Bull has secured a €387.8 million ($445.1 million) contract to build the next-generation Lumi-AI supercomputer at CSC Finland's Kajaani datacenter, displacing Hewlett Packard Enterprise as system integrator.

The award came from the EuroHPC Joint Undertaking with co-funding from Finland, Czechia, Denmark, Estonia, Norway, and Poland. The new partition replaces the original Lumi cluster's GPU component, which HPE delivered in late 2021 with AMD Instinct MI250X accelerators.

What does the Lumi-AI partition deliver?

The upgrade targets 10X the AI training throughput and "about 2X" the FP64 HPC performance of the current Lumi-G partition, according to CSC Finland. Industry analysis projects the configuration at 1,240 nodes with 4,960 AMD MI430X accelerators, yielding a peak of 1.43 exaflops at FP64 precision.

Each node pairs four MI430X accelerators with a single 256-core Venice Epyc 9006 processor, yielding 317,440 CPU cores across the partition, a 66.7 percent increase over Lumi-G's 190,592 cores. The MI430X delivers 288 teraflops on FP64 workloads across both vector and matrix units.

The MI250X it replaces offered 47.9 teraflops on vector units and 95.7 teraflops on matrix units, with INT4 and INT8 precision grafted onto FP16 silicon. The MI430X adds native FP4 and FP8 support, multiplying performance gains as precision drops.

How does the cost compare?

At $445.1 million against 1.43 exaflops peak, the Lumi-AI partition costs roughly $312 per teraflops, undercutting the original Lumi cluster's 2021 price tag of $431 per teraflops by 27.6 percent.

• Contract value: €387.8 million ($445.1 million)
• Peak performance: 1.43 exaflops FP64
• Cost per teraflops: $312
• Projected node count: 1,240
• Projected GPU count: 4,960 AMD MI430X

Power draw for the new partition is estimated at 6.2 megawatts, against 7.1 megawatts for Lumi-G. Energy efficiency improves from 53.43 gigaflops per watt to roughly 75 gigaflops per watt on Linpack-class workloads.

Why did Bull beat HPE?

EuroHPC procurement prioritizes European suppliers. Bull, the French systems vendor behind the BullSequana XH3500 platform, qualifies under that criterion. Hewlett Packard Enterprise, which built the 2021 Lumi cluster, does not.

Bull's BXI fabric is being adapted to Ultra Ethernet Consortium specifications, giving BullSequana XH3500 customers a path to UEC switch hardware and creating differentiation that HPE's Slingshot interconnect cannot match as cleanly. The interconnect alignment reinforced the integrator selection.

When does the system go live?

Installation is scheduled for the second half of 2027 in a new datacenter building at Kajaani, and the Lumi-AI GPU partition will form one element of the broader Lumi AI factory, a term borrowed directly from Nvidia's marketing vocabulary.

The factory also includes a quantum subsystem branded Lumi-IQ. IQM Quantum Computers, a Finnish vendor that has raised €600 million in private capital and carries a $1.8 billion valuation, will supply a 150-physical-qubit Halocene H4 system in 2027 supporting 5 logical qubits.

A 2028 upgrade will improve error correction on the H4 platform. In 2029, CSC Finland will install a Halocene H5 machine delivering at least 9 logical qubits, implying roughly 270 physical qubits.

What competitive dynamics emerge?

AMD's Altair architecture gives the Lumi-AI partition native low-precision math that the previous Aldebaran silicon lacked, closing the FP4 and FP8 gap while positioning the MI430X for both HPC simulation and AI training workloads simultaneously. CSC Finland's MI430X deployment gives AMD a flagship EuroHPC reference site and gives Bull a marquee European AI factory contract to carry into subsequent bids against HPE and Lenovo.