EuroHPC Signs $450 Million Deal for AMD-Powered LUMI-AI
Hardware / analysis
EuroHPC Signs $450 Million Deal for AMD-Powered LUMI-AI
The €387.8 million contract promises a tenfold jump in AI capacity over the current LUMI supercomputer by 2027, though EuroHPC did not say in what unit that capacity is measured.
The EuroHPC Joint Undertaking signed a contract on Aug. 31, 2026, to deploy the LUMI-AI supercomputer, a EUR 387.8 million ($450 million) system built on AMD chips that have not yet shipped anywhere else. The system is expected to be installed and available to users in 2027, EuroHPC said in its announcement.
EuroHPC's sixth AI Factory procurement, in the words of executive director Anders Jensen, will run on AMD Instinct MI430X GPUs, a chip AMD has not yet shipped in any production system, paired with sixth-generation, 256-core AMD EPYC processors. The system will be hosted at the CSC IT Center for Science data center in Kajaani, Finland. EuroHPC's own announcement describes the MI430X as AMD's upcoming design, not a part with an existing shipping history anywhere else.
What EuroHPC actually signed on Aug. 31
The budget splits evenly: half from the EuroHPC JU's Digital Europe Programme funding, half from the LUMI AI Factory consortium of Finland, Czechia, Denmark, Estonia, Norway and Poland. Bull, the French systems vendor, won the contract after a call for tender EuroHPC opened in May 2025, following Finland's selection as host site in December 2024. Bull will build the liquid-cooled system on its BullSequana XH3500 architecture, part of the same BullSequana product line the company says underpins Europe's exascale-class systems, according to DataCenterDynamics. Supercomputing time will be allocated between EuroHPC and the consortium in proportion to what each side paid, and the LUMI AI Factory, already running on the current LUMI machine since April 2025, has been serving startups and SMEs in manufacturing, health and life sciences, and communications technology, alongside climate research, materials science and language-technology users EuroHPC named as additional beneficiaries.
| Bull AI Factory contract | Value | Funding | Duration |
|---|---|---|---|
| LUMI-AI (Finland) | EUR 387.8 million ($450 million) | 50% EuroHPC JU, 50% LUMI AIF consortium | Installed by 2027 |
| Mimer (Sweden) | EUR 30 million ($35 million) | Five-year contract | Ongoing |
A tenfold claim with no baseline attached
EuroHPC said LUMI-AI is expected to deliver a tenfold increase in AI capacity over the current LUMI supercomputer, which DataCenterDynamics has reported peaks at 380 petaflops. Neither EuroHPC nor Bull said in what unit the promised tenfold gain is measured, at what numerical precision, or on which workload. A petaflops figure for dense scientific computing and a throughput figure for low-precision AI inference are not comparable numbers, and a vendor claim that does not specify which one it means is not a number a reader can check against anything. The same gap showed up when Samsung claimed an eightfold speed increase for its zHBM memory without a shipping date attached, and when Intel built a chip around tokens per watt and then declined to publish the tokens-per-watt figure.
Who builds it, and who owns the builder
Bull is no longer a private contractor bidding for public money at arm's length. The French state completed its acquisition of Bull, the AI, HPC and quantum unit of Atos, in April 2026, following a EUR 404 million ($469 million) share purchase agreement signed July 31, 2025, DataCenterDynamics reported. A state-owned company is now building a state-and-consortium-funded supercomputer, on top of the EUR 30 million, five-year contract Bull separately signed to supply Sweden's Mimer AI factory using the same BullSequana XH3500 architecture. Bull CEO Emmanuel Le Roux called LUMI-AI a milestone for both European AI and for Bull itself, building on what he described as the company's track record in exascale systems. IBM will supply the storage layer, based on IBM Storage Scale, and Nokia will handle data center networking alongside Bull's own BXI interconnect. CSC managing director Kimmo Koski said LUMI-AI's capabilities, combined with a separate quantum computer named LUMI-IQ that the LUMI AI Factory is also building, are meant to help European researchers and industry tackle what he called unforeseen research and innovation challenges in health, climate and other fields.
- AI Factories19 count
- AI Factory Antennas13 count
- Supercomputers procured12 count
Source: EuroHPC Joint Undertaking, accessed 2026-09-11
What's still unbuilt
LUMI-AI is one system inside a network EuroHPC says now spans 19 AI Factories and 13 AI Factory Antennas, on top of 12 supercomputers the joint undertaking has procured altogether, including the JUPITER and Alice Recoque exascale machines. At the end of July 2026, EuroHPC opened a separate call for tenders for what it calls AI Gigafactories, a larger tier still, a new pillar added to the joint undertaking's mandate under an EU regulation adopted in January 2026 that also expanded its quantum computing remit. Every number in that network describes money committed or contracts signed, not hardware running today. LUMI-AI itself will not take a single job until 2027, on chips AMD has not put into production anywhere else first, in a contract that took more than two years to move from Finland's December 2024 selection as host to a signature on Aug. 31, 2026. What would change this read is a slip in that 2027 date, or an independent benchmark once the system is live that finally puts a unit on the tenfold claim EuroHPC is making today.
Sources
More in Hardware
- 01Waymo Targets Singapore for 2028, Two Rivals Already Carry RidersWeRide and Pony AI have carried invited and paying riders through Singapore's Punggol district since April, roughly two years before Waymo's own timeline puts a rider in one of its cars there.
- 02Royal Enfield Prices Flying Flea at €5,990 Abroad, ₹2.79 Lakh at HomeNew Atlas pegs the electric motorcycle's April price in India at roughly $3,000 by direct conversion, and Royal Enfield has already lived through the same gap once with a gasoline model.
- 03Nvidia Won't Call Its Working Rust GPU Track Production-Readycutile-rs already backs an open-source LLM server and a Hugging Face testbed, but Nvidia's Sept. 8 announcement stops short of endorsing either new track for production.
- 04Arm Reuses the Total Design Name for Robots, Not Yet the SiliconThe original Total Design already has a customer-ready chiplet on TSMC's N2 process; the physical AI version Arm announced Sept. 8 is a set of robot-capability definitions.