EU Orders AI Supercomputer from Bull as Demand Overwhelms Capacity

PARIS: The European Union has awarded French state-owned Bull a €387.8 million contract to build a new AI-focused supercomputer, a milestone order that arrives as EuroHPC officials say demand for European AI computing power now exceeds available capacity. The machine, called LUMI-AI, will be installed alongside the existing LUMI supercomputer in Finland and is expected to become operational in the second half of 2027.

The contract is the largest ever received by Bull, which was carved out of debt-laden Atos earlier this year. It marks the sixth system ordered under EuroHPC's AI Factories programme, a network backed by an €8.2 billion budget through 2027 that has established 19 AI Factories around 12 supercomputers across the bloc. "Right now, we have to reject some applications because we cannot satisfy all these demands," EuroHPC infrastructure head Evangelos Floros told reporters, adding that capacity constraints should begin easing as additional systems come online from late this year.

The hardware choices are deliberate. Bull will integrate AI chips from AMD, Nvidia's main competitor, alongside storage from IBM and networking technology from Nokia. The configuration represents an effort to build European AI infrastructure that is not wholly dependent on a single American chip supplier, even as U.S. components remain central to the system.

EuroHPC infrastructure is already being used by European AI developers. OVHcloud has told Reuters it completed initial training of a foundation model on Germany's Jupiter supercomputer, while Italian startup Domyn, which leads the Commission-backed EUROPA consortium, is using EuroHPC capacity to train its own model. The LUMI-AI order extends that foundation, giving European startups and research institutions another large-scale training option as they compete with American and Chinese rivals.

BuiltWorld AI Operational Take: The LUMI-AI order is a test case for whether Europe can build frontier-class AI infrastructure with a multi-vendor hardware stack rather than a single-supplier architecture. For project teams, the operational challenge is integration. Combining AMD silicon, IBM storage, and Nokia networking into a cohesive supercomputing fabric takes longer to commission and tune than a homogeneous Nvidia build. The 2027 launch date leaves margin for that work, but any slippage in component delivery or integration testing will push Europe's capacity relief further out at a time when its own officials admit demand is already outstripping supply.