Bull has been selected to deliver the €387.8 million LUMI-AI supercomputer in Finland, providing the principal computing platform for a European AI Factory intended to serve industry, research organisations, start-ups, and public-sector users.
The EuroHPC Joint Undertaking has signed the procurement contract for the system, which will be installed at CSC – IT Center for Science’s data centre in Kajaani. Delivery, installation, and maintenance are included in the overall contract, with the machine expected to become available during 2027.
LUMI-AI is designed to provide around ten times the artificial-intelligence capacity of the current LUMI supercomputer and almost twice its conventional high-performance computing capability. The increase reflects the growing difference between architectures optimised for AI workloads and the floating-point performance traditionally used to characterise scientific supercomputers.
The machine will be built around Bull’s liquid-cooled BullSequana XH3500 architecture and will use AMD Instinct MI430X GPUs alongside sixth-generation AMD EPYC processors with 256 CPU cores. Bull will also supply its BXI high-performance interconnect, while IBM will provide storage based on Storage Scale technology and Nokia will contribute data-centre networking.
Those components have to operate as a balanced system. Modern AI installations can contain huge numbers of accelerator cores, but their usefulness depends on moving data quickly enough between processors, memory, storage, and network interfaces to keep those accelerators occupied rather than waiting for input.
The same problem applies to conventional simulation workloads. Computational fluid dynamics, materials modelling, weather models, structural analysis, and other scientific applications can require large numbers of processors to exchange intermediate data repeatedly, which means interconnect performance and latency can become as important as the nominal speed of the individual chips.
Cooling is another increasingly dominant part of the design. As accelerator power density has risen, removing heat through room-level air systems alone has become progressively less practical for the highest-density machines. LUMI-AI will use liquid cooling integrated into the Bull platform, moving heat away from processors closer to the source.
That allows computing hardware to be installed more densely, but it also turns cooling into a direct part of the computing infrastructure. Pumps, manifolds, heat exchangers, water quality, leak detection, controls, and maintenance procedures all have to operate reliably around some of the most expensive electronic equipment in the facility.
CSC says LUMI-AI will act as the computing backbone of the LUMI AI Factory and will support AI training, inference, large scientific simulations, and data-intensive applications. The system is intended to handle large, dynamic, and sometimes confidential datasets, making multi-user security and resource allocation important alongside raw performance.
Manufacturing is one of the programme’s stated priority areas. Since April 2025, the existing LUMI AI Factory has provided AI computing resources and support to European small and medium-sized businesses and start-ups, with manufacturing, health and life sciences, and communications technologies among its strategic sectors.
For industrial users, access to a shared machine at this scale can remove the need to justify dedicated infrastructure for workloads that exceed conventional enterprise systems. Large engineering simulations, digital twins, specialist AI models, process optimisation, computer vision, and materials research can all demand bursts of computing capacity that would otherwise require substantial capital investment.
Shared infrastructure creates its own operational questions. Commercial users need predictable access, protection for proprietary data, software environments suited to different engineering tools, and sufficient technical support to move workloads onto an unfamiliar supercomputing platform. The value of the hardware therefore depends heavily on how easily organisations can use it.
The consortium behind LUMI-AI comprises Finland, Czechia, Denmark, Estonia, Norway, and Poland. EuroHPC will fund half of the €387.8 million total through the Digital Europe Programme, while the participating consortium will provide the remaining 50%.
LUMI-AI will also sit alongside the planned LUMI-IQ quantum computer, creating an experimental environment in which quantum systems can be linked with conventional HPC and AI resources. Practical industrial quantum computing remains far less mature, but colocating the technologies allows researchers to test hybrid workflows without pretending that one architecture is ready to displace the others.
The Kajaani installation is part of a wider European build-out. EuroHPC is overseeing 19 AI Factories and has already signed several procurement contracts for AI-optimised supercomputers, while separate AI Factory Antennas are extending access and support services across more countries.
For Bull, the contract puts its system architecture, cooling technology, and interconnect hardware at the centre of one of Europe’s largest current computing procurements. The supplier still depends on an international component ecosystem for processors, storage, and networking, but the integration responsibility sits with the system vendor.
That responsibility becomes increasingly significant as computing installations resemble industrial plants rather than racks of general-purpose servers. Electrical capacity, cooling-water systems, switchgear, backup infrastructure, controls, cybersecurity, storage, and network design all have to scale alongside the processors.
Installation during 2027 will provide the first practical test of the design. The headline specifications are substantial, but users will judge the machine by available capacity, job throughput, software performance, reliability, and the amount of useful work completed per unit of electrical power. Expensive silicon is easy to count; keeping it productively occupied is the harder engineering problem.
