Europe is spending €387.8 million on a new AI supercomputer in Finland as demand for public computing capacity outstrips supply and the continent tries to narrow its infrastructure gap with the United States and China.
The EuroHPC Joint Undertaking and the six-country LUMI AI Factory consortium have selected French computing company Bull to deliver LUMI-AI. The system will be installed at CSC's new data center in Kajaani, Finland, and is scheduled for deployment in the second half of 2027.
LUMI-AI is expected to provide ten times the artificial-intelligence capacity of the current LUMI supercomputer while nearly doubling its traditional high-performance computing capability. It will use next-generation AMD accelerators and processors, IBM storage and Nokia networking.
The contract is Bull's largest order and covers the machine's acquisition, delivery, installation and maintenance. It also marks the sixth AI-optimised supercomputer ordered under Europe's AI Factories programme.
LUMI-AI supercomputer: key facts
- Contract value: €387.8 million, approximately $449 million at the exchange rate cited by Reuters
- Location: CSC data center, Kajaani, Finland
- Expected deployment: Second half of 2027
- Supplier: Bull
- Owner: EuroHPC Joint Undertaking
- Funding: Split equally between EuroHPC and the LUMI AI Factory consortium
- Consortium countries: Finland, Czechia, Denmark, Estonia, Norway and Poland
- AI performance target: Ten times the current LUMI system's AI capacity
- Traditional HPC target: Nearly twice the current LUMI system's capability
- Architecture: BullSequana XH3500
- Accelerators: AMD Instinct MI430X GPUs
- Processors: Sixth-generation AMD EPYC 256-core CPUs
- Storage: IBM Storage Scale-based system
- Networking: Nokia data-center technology and Bull's BXI interconnect
- Power: 100% renewable electricity
- Cooling: Direct liquid cooling with heat recovery for Kajaani's district-heating network
Why Europe is building LUMI-AI now
The immediate reason is capacity. EuroHPC infrastructure head Evangelos Floros told reporters that some applications for computing resources currently have to be rejected because the network cannot satisfy all demand.
Training and operating modern AI models require large clusters of specialised accelerators. The biggest US technology companies can build private data centers and reserve enormous chip capacity, but startups, universities and smaller companies often cannot afford comparable infrastructure.
EuroHPC's answer is a shared network in which approved European developers, researchers and businesses can access advanced computing resources. The organisation has a budget of €8.2 billion through 2027 and has established 19 AI Factories around 12 supercomputers, according to Reuters.
LUMI-AI is intended to expand that network at a time when access to compute can determine whether an AI project moves from research to a usable product.
What hardware will LUMI-AI use?
LUMI-AI will be built on Bull's new BullSequana XH3500 architecture, a liquid-cooled platform designed for dense AI and high-performance computing workloads.
Its main accelerators will be AMD Instinct MI430X GPUs. These chips will be paired with sixth-generation AMD EPYC processors containing up to 256 CPU cores. AMD is Nvidia's principal challenger in data-center AI accelerators, so the selection gives Europe a major non-Nvidia installation at a time when governments and cloud providers want more competition in the chip market.
IBM will provide high-performance storage based on IBM Storage Scale. Nokia will contribute data-center networking, which will complement Bull's BXI high-bandwidth, low-latency interconnect.
The combination is designed to support several types of work rather than only the training of large language models. Planned workloads include:
- Training and fine-tuning foundation models
- Running AI inference at scale
- Climate and weather modelling
- Health and life-sciences research
- Industrial engineering and manufacturing simulations
- Materials science
- Communications research
- Large scientific calculations requiring high numerical precision
The system will also use a multi-tenant architecture, allowing separate organisations to work on the machine while protecting their data and computing environments.
LUMI-AI versus the current LUMI supercomputer
The existing LUMI machine has been operating in Kajaani since 2021. EuroHPC lists it as a pre-exascale Cray EX system with 386 petaflops of sustained performance.
LUMI has supported thousands of scientific projects and has become one of Europe's most important shared computing platforms. However, it was designed before the current wave of generative AI drove demand for specialised accelerator capacity.
LUMI-AI is not simply a routine replacement. Its claimed tenfold jump in AI capability is much larger than its nearly twofold increase in conventional HPC performance because newer GPU architectures are optimised for the lower-precision calculations widely used in machine learning.
That distinction matters when comparing supercomputers. A machine can show a dramatic increase in AI throughput without delivering the same multiple in the double-precision calculations used for some scientific simulations.
EuroHPC and CSC have not yet published a final measured benchmark for LUMI-AI. The tenfold and twofold figures are performance expectations from the procurement announcement, not independent test results from an operating machine.
Where will LUMI-AI be located?
The system will be installed at CSC's new data center in the Renforsin Ranta business park in Kajaani, a city in central Finland. It will sit within the wider LUMI AI Factory ecosystem and near the current LUMI infrastructure.
Finland offers several advantages for large-scale computing: a cool climate, access to renewable electricity, established research networks and the ability to reuse data-center heat locally.
The consortium is led by Finland and includes the Czech Republic, Denmark, Estonia, Norway and Poland. The participating countries will share responsibility for the project and access to its resources, while EuroHPC will own the machine.
Half of the €387.8 million funding will come from EuroHPC, with the other half supplied by the consortium.
How LUMI-AI will reuse its heat
AI data centers consume large amounts of electricity and convert much of it into heat. LUMI-AI will use Bull's direct liquid-cooling system, circulating warm water close to heat-producing components instead of relying only on energy-intensive air conditioning.
The captured heat will be fed into Kajaani's district-heating network. CSC says the system will operate entirely on renewable electricity and is designed to be among the world's most energy-efficient large AI infrastructures.
Heat reuse does not eliminate the project's environmental impact. Manufacturing processors, constructing the data center and supplying electricity all carry costs. But the design can reduce cooling overhead and replace some heat that would otherwise need to be produced separately for local buildings.
The real efficiency will become clearer after the machine is operational and measured under normal workloads.
LUMI-AI will connect with a quantum computer
CSC plans to complement the supercomputer with LUMI-IQ, a quantum-computing platform supplied by Finnish company IQM Quantum Computers.
Both systems will be located in the same data center and integrated to support hybrid workloads. Researchers could use classical high-performance computing for most of a simulation, AI to identify patterns or optimise parts of the process, and quantum hardware for selected experimental calculations.
Quantum computing remains at an early stage and will not replace the supercomputer. The value of the combination is experimental: it gives researchers a controlled environment for learning which problems, if any, benefit from distributing work across AI, conventional HPC and quantum systems.
Who will be able to use LUMI-AI?
The LUMI AI Factory is designed for European startups, small and medium-sized companies, universities, research institutes and larger industrial users. It combines computing access with data services, training, software tools and expert support.
Access will not mean unlimited free computing for every applicant. Projects typically pass eligibility and resource-allocation reviews, and current EuroHPC demand already exceeds available capacity.
The programme is particularly significant for organisations that have promising AI research but cannot sign enormous cloud contracts or build their own GPU clusters. Shared public infrastructure can lower the cost of experiments while keeping sensitive European data and model development within a regional governance framework.
European companies are already using the network. Reuters reported that French cloud provider OVHcloud completed initial training of a foundation model on Germany's JUPITER supercomputer, while Italian AI startup Domyn is using EuroHPC resources for model training.
What “sovereign AI” means in this project
European officials frequently describe LUMI-AI as sovereign infrastructure. In practical terms, that means Europe wants the ability to train important models, conduct research and run sensitive workloads without relying entirely on cloud platforms controlled outside the region.
The machine will still contain global technology. Its accelerators and processors come from US-based AMD, while IBM supplies storage. Nokia and Bull add European networking and system-design components.
LUMI-AI therefore does not create complete semiconductor independence. It does, however, give European institutions more control over where computing happens, who receives access and which legal and data-protection rules apply.
That is a more realistic form of technological autonomy than trying to manufacture every component within one region.
Can LUMI-AI close Europe's gap with the US and China?
One €387.8 million system will not erase the difference. American hyperscalers are investing tens of billions of dollars in data centers, and China continues to expand domestic AI infrastructure despite restrictions on advanced chip imports.
Europe's strategy is a distributed public network rather than a single enormous private cluster. Its success will depend on more than the advertised performance of each machine.
The important measures will include:
- How quickly researchers can obtain computing allocations
- Whether startups can move from experiments to commercial products
- How reliably the system operates at high utilisation
- The quality of software and technical support
- Whether European models and scientific discoveries emerge from the investment
- How well the network connects with private cloud and data-center capacity
The fact that EuroHPC is rejecting applications shows the demand exists. The larger question is whether new systems can arrive fast enough in a market where AI hardware changes rapidly.
What happens next
Bull and its partners must manufacture, integrate and install LUMI-AI before the planned second-half 2027 deployment. CSC will prepare the Kajaani facility, including power, networking, liquid cooling and heat-recovery systems.
More detailed specifications and access schedules are expected closer to launch. Performance claims will also need validation once the complete system is operating.
For now, the contract is a major industrial win for Bull and a clear signal that Europe considers shared AI computing a strategic asset. LUMI-AI will not match the combined scale of US hyperscalers, but it could give European researchers and companies access to hardware that would otherwise remain out of reach.
Thumbnail: official visualization of the planned LUMI AI Factory data center by NOAR – Nordic Architects, published by CSC.
