Korea expects a top-ten TOP500 ranking for Hangang, but no HPL result was identified in the announcements reviewed. Service is targeted for December.
South Korea's Ministry of Science and ICT (MSIT) and the Korea Institute of Science and Technology Information (KISTI) opened the call for users of the country's sixth national supercomputer on September 9, 2026, with official service targeted for December. The HPE-built system, named Hangang (한강, after the Han River), has a theoretical peak of 614 petaflops, 24 times that of Nurion, the fifth national system, according to the ministry's September 9 release. Citing the most recent TOP500 list, the ministry estimates that Hangang will be listed as a top-ten-class supercomputer.
That ranking is an estimate. No result from HPL (High-Performance Linpack, the dense linear-algebra benchmark the TOP500 uses to rank systems) was identified in the announcements reviewed for this article. The machine reaches users after six tender rounds, four of which failed before the ministry raised the project budget by 53% in November 2024.
HPE's May 2025 announcement describes the system, which HPE calls KISTI-6, as an HPE Cray Supercomputing EX4000 with two partitions, one built on NVIDIA GH200 Grace Hopper Superchips and the other on 5th Gen AMD EPYC processors. Nodes connect over HPE Slingshot 400, the 400-gigabit-per-second generation of HPE's interconnect. SCN has covered controlled congestion tests of the earlier Slingshot-11 fabric on LUMI; those results come from a different installation and generation and have not been reported for Hangang. Storage is HPE Cray Supercomputing Storage Systems E2000, mixing NVMe and NL-SAS drives under HPE's Data Management Framework. Cooling is 100% fanless direct liquid cooling, in a government data center that HPE says was adapted for it. HPE quoted a rounded 600 petaflops of peak performance. The ministry's September figure is 614, with 205 petabytes of storage and a network rated above 400 gigabits per second.
The ministry's May 2025 plan, set out in its contract release, listed a CPU partition of 800 nodes, each with two 128-core AMD EPYC 9745 processors, and a GPU partition of 2,084 nodes with four GH200 Superchips apiece, plus 20 "GPU-fat" nodes carrying eight NVIDIA H200 SXM GPUs each. The planned GPU count totals 8,496, matching the September announcement. The listed Grace CPUs and CPU-only partition total 9,936, but the release does not specify how host CPUs for the H200 nodes are counted, so these documents cannot reconcile the full CPU inventory. The 2025 release also noted that the configuration could change during construction. The September release defines one allocation node as four GH200s.
KISTI and HPE concluded the contract on May 12, 2025, the ministry said in the same release, at 382.5 billion won, including 78 billion won for five years of maintenance. The Public Procurement Service said two companies bid. The project budget runs to 2031 and totals 448.3 billion won; the release assigns 315.4 billion won to building the system and the rest to facility work such as power distribution, commercial software, and maintenance, without reconciling those lines against the contract value. At the European Central Bank's reference rates for September 9, 2026, a cross rate of 1,336.2 won per US dollar (via Frankfurter), the contract comes to roughly $286 million and the project budget to about $335.5 million.
The ministry's November 2024 release says Hangang was originally due to enter service in 2025, and that four tender rounds failed as demand for GPUs surged with the generative AI boom and prices spiked.
On November 4, 2024, the National Supercomputing Committee, chaired by the science minister, raised the budget from the preliminary feasibility study's 292.9 billion won to 448.3 billion won, an increase of 155.3 billion won, or 53%. The ministry said it based the new figure on large systems built in Korea and abroad and on quotations from the main manufacturers, following a Ministry of Economy and Finance re-review of the project plan. The terms changed too. Construction and maintenance were merged into a single tender, and advance payments and financing were reworked, changes the ministry described as a way to encourage manufacturers to bid. SCN's August analysis of Korea's sovereign AI program described a similar pattern, with the state holding firm where it could act by administrative decision and giving ground where it needed private bidders.
A December 2024 re-tender scheduled main-system delivery for October 2025 and beta service for March 2026, Aju Business Daily reported. HPE was named preferred bidder over SK C&C in March 2025, according to Aju. The ministry's May 2025 release counted six tender rounds in total and set the official service start for the first half of 2026. By March 2026 etnews was reporting operation in the second half of the year, and KISTI president Lee Sik told Herald Business in May that the system would be built within the year. Yonhap's report on the user call cites a July 2026 target, about five months before the December date.
The September release says HPE completed installation and basic checks in August and that performance and stability testing is under way. Park Chan-yeol, who heads KISTI's National Supercomputing Division, told ZDNet Korea that performance and function tests would begin soon and take roughly a month, and that the December date depends on them finishing on time. KISTI has not announced an HPL submission schedule in the materials reviewed. A November entry would need to meet the TOP500's October 24 submission and November 1 update deadlines. That list is released at SC26 in Chicago, which runs November 15 to 20.
On the June 2026 list, Eni's HPC6 ranked eighth at 477.9 petaflops measured; Fugaku was ninth at 442.01, and Alps was tenth at 434.9, measured against a 574.84 peak. The two GH200 systems in that top ten ran HPL at 75.7% of peak (Alps) and 81.5% (JUPITER Booster, 1,000 measured against 1,226.28). Applied to 614 petaflops, that range gives roughly 465 to 500 petaflops, which would have placed eighth or ninth in June. This is analysis, not a forecast. It assumes a whole-system run and a submission before the deadlines, and November's field is not yet settled. Oak Ridge's Lux, which SCN examined in August, has a published FP64 peak of 317 petaflops, below the measured score of June's tenth-place system.
Korea's highest-ranked system on the June list was NHN Cloud's NIPA-CL1, which measured 137.4 petaflops on NVIDIA B200 GPUs and placed 20th; The Elec reported it uses 4,080 of them. Nurion, the national machine Hangang replaces, sat at 144th with 13.93 petaflops measured. The ministry's 2025 contract release cited Nurion's overload as one reason for replacing it.
Other sovereign programs handle headline performance differently, and at different stages. EuroHPC's LUMI-AI contract, signed on August 31 for a system due in 2027, named its accelerator but published no absolute performance figure; CSC told SCN it had not disclosed the accelerator count because AMD had not released the final MI430X specification. Japan's FugakuNEXT, as SCN reported, emphasizes application performance rather than TOP500 position, continuing RIKEN's earlier approach.
IonQ and KISTI finalized an agreement on December 23, 2025, to deliver a 100-qubit Tempo system for on-site integration with Hangang, with Megazone Cloud as a partner; the contract value was not disclosed. A three-way agreement with NVIDIA followed at GTC on March 18, 2026, covering plans to use CUDA-Q and NVQLink for a hybrid quantum-classical environment, etnews reported. Lee Sik told Herald Business in May that the IonQ system would be installed in the second half of 2026. The ministry's September 9 release does not mention the quantum system.
Hangang will run three programs, according to the ministry's September 9 release. A National Strategy Program gets about 30% of resources for national projects judged urgent and important, such as securing strategic technologies, with K-Moonshot and the seven SEED national projects first in line. An R&D Innovation Program gets about 55% and is open to researchers from universities, institutes, and industry. It has three tiers: grand-challenge projects at 50 nodes' worth of resources, creative research at 2 to 49 nodes, and a new exploratory tier of one node, granted after minimal review, for testing early ideas. A new Industrial Innovation Program takes about 15% and pairs compute with KISTI consulting and technical support. The ministry says it targets small and mid-sized companies and startups with R&D capability that have struggled to secure computing infrastructure.
The ministry has described the split differently in the past. Its November 2024 plan gave priority shares of 30% to AI-based core technology and 20% to underserved industrial fields, and the May 2025 contract release used another breakdown by field. The new programs use different categories from those outlined, so the shares do not establish a like-for-like change in industry access.
Pilot-service applications run from September 23 to October 8, with pilot use from late October to mid-November, the September 9 release says. Applications for official service open September 28 and close at 18:00 KST on October 16, through the National Supercomputing Center's website (ksc.re.kr). Proposals undergo document review and evaluation by a committee, and selected users receive allocations in December.
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.
