EuroHPC's EUR 29.76M Bull contract deploys a new AI-optimised supercomputer in Linköping - a tenth of IT4LIA's budget, with the services model as the differentiator.
Editor's note (July 16, 2026): An earlier version of this article's headline and framing characterized the EUR 29.76M Mimer contract as buying a services layer rather than a new system. Per EuroHPC JU's April 21 announcement, the contract covers the acquisition, delivery, installation, and maintenance of a new AI-optimised supercomputer at Linköping University, with services and software layers built around that new hardware. The headline, dek, and several passages have been revised accordingly. Our analysis of Mimer's scale relative to peer AI Factories, and of its services-first access model, is unchanged.
Sweden already had GPUs. That matters for reading what EuroHPC JU actually bought on April 21.
The agency signed a EUR 29.76M contract with Bull to deploy the Mimer AI Factory in Linköping, hosted by NAISS at Linköping University with RISE. Installation begins this year. The press release leans on the usual categories: life sciences, materials science, autonomous systems, gaming. The target user base is SMEs and startups, with the wider research community sharing access.
What the announcement underplays is more interesting. Mimer is not Sweden's JUPITER. It is a new, deliberately modest machine wrapped in an unusually developed services envelope.
Start with the budget. At EUR 29.76M, split 50/50 between EuroHPC's Digital Europe Programme and the Swedish Research Council, Mimer lands well below the current European AI infrastructure tier. JUPITER ran EUR 273M. LUMI passed EUR 202M. MareNostrum 5 cost EUR 151.41M for the machine alone, with the AI upgrade adding another EUR 129M on top. IT4LIA, Italy's AI Factory signed days earlier, came in at EUR 290M and was credited with 160+ exaflops of peak AI inference. Mimer is roughly a tenth the size of its Italian cousin on paper. Treat the announcement accordingly.
That framing matters because the "AI Factory" label is doing real work across the program. It is not an architecture class. It is a procurement wrapper EuroHPC stretches across everything from brand-new flagship machines to platform upgrades to pure services ecosystems. Mimer is closer to the last of those. The gap between the Factory branding and what any given contract actually funds keeps widening.
Architecturally, Bull is disclosing the scaffolding and almost nothing else. The system will run on the BullSequana XH3500 platform with direct liquid cooling and the BullSequana AI software stack, delivered with cloud-style access. GPU count, memory capacity, and peak FLOPS are not in the public release. For a machine sold on its usefulness to SMEs rather than its top-line numbers, that may be deliberate. It still leaves the technical substance opaque.
The Bull name here carries new weight. On January 29, 2026, Bull was relaunched as an independent brand after a French-state sovereignty acquisition separated it from Atos/Eviden. Mimer is one of the first major EuroHPC wins under that relaunch, and the sovereign-AI framing is being applied hard: European software stack, European procurement, European operating model. The framing is fair as far as it goes. It is worth saying that "sovereign" in this context means European control of the platform, software, and access policies, not proof of a European silicon supply chain. That distinction matters and is being elided across the program, not just here.
The more honest question is what Mimer actually adds to Sweden's stack. The contract does buy new compute - a BullSequana XH3500 system whose acquisition, delivery, installation, and maintenance account for the full EUR 29.76M. What it doesn't buy is capacity Sweden was short of: Linköping already hosts Berzelius, Sweden's premier AI/ML system, with 94 DGX-A100 nodes and 16 DGX-H200 nodes on shared VAST storage, funded by the Knut and Alice Wallenberg Foundation and running since 2021. Tetralith sits alongside it as the older general-purpose CPU cluster. KTH/PDC runs Dardel, a Cray EX with AMD EPYC plus MI250X GPUs and a smaller Grace Hopper partition. NAISS also holds a Swedish share in LUMI. The country is not GPU-starved.
What Sweden lacked was an AI-Factory-shaped service envelope: a procurement aimed squarely at SMEs and startups, sensitive-data workflows, and industrial onboarding, with a pricing and access model the academic systems were never built to offer. Mimer is already publishing that service model in the open, with access tiers that start at 5,000 GPU hours for quick experimentation and scale up to 22 million GPU hours for large-scale industrial workloads. That is the real news. Mimer is as much a services catalogue as it is a system.
The national subtext reinforces the point. Sweden's upcoming Arrhenius machine, a separate EUR 68.5M EuroHPC mid-range system clocking 40+ petaflops, is also heading to Linköping. Between Arrhenius and Mimer, Sweden is quietly consolidating its national HPC/AI footprint around a single site rather than extending the older distributed model. That is a procurement strategy worth watching.
The sustainability line deserves a caveat. EuroHPC pitched Sweden as attractive on GHG emissions and electricity cost per kWh, the same logic that proved out for LUMI in Finland. Fair enough for Arrhenius, which leans on it explicitly. Mimer's announcement, though, does not disclose power draw, PUE, or waste-heat reuse. The Nordic operating environment helps this system too, but the energy case for Mimer specifically is asserted, not quantified.
None of this makes Mimer a bad procurement. A smaller, services-shaped AI Factory aimed at SMEs and startups, with a public access tier structure and a relaunched European vendor behind it, is a reasonable thing for Sweden to buy. It is just a different thing than the headline suggests. Read Mimer as a practical access layer for the Swedish and European AI base, not as a new flagship, and the contract makes sense. Read it the other way and the numbers will not add up.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
