Finland's new 33.9 PF Roihu supercomputer triples national capacity and shares its hydroelectric Kajaani site with EuroHPC's LUMI - a two-tier sovereign-plus-EuroHPC campus under one operator.
Reported analysis. Specifications, contracts, timelines, and quotes are drawn from CSC, Bull, the EuroHPC JU, and LUMI public materials with sources linked. The framing of Kajaani as a two-tier sovereign-plus-EuroHPC campus and the read on Bull's first post-spin-off year are our interpretation.
Finland's national computing center, CSC, switched on its new flagship in Kajaani on May 21. The town has fewer than 40,000 residents, sits about 600 kilometers north of Helsinki, and has built itself out into a Nordic data-center cluster over the past decade. The system is Roihu, Finnish for "blaze," continuing CSC's single-word native naming pattern after Sisu, Mahti, and Puhti. Pilot users have been on the machine since April 29; general availability is set for June 2026, and Mahti retires in August.
Roihu replaces both Mahti and Puhti, triples Finland's national supercomputing capacity, and lifts the country's GPU performance more than tenfold against the predecessor pool. Bull, the European HPC vendor that became French-state-owned on March 31, delivered the BullSequana XH3000 under an initial €30 million contract, with options that could push the framework to €60 million.
That is what most outlets carried. The more interesting fact sits underneath the inauguration. Not the machine, the campus.
Roihu shares the Renforsin Ranta industrial park with LUMI, the EuroHPC Joint Undertaking's pre-exascale supercomputer operated by a consortium of eleven European countries (eight EU member states plus Iceland, Norway, and Switzerland). LUMI is community infrastructure. Roihu is wholly Finnish: nationally funded, nationally owned, nationally governed by CSC. Kajaani now runs both tiers of Europe's two-tier supercomputing model, the EuroHPC tier and the sovereign national tier, under a single operator on one private industrial grid drawing from one primary energy source.
That source is 100% hydroelectric, supplied via Vattenfall on the loop that powers the business park. LUMI's waste heat already covers roughly 20% of Kajaani's annual district-heat demand via the Loiste Lämpö network. A new 30 MW Caverion heat pump plant (12 MW commissioned in May 2026, 18 MW scheduled for 2027) is expected to push the data centers' contribution above 80% of city district-heat demand once fully online.
Our read: no comparable EU campus we are aware of co-locates both EuroHPC and a sovereign national flagship under one operator, on a single site, with a fully renewable supply and district-heat reuse at this scale. Sweden's EuroHPC AI Factory at Linköping is centered on a services-layer build-out for Mimer, not a parallel national flagship. France separates EuroHPC at CEA from national-tier systems through GENCI. Spain's MareNostrum 5 sits under the EuroHPC umbrella with national co-investment, not as two formally separate flagships. Italy's Bologna Technopole, with LEONARDO near CINECA's national-tier resources, is the closest analog. We have not done an exhaustive survey and welcome corrections.
Pekka Lehtovuori, CSC's Director of Advanced Computing Facility, Operations & Systems, framed Roihu's ownership as a deliberate sovereignty signal: "Ownership and governance of Roihu remain in Finnish hands. This strengthens national resilience and security of supply." Minister Mari-Leena Talvitie (Science and Culture) tied the procurement to Finland's national target of lifting R&D spending to 4% of GDP, and CSC's managing director Kimmo Koski cited a Taloustutkimus study finding that every euro invested in CSC's high-performance computing services returns 25 to 37 euros to Finnish society.
Roihu is a hybrid CPU/GPU BullSequana XH3000, direct-to-chip liquid-cooled throughout. Per CSC's Roihu page, the CPU partition runs 486 nodes with two AMD EPYC 9965 "Turin" processors each: 192 Zen 5c cores per socket, 384 per node, roughly 186,624 cores across the partition. The GPU partition has 132 nodes with four Nvidia GH200 Grace Hopper superchips each, for 528 GH200s total. Storage is 6.0 PiB of flash scratch plus 0.5 PiB of flash for project and home directories on two independent parallel file systems; CSC says the scratch tier delivers about 10× Puhti's I/O throughput.
Per docs.csc.fi, the interconnect is InfiniBand NDR at 200 Gb/s per CPU node, the operating system is Red Hat Enterprise Linux 9, and the programming environment is built around the GNU and AOCC compiler stacks, OpenMPI as the main MPI library, and CUDA plus the Nvidia HPC SDK for the GPU partition. The InfiniBand choice will weigh on how Roihu's AI workloads scale at partition level, given recent production-scale measurements showing large performance gaps between InfiniBand and Slingshot under AI traffic.
Public materials mix metrics on performance. CSC's procurement-era messaging quotes a 49 PF "theoretical peak"; the detailed systems documentation lists 10.5 PF HPL on the CPU side and 23.4 PF HPL on the GPU side, for a 33.9 PF HPL aggregate. Both are defensible. Theoretical peak and Linpack are different measures, routinely reported side by side, but they are not interchangeable, and several outlets have cited the 49 PF figure without flagging that the HPL number is lower.
What is not in the public record is an explicit EuroHPC funding share for Roihu. CSC's language consistently frames Roihu as a national procurement, and the Lehtovuori and Talvitie quotes strongly imply 0%, but no funding-decision document we can cite directly has surfaced.
Roihu is the first national-flagship system Bull has delivered since its spin-out from Atos and sale to the French State closed on March 31, 2026, at a €404 million enterprise value (€300M base plus up to €104M in earn-outs tied to FY2025/FY2026 profitability). That ended a six-year stretch with Bull's HPC engineering inside Atos, rebranded Eviden. Eviden kept Atos's cybersecurity, mission-critical systems, and Vision AI; HPC, Quantum, and AI compute went with Bull, now a €720M-revenue independent entity, ~3,000 staff across 32 countries, with what Bull describes as "the only HPC manufacturing plant in Europe" at Angers.
The engineering team did not change. The people who designed and integrated Roihu also delivered CSC's previous two flagships, Puhti in 2019 and Mahti in 2020, under the Atos badge, along with LEONARDO at CINECA, the CPU partition of LUMI, and MareNostrum 5 at BSC. The structural change is corporate: sovereign ownership, a single-focus mandate, an independent balance sheet. Whether a state-owned, single-focus Bull executes differently than Eviden-under-Atos did, the next year of deliveries will show. Roihu alone does not settle it.
Bull's 2026 order book is worth tracking on its own terms. Alice Recoque, the EuroHPC JU's next exascale system, was contracted to Bull on November 18 at SC25: a €354.8 million JU contract covering acquisition, delivery, installation, and maintenance, funded 50/50 by EuroHPC and the Jules Verne consortium led by France. AMD/Eviden messaging around the award also references a larger €554 million total project cost over five years that bundles operating costs beyond the system contract; the two figures are different and should not be conflated. Alice Recoque is a BullSequana XH3500 platform with AMD MI430X accelerators, installing at CEA's TGCC for user availability in 2027.
European exascale to date is JUPITER at Jülich, Europe's first HPL-exascale system, delivered by Eviden/Bull-era integration with Atos still in the parent, and confirmed at 1.000 EFlop/s on the November 2025 TOP500. Isambard-AI in Bristol, on HPE Cray hardware, is commonly described as exaflop-scale AI performance rather than HPL exascale, and the two should not be lumped together. Bull's first exascale system under the new corporate structure is Alice Recoque, and it is not on the floor yet. A fair read on Bull's current position: the sovereign-EU-vendor narrative is plausible; the exascale-share narrative is not yet supported by deliveries.
Roihu is being designed to a workload mix rather than a benchmark ranking. AI model training (mostly small-to-medium models), molecular dynamics, drug-compound library screening, climate and glacier scenarios, fluid dynamics, weather modeling, materials and life sciences, HPC pedagogy at higher-education scale, and a confidential-data processing capability CSC says is in active development. None of those rewards a top-of-list HPL number specifically, and the architectural choice tracks with Japan's decision to step FugakuNEXT away from the TOP500 chase in favor of throughput on production scientific workloads. The sensitive-data path is worth watching. Operational overhead for restricted workloads on a shared national flagship has grown under the NIST SP 800-234 HPC security overlay and equivalent European frameworks now landing in 2026. A hybrid quantum-HPC integration path is in scope but not yet specified to system level.
Mahti decommissions in August 2026; Puhti's compute resources go offline around July, with login and storage retained until August. "We'll donate the systems to suitable recipients for continued use or spare parts," CSC's Sebastian von Alfthan said in the center's September 2025 installation update.
The structural question Roihu sets up is whether the Kajaani two-tier model is a Finnish artifact (a country with the hydroelectric supply, cool ambient, land, and political will) or a template other EU member states can credibly copy. The answers will come from the next AI Factory awards under EuroHPC's 2026 calls, and from what Bull's sovereign structure does, or fails to do, to the European exascale order by 2028.
