, so a shorter wait for them depends on Bull actually reaching 24 racks a month next year.
- exposure Signing with Bull now means signing with a company the French state bought for strategic reasons, whose nuclear research agency has its own order on the same floor.
- decision Tenders that require European content can point to 70% European-made components in current builds, but the figure counts component origin, so buyers still have to name which parts must be European.
I take the rack count Bull doubled as a measure of integration capacity. The Angers line takes processors from Nvidia, AMD or Intel, or European components, depending on what each customer specifies [14]. Bull's own overview of the site describes teams adapting systems to customer requirements and validating configurations before delivery [16]. The June agreement with Foxconn keeps the same division of labour. Nvidia NVL systems are to be manufactured at a plant in the Czech Republic, with final assembly in Angers [17].
At 12 racks a month the line can produce 144 racks a year, against 72 before [1]. Alice Recoque is 94 racks [4]. That one system is about 65% of a year's output at the new rate [2]. On the old line it would have needed nearly 16 months of everything Angers could build. On the new one it needs just under eight [3]. These ratios assume every rack takes the line roughly the same time to build. Rack count is the only throughput unit in the reporting.
Finland's LUMI-AI is being assembled alongside it and, at 388 million euros, is Bull's largest contract to date, according to Reuters [4]. The reporting does not include LUMI-AI's rack count or any lead time Bull quotes to commercial buyers. For a team buying an HPC or AI system in Europe, a shorter wait needs capacity left over after those two builds. Executives told Reuters the line has room to reach 24 racks a month next year if demand continues [3]. At that rate Alice Recoque would be about a third of a year's output [4]. Chief executive Emmanuel Le Roux told Reuters that Bull received more orders in 2026 than in any previous year [20] [12].
The state's interest in owning a factory is older than the AI build-out. Reuters reported that France pursued Bull's advanced computing business for strategic reasons. Le Roux pointed to the 1980s, when the United States halted supercomputer shipments to France's nuclear agency [8]. Much of the money behind current orders is public. The EU committed 7 billion euros to its supercomputer network over 2021 to 2027 [9].
In my view a parallel line is the right place for the 80 million euros [2]. Bull's largest orders are national machines like Alice Recoque and LUMI-AI [4]. A factory that can integrate only one of them at a time turns each tender win into a wait for the next customer, and executives said that was the position before the expansion [5]. Bull has delivered at this scale before. It built JUPITER, which Reuters described as Europe's first exascale supercomputer, now operating at Germany's Juelich research center [18]. OVHcloud and Domyn are attempting to train frontier-class AI models on it [19].
What to watch
- Whether Bull commits to the step from 12 to 24 racks a month next year, which executives tied to demand continuing.
- LUMI-AI's rack count and delivery schedule, which set how much of the new capacity is left for other customers.
- Whether the first staged Alice Recoque deliveries arrive from the end of 2026 as planned.