NTUA opened a 48-GH200 campus supercomputer in beta as the EuroHPC machine on its own Lavrion land, set to start at the end of September, Greece's minister said.
The National Technical University of Athens (NTUA) announced on September 16, 2026, that installation of a new central computing infrastructure is complete and open to its community: an OpenStack private cloud and ICARUS, a 24-node HPE cluster with two NVIDIA GH200 Grace Hopper Superchips in each node. The university's release (in Greek) counts 48 combined CPU-GPU units, states a total performance of more than 2 petaflops at FP64 precision, and calls ICARUS the most powerful supercomputing unit at a Greek university and one of the most powerful at any European university.
The operator's own documentation gives a different headline figure. The ICARUS docs landing page lists 3,456 Arm CPU cores, 48 Grace Hopper GPU devices, 24 GH200 compute nodes, and 3.2 PFlops FP64 Tensor Core, above a banner asking users to acknowledge that the system is in beta mode. Both numbers are NTUA's. Neither is a measured result.
Nine days on, the more consequential fact sits about 60 kilometers southeast of Athens. Greece's national EuroHPC supercomputer, DAEDALUS, is being built at the Lavrion Technological and Cultural Park, which belongs to NTUA. It ran the High-Performance Linpack benchmark by June and placed 31st on the TOP500. As of late September, EuroHPC JU's Greece page still describes it as currently being deployed, and Greece's digital governance minister said on September 1 that it would enter regular operation at the end of the month. The university that hosts Greece's national GH200 machine has procured a 48-superchip machine on the same silicon, from the same vendor, and opened it to its own faculty first.
NVIDIA publishes two FP64 peaks for the Hopper GPU inside a GH200: 34 teraflops for the vector units and 67 teraflops for the Tensor Cores. NTUA reproduces both on its Grace Hopper reference page, and adds a caveat that most operator pages omit: "all figures are vendor peaks, not measured application performance." The same page warns that Tensor Core figures should not be read as general-purpose CPU or GPU application throughput.
Multiply through, and the arithmetic is straightforward. Forty-eight superchips at 67 TFLOPS FP64 Tensor Core is 3.2 petaflops, per the docs. Forty-eight at 34 TFLOPS FP64 vector is 1.6 petaflops. The release's "more than 2 petaflops at FP64" does not fall out of either product. Adding the Grace CPUs' own double-precision throughput closes some of the distance, but NTUA has not published what the release figure counts, and SCN has not tried to reconstruct it. What can be said is that both are theoretical peaks for a machine that has not published an HPL run, and that the higher one is the Tensor Core figure. Readers who want the longer argument about what FP64 Tensor Core throughput means for simulation codes will find it in SCN's coverage of how the establishment is adapting science to AI silicon.
For context on scale: NTUA has published no Linpack result, and ICARUS is not on the June 2026 TOP500 list. That is normal for the campus tier. SCN's coverage of Maui's Makau, whose 2.64-petaflop peak sat just below the June 2026 entry threshold, describes what a machine at roughly this scale is for when a ranking is not the goal.
The practitioner-facing detail is in NTUA's docs, and it is unusually complete for a system nine days past its announcement.
The system messages page, timestamped September 9, 2026, records that ICARUS is operating in beta with 20 of the 24 Grace Hopper nodes available, providing 2,880 CPU cores and 40 GH200 GPUs. The current allocation call is labeled "Production Call (BETA)" and is limited to one node per job, three concurrent nodes per project, a 12-hour maximum walltime, and a one-month allocation. The terms of usage state that user data is not backed up and that no storage quotas are currently enforced; a retention period after project end has not yet been set. Compute nodes have no internet access, and NTUA's FAQ notes that an x86-64 binary will not run natively on the Arm CPUs.
That is what "installation complete" means in practice at a campus center in its first month, and NTUA has documented it more candidly than many larger facilities do. But the press release and the operator pages describe the same week from two different altitudes, and a reader who sees only the first will leave with the wrong picture of what a faculty group can do on ICARUS today.
The release attributes ICARUS to the Ministry of Education's "Universities of Excellence" program. Rector Ioannis Chatjigeorgiou, quoted by the Athens-Macedonian News Agency and carried by Skai on September 19, said the new acquisitions came "thanks to funding from the 'Universities of Excellence' Program" of the ministry (SCN's translation from Greek).
Universities of Excellence is a Greece 2.0 action, meaning European Union Recovery and Resilience Facility money routed through the national plan. The Greece 2.0 project page gives the program a total budget of EUR 229.0 million, of which EUR 176.6 million is RRF funding, running from September 6, 2022, to June 30, 2026. The program is a competitive call for research-infrastructure upgrades at selected universities. The ICARUS line item is not published, and NTUA's release does not give a price. The program end date explains the timing: an RRF action that closed on June 30, 2026, sets a ceiling on when the purchase could have been made.
DAEDALUS is the larger half of this story, and its timeline rests on dated public statements.
EuroHPC JU signed the procurement contract with HPE on March 28, 2025, after a tender launched in June 2024. The system was specified at over 89 petaflops, an absolute figure that EuroHPC does not always publish at contract stage; the LUMI-AI award went out without one. Total acquisition cost is EUR 36 million, 35 percent from EuroHPC JU and 65 percent from Greece 2.0. Cyprus, Montenegro, and North Macedonia joined the consortium. The contract announcement names the former power station building at the Lavrion Technological Cultural Park of the National Technical University of Athens as the site, with GRNET, the national research network, as operator under the Ministry of Digital Governance. The contract-day language said DAEDALUS "will be accessible in early 2026" to users across Europe, and EuroHPC's executive director Anders Jensen said he expected the system to be operational in 2026.
Greek News Agenda, the Greek foreign ministry's English-language news service, reported on March 4, 2026, that the machine, with more than 2,000 GH200 superchips in direct-liquid-cooled cabinets, had been assembled and tested at HPE's facility in the Czech Republic, would be installed at Lavrion in the summer of 2026, and would reach full operation and service through the Pharos AI Factory in the autumn.
By June, it had run Linpack. The TOP500 entry lists DAEDALUS as an HPE Cray EX254n with 411,264 cores, Slingshot-11, an Rmax of 85.69 petaflops and an Rpeak of 115.08, at 31st on the June 2026 list. EuroHPC's June 23 announcement said the machine was "expected to become fully available to European users shortly." GRNET's release the next day said it was expected to become operational during 2026.
On September 1, Digital Governance Minister Dimitris Papastergiou told the University of Western Macedonia's rectorate in Kozani that all project procedures for DAEDALUS would be completed at the end of September so the system could enter regular operation, PowerGame reported (SCN's translation from Greek). The same report says a second national AI-oriented supercomputer, budgeted at EUR 43 million, is planned for Kozani, with tender documents due between October and December and service possible by the end of 2027.
Three weeks on, the EuroHPC AI Factories page for Greece still lists DAEDALUS as being deployed, and the Pharos AI Factory featured the system in audiovisual material at EuroHPC User Days in Dublin on September 22 to 24. The machine exists and has run its benchmark. As of September 25, the minister's end-of-September date is the only specific one on the record, and it falls this week. SCN found no published access call for the national GH200 system on NTUA's land: GRNET's English news page, its HPC portal (where production calls 17 through 20 name ARIS, the predecessor system) and EuroHPC's access-calls page carried none when SCN checked the same day. Meanwhile, the campus GH200 system is taking faculty proposals under a beta call.
The structure will look familiar to readers who followed Finland's Kajaani campus, where a sovereign machine and a EuroHPC machine sit on one site with different owners and different access rules. Greece now has a version of the same layering, at a much smaller scale on the campus side and with the two tiers 60 kilometers apart rather than in one hall. The Greek case adds one detail Finland did not have: the same integrator and the same superchip top and bottom.
Analysis follows; the sourced facts are above.
The obvious reading of ICARUS is as a small machine beside a large one. Forty-eight GH200 against more than 2,000 is a ratio of about one to 43. The ratio says little about what a national technical university gets for the purchase, because aggregate flops were never the shopping list.
The first thing it buys is memory per problem. Each GH200 exposes up to 624 GB of combined CPU and GPU memory in a coherent address space, according to NTUA's reference page, so a two-superchip node can hold a problem of roughly 1.2 TB without a domain decomposition across the network. For computational mechanics, CFD and multiphysics at the scale of a faculty research group, the constraint has usually been fitting the mesh into GPU memory. On that axis, a 100 GbE fabric with no InfiniBand is a defensible design at 24 nodes; it would not be at 500.
The second is a software stack that its students have to learn. Because the CPUs are Arm, existing x86 builds do not run, and NTUA's FAQ says so. A cohort of engineering students at a national technical university will now compile for aarch64 and CUDA 13 by default, on the same superchip that JUPITER, Arrhenius and DAEDALUS run. That is the Software Ceiling at the campus tier: the porting burden lands on graduate students and lab codes, and the payoff is that whatever they build ports upward to the national machine without an architecture change.
The third is governance. ICARUS allocates through a faculty proposal call, an NTUA evaluation team, and core-hour accounting that NTUA sets, with GPU time converted to equivalent core-hours at a rate the billing policy says is specified by the ICARUS offering but does not publish. A EuroHPC machine cannot offer a single institution that kind of control over its own queue. For a rector, that is the difference between owning a resource and applying for one, and the release's emphasis on "research autonomy" reads that way.
What 48 superchips do not buy is a place in any ranking, headroom for multi-node training runs at the fabric speeds a Slingshot or InfiniBand system provides, or the current generation of NVIDIA silicon. GH200 is the Hopper generation, and Blackwell-based systems have been installed at national facilities since 2025; Taiwan's NCHC reported two GB200 NVL72 systems deployed in November of that year. For a buyer working against an RRF program that closed on June 30, 2026, the mature server generation was the one that could be procured, delivered, and installed inside the window. That is a rational choice at this tier, where SCN's coverage of Sweden's Mimer found a machine sized for services and access, with no ranking in view.
NTUA's claim to have one of the most powerful supercomputers at any European university is not one SCN can test. No registry distinguishes machines that a university owns from machines that a university hosts on behalf of a national program. Arrhenius at Linköping, which the June 2026 TOP500 placed at 42nd, has 382 Grace Hopper nodes of four GH200 each according to KTH's PDC newsletter, or 1,528 superchips; it is a EuroHPC mid-range system operated by NAISS, not a Linköping asset. JUPITER at Jülich has about 24,000 GH200. DAEDALUS itself sits on university land. Against machines a European university has bought with its own budget line and allocates itself, SCN has no comparable list, and the claim stands as NTUA's.
What the two Greek tiers do show, without needing that claim, is a single vendor and a single superchip at both levels of a national research-computing stack. HPE won DAEDALUS in an open EuroHPC tender. ICARUS is HPE ProLiant with HPE Cray Lustre storage, procured through a route NTUA has not published. Both are NVIDIA Grace Hopper. Both are paid for with Recovery and Resilience Facility money.
That alignment also creates practical benefits and a dependency. The benefits are shared HPE support relationships, a software environment that ports from campus to national scale, and a student pipeline trained on the hardware it will meet at the next tier up. The dependency is that Greek sovereign compute, campus and national, runs on NVIDIA's Grace CPU, built from Arm's Neoverse V2 cores, NVIDIA's Hopper GPU, and a US integrator. Most of the EuroHPC GH200 fleet runs the same NVIDIA silicon.
ICARUS becomes a working facility when its beta limits lift: all 24 nodes in service, allocations longer than a month, walltimes longer than 12 hours, and a published GPU-to-core-hour weight. A benchmark of any kind, HPL or an application run, would let a measured number replace the two theoretical FP64 figures. NTUA's docs place the machine on the Zografou campus; nothing published says what it draws or how it is cooled.
For DAEDALUS, the marker is simpler: a first EuroHPC access call naming the system, the step KISTI took for Korea's Hangang months ahead of service. Until one appears, Greece's national GH200 capacity remains, in EuroHPC's own phrase, being deployed, and the 40 superchips NTUA has in service are the ones a Greek university researcher can submit a job to today.
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.
