EuroHPC's Largest AI Factory Supercomputer Contract Has No Published Absolute Performance Figure

The EuroHPC Joint Undertaking signed a procurement contract with Bull on 31 August 2026 for LUMI-AI, an AI-optimised supercomputer to be installed at CSC's data center in Kajaani, Finland. The system will be "co-funded with a total budget of EUR 387 800 000" for acquisition, delivery, installation and maintenance, split evenly between the Joint Undertaking and a hosting consortium of Finland, Czechia, Denmark, Estonia, Norway and Poland. The machine will run AMD Instinct MI430X GPUs and 6th Gen AMD EPYC 256-core processors. The JU expects it "to be installed and made available to users in 2027."

That is the largest budget EuroHPC has attached to an AI Factory procurement, about 1.34 times the next largest. It is also the only one of the six that names its silicon and publishes no absolute measure of what that silicon adds up to.

Instead, the announcements publish two ratios. The Joint Undertaking's release says the system "is expected to deliver a tenfold increase in artificial intelligence (AI) capacity compared to the current LUMI supercomputer." The host releases from CSC, LUMI and Bull add a second: "nearly double the high-performance computing (HPC) capability." Neither ratio names the data type it is measured in or defines its baseline, and no exaflops figure, GPU count or node count appears in any of them.

SCN put six questions to CSC on the day of the announcement. Pekka Manninen, director of the LUMI AI Factory at CSC, answered five on the record and declined one: the accelerator count. His reason for declining: AMD has not finished specifying the chip the machine is built on.

Six contracts, uneven disclosure

Anders Jensen, the Joint Undertaking's executive director, said the JU "has now signed its sixth AI Factory procurement contract for a next-generation AI-optimised supercomputer." The count reaches six by including the MareNostrum 5 AI upgrade, which extends an existing machine at the Barcelona Supercomputing Center rather than procuring a new one.

Every figure below comes from the Joint Undertaking's own release; host and integrator detail is discussed in the text.

* Each of the six releases states a budget covering the same four scope categories: acquisition, delivery, installation and maintenance. The MareNostrum 5 figure is the only approximate one, given as "around EUR 129 000 000."

Three of the six releases carry a peak AI performance figure, and they are the three that name NVIDIA hardware. The three without a figure are the two Bull contracts and the MareNostrum upgrade, and two of those, MareNostrum 5 and Sweden's Mimer, name no accelerator at all.

The exaflops figures that do exist are not a ranking. NVIDIA rates the GB200 NVL72 at "1,440 | 720 PFLOPS" of NVFP4 across 72 GPUs, sparse and dense respectively. MeluXina-AI's disclosed 1,008 GPUs and "over 20 Exaflops" are consistent with the sparse figure, though the JU release states neither data type nor sparsity. HammerHAI and IT4LIA publish no GPU count, so counts there can only be back-solved under assumptions. All three figures are floors, so budget divided by performance yields an upper bound: roughly €1.81 million per exaflop at IT4LIA, €3.67 million at HammerHAI, €4.00 million at MeluXina-AI. A 2.2x spread on broadly comparable NVIDIA silicon suggests the denominators are not scoped alike.

One thing the six systems have in common appears in none of their announcements. Every EuroHPC AI Factory system will carry a small prototype of European AI processors, Manninen told SCN. He named no supplier, size, or timeline.

CSC explains the missing count

Asked how many MI430X the system will hold and how they will be partitioned, Manninen told SCN: "We did not disclose the exact number yet, as AMD has not released the final exact spec of MI430X." The numbers stay "a bit vague for some time still," he said, and node counts will be published closer to general availability.

That is a fact about the part, not about the procurement. AMD's product page states that MI430X GPUs "are expected to be available in 2027," and every performance number on it is prefixed "up to," described as peak theoretical, and footnoted as subject to change on release.

Where the €387.8 million came from

The figure matches the budget EuroHPC published before it had a bid to read. The call for tender, reference EUROHPC/LUX/2025/CD/0001, opened on 23 May 2025, and its description reads: "The estimated total value for the call is EUR 387 800 000." A Wayback Machine snapshot taken on 3 June 2025, fifteen months before the award, carries that sentence verbatim. Europe's second exascale supercomputer followed the same path: the Alice Recoque call of 9 September 2024 gave an estimated total value of EUR 354 800 000, and the contract signed with Eviden on 18 November 2025 states the same number as its total budget.

The CD in the tender reference stands for competitive dialogue, which the Alice Recoque call page describes in the Joint Undertaking's own words: after the participation deadline, the JU will "initiate a competitive dialogue," evaluate "the market's technical offerings," and only then invite tenders. SCN's reading is that a published envelope frames what vendors propose. Neither the JU nor Bull has described how closely the final bid tracked the estimate.

Set the two contracts side by side, and the resemblance runs deep. Both are built on MI430X GPUs paired with the 256-core sixth-generation EPYC processors AMD code-named Venice and now sells as the EPYC 9006 series, a part SCN assessed for its core-density case in agentic infrastructure. Both use Bull's BXI interconnect, and both budgets cover the same four headline scope categories. Even the platform is shared: Alice Recoque's architecture is the Eviden Sequana XH3500, LUMI-AI's the BullSequana XH3500. So is the seller, the same Advanced Computing business through its ownership transition. It signed as Eviden in November 2025, four months before the French State completed its acquisition of Atos's advanced-computing activities, and as Bull in August 2026, into a Kajaani campus SCN examined during the Roihu procurement.

The AI Factory system carries €33 million more than the exascale flagship. Because both signed values mirror envelopes EuroHPC published before final bids, that difference reflects two differently sized procurements, not that Bull charged €33 million more for comparable hardware.

What the two ratios are measured in

LUMI runs the AMD Instinct MI250X today; LUMI-AI will run the MI430X. Both headline ratios now have a stated basis, and the two are not the same kind of measurement.

The tenfold claim, Manninen told SCN, is "FP4 peak flop/s, 'AI inference flops'": an MI430X FP4 peak set against an MI250X FP16 peak, because CDNA 2 has no FP4 rating to compare with. Neither FP8 nor FP4 appears anywhere on AMD's MI250X page.

Asked whether "nearly double" is measured against LUMI's peak or its Linpack result, and against which partition, he answered: "Just the GPU partition, and HPL." That anchors the capability ratio to an achieved benchmark result. LUMI's own documentation gives the LUMI-G partition as 11,912 MI250X modules with an aggregated HPL result of 379.70 PFlop/s, matching the Rmax on LUMI's TOP500 entry.

Manninen answered two separate questions and did not characterize the relationship between them.

Why the two ratios cannot reveal a GPU count

Readers will reach for the arithmetic, so it is worth doing in public and showing where it breaks. The table below is not an estimate of LUMI-AI's configuration. It demonstrates that the published ratios cannot be reconciled into a single machine size, because each rests on a different definition of its inputs.

Three caveats come first. The rows are not three estimates of one quantity; each carries a different assumption, and the assumptions are mutually incompatible. Nor do the derivations share a baseline machine, since the FP16 aggregate uses all 11,912 MI250X modules in LUMI-G while the TOP500 figures appear to describe a smaller submitted configuration. SCN also publishes no accelerator count. CSC's reason for withholding one applies at least as forcefully to an estimate made from outside.

The spread does not mean one accelerator yields two machine sizes. It means the inputs are defined differently, so back-solving from them produces counts that differ by nearly a factor of two. Achieved throughput and theoretical peak are not comparable quantities.

The part itself is still moving

AMD's published specifications for the two accelerators, retrieved on 31 August 2026, show how one upgrade produces a ratio near two and a ratio near ten.

One of those numbers has already changed. HLRS's Herder announcement on 2 December 2025 listed MI430X memory bandwidth as "19.6TB/s," a figure CEA's Alice Recoque page still carries, while AMD's page now lists 23.3 TB/s. All three state 432 GB of HBM4. Specifications published ahead of availability move, and SCN has tracked the same divergence on the NVIDIA side.

AMD positions the MI430X for "scientific HPC and sovereign AI" and claims 8.7 times the peak FP64 of NVIDIA's Vera Rubin, a figure footnoted as "based on engineering projections by AMD performance labs in July 2026" against a competitor part that has not been released either. Keeping double precision in hardware runs counter to a trend SCN documented in the industry's work reconstructing FP64 on AI silicon: AI-first accelerators have increasingly reduced native FP64, pushing scientific codes toward mixed-precision and emulation techniques.

LUMI-AI is the fourth announced MI430X supercomputer and the third in Europe. AMD's product FAQ, which predates this contract, names only Discovery, Alice Recoque and Herder. None of the four host institutions publishes an accelerator count of its own, which places CSC's position in line with the rest of the field rather than apart from it. Fabric splits the group: BXI on Alice Recoque and LUMI-AI, the Cray GX5000 platform's fabric on Herder and Discovery, and NVIDIA Quantum-X800 InfiniBand on two of the AI Factories, HammerHAI and IT4LIA, a divergence SCN examined in production-exascale interconnect behavior under AI traffic.

The Kajaani end of that schedule has both dates. LUMI-AI is a new build inside the same paper machine hall that houses the current LUMI, Manninen told SCN, in a facility CSC's public materials describe as "a completely new data center" on the site of a former paper mill. The two will overlap through 2028, he said, with the current machine planned for decommissioning late that year.

What remains unknown

CSC's answers close most of the file. The accelerator count and its partitioning were asked and declined, and the stated reason does not resolve until AMD finalizes the MI430X. The basis of the capability ratio is narrower than it was but still unsettled: CSC confirmed the baseline is LUMI-G's Linpack result, and did not say whether LUMI-AI's doubled figure is a peak or a projected benchmark result, which is the difference between the two back-solved readings of that ratio.

The disclosure asymmetry has a shape but no established cause. "Nearly double" appears only in the host releases; the Joint Undertaking published the AI ratio, no capability ratio, and no exaflops, GPU, or node figure of any kind. The sources reviewed for this article do not state why.

The procurement is signed, and the supercomputer is due in service in 2027. The accelerator it is built on is also due in 2027, and the published specifications have already changed once.