On November 2, 2026, South Korea inaugurated Hangang, its sixth national supercomputer, at the Korea Institute of Science and Technology Information (KISTI) in Daejeon. The machine is not fully open yet. It is running a pilot for selected researchers, with full service scheduled for December 16, 2026.
Hangang is rated at 614.7 petaflops of theoretical peak performance. That number matters because it decides what scientists can attempt. A climate group can test more storm paths. A drug team can screen more compounds. A battery lab can model more atoms before building a physical cell.
What Switched On in Daejeon on November 2
The inauguration took place at KISTI headquarters in Yuseong-gu, Daejeon. Around 200 people from industry, academia, and government labs attended. The hosts were the Ministry of Science and ICT and KISTI.
KISTI President Lee Sik said the center would focus on stable operation and user support so domestic researchers can pursue challenging work. Kim Sung-soo, head of research and development policy at the ministry, described Hangang as core infrastructure for combining AI with science and technology.
The name Hangang comes from the Han River that flows through Seoul. The system succeeds Nurion, the fifth national machine introduced in 2018. Nurion supported 1,730 research projects over eight years and contributed to 1,520 papers in SCI-indexed journals, including 70 in Nature, Science, Cell and their sister journals.
What 24 Times the Last Machine Really Means
Hangang’s practical gain over Nurion is scale. A simulation that took a month could finish in a day or two, which lets teams test more ideas in the same grant period.
Storage rises to 205 petabytes. That space holds the vast datasets modern science creates, from weather observations to molecular libraries. Network speed exceeds 400 gigabits per second, which lets thousands of processors exchange results without waiting. If the network were slow, even fast chips would sit idle.
| System | Theoretical performance | Storage | Network |
|---|---|---|---|
| Nurion (No. 5, 2018) | 25.7 petaflops | 20 petabytes | 100 Gbps class |
| Hangang (No. 6, 2026) | 614.7 petaflops | 205 petabytes | More than 400 Gbps |
| Change | 24 times higher | 10 times larger | 4 times faster |
The ministry expects Hangang to rank in or near the global Top 10 when the next Top500 list is released at the International Supercomputing Conference SC2026. That is an expectation, not a confirmed rank. A formal High Performance Linpack result, which the Top500 ranking uses for ordering systems, had not been published in the announcements reviewed for this article.
Two Kinds of Chips, One Shared Instrument
Hangang carries 8,496 graphics processing units, including Nvidia GH200 Grace Hopper Superchips, alongside 9,936 central processing units. The two types do different work. CPUs handle large-scale computational science, such as fluid dynamics for a rocket engine. GPUs accelerate AI research, such as training a model that predicts new materials.
The GH200 pairs a Grace CPU with a Hopper GPU in one package with fast chip-to-chip memory access. That design helps AI-for-science codes that move constantly between simulation and learning. HPE built the system as an HPE Cray Supercomputing EX4000, with one partition built around GH200s and a second built around 5th Gen AMD EPYC processors, linked by HPE Slingshot 400 interconnect.
Cooling changed as well. Nurion used indirect liquid cooling, where cold water absorbed heat at the back of the cabinet. Hangang uses direct liquid cooling, where coolant circulates over the chips themselves. KISTI reports power usage effectiveness improved to about 1.15 from 1.35, which means less electricity is wasted on cooling. For context on why dense AI systems now live or die by this plumbing, see our look at liquid-cooled supernodes and power usage effectiveness.
Who Actually Gets Time on Hangang
Hangang is a shared lab instrument, not a single-team machine. KISTI will divide resources into three programs.
- R&D innovation, about 55 percent. Open to universities, institutes, and industry for computational science, AI, and data analysis. It includes grand-challenge projects at larger allocations, creative research at mid-size allocations, and a new one-node exploratory tier for testing early ideas.
- National strategy, about 30 percent. Reserved for urgent national projects, including K-Moonshot and strategic technology programs selected through government review.
- Industrial innovation, about 15 percent. Paired with KISTI consulting and technical support for small and mid-size companies and startups that cannot afford their own cluster.
The pilot reflects that mix. KISTI opened researcher recruitment in September, accepted pilot applications from September 23 to October 8, and is running pilot service for 189 projects across 64 institutions through November 22. Requested resources already amount to about 85 percent of what Hangang can provide.
The First Science: From Drug Candidates to Batteries
K-Moonshot is the flagship national use. The program uses advanced AI models to tackle major scientific challenges in areas such as advanced biology, future energy, and space. Hangang provides the shared compute behind it.
Two workloads show what that means for daily research. The first is AI drug discovery. Researchers now screen large compound libraries in software to narrow down candidates before any lab work begins. Hangang lets them analyze huge datasets, test how molecules bind, and compare design options entirely in simulation.
The second is materials development, including next-generation batteries. Here AI models and atom-level simulations work together. A model suggests a chemistry. A simulation checks stability. The lab then builds only the most promising options. Similar CPU-plus-GPU workflows already support vehicle crash testing, scientific weather forecasting, and liquid rocket engine combustion work that Nurion carried for years.
That hybrid approach, simulation plus learning, is now the global pattern for AI supercomputers. The United States followed the same path with systems such as Perlmutter, which paired large GPU partitions with fast storage to move from months of analysis to hours for cosmology and materials work. Read our earlier explainer on how Perlmutter sped cosmology and battery-interface science.
Why It Took Four Years and What It Cost
Hangang was a long build. The project passed a preliminary feasibility study in August 2022 with a budget of 292.9 billion won. Four bidding rounds then failed as generative AI demand drove GPU prices beyond the original budget.
In November 2024 the government revised the total project budget to 448.2 billion won. HPE was named preferred bidder in March 2025, and KISTI signed a build contract on May 12, 2025 for 382.5 billion won, including 78 billion won for five years of maintenance. The larger project figure runs to 2031 and covers facilities, power distribution, commercial software, and maintenance.
KBS reports total construction spending of 448.2 billion won, while English-language business coverage cites 382.5 billion won for the build contract and around 370 billion won in other tallies. The difference reflects contract value versus full program cost. We use the ministry figures above to keep the two separate.
This history matters for users. Failed bids pushed the schedule from an original 2025 target to pilot service in fall 2026. The lesson shaped the new allocation model, with a low-barrier exploratory tier so small teams can test ideas without waiting for a large grant.
What Happens Next: December 16 and a Top 10 Test
Full service is scheduled for December 16, 2026. Pilot users keep running through late November while KISTI completes performance and stability tests. Applications for regular service closed on October 16 through the National Supercomputing Center website, with selection by document review and committee evaluation.
The next public milestone is SC2026 in Chicago on November 15 to 20, where the Top500 will publish its November list. The ministry has said Hangang should place in the top-10 class based on the June 2026 list, where tenth place measured 434.9 petaflops. That comparison is illustrative. Real placement depends on a submitted whole-system run and a field that is not yet settled.
Looking further ahead, KISTI plans to connect Hangang to a 100-qubit IonQ Tempo quantum system in the first half of 2027, with a joint environment using Nvidia Grace Hopper tooling for hybrid quantum-classical work. The idea is simple. The supercomputer handles large accurate calculation. The quantum system handles optimization, or finding the right path quickly. Drug discovery is the cited example, where a quantum step narrows protein structures and the supercomputer computes binding energies. That hybrid link is planned, not yet live.
The contrast with other national strategies is sharp. China recently pushed a CPU-only exascale path to reduce dependence on foreign accelerators, as detailed in our breakdown of LineShine and compute sovereignty. Korea took the opposite route, buying best-available GPUs and focusing innovation on open access and efficient operation.
Hangang Supercomputer FAQ: Speed, Access, Top500 Rank
How fast is the Hangang supercomputer?
Hangang is rated at 614.7 petaflops of theoretical peak performance, or 614.7 quadrillion calculations per second. That is 24 times the theoretical performance of Nurion. Measured Linpack performance for ranking will be lower, and KISTI had not published a Top500 submission result at the time of the November 2 inauguration.
When will Hangang enter full service?
Pilot service began after September recruitment and runs through November 22 for 189 projects. Full service is scheduled for December 16, 2026. Some early ministry summaries listed December 14, but KISTI and later ministry briefings confirm December 16 as the operational date.
Who can use Hangang?
Universities, government-funded institutes, industry teams, and small businesses can apply through KISTI. About 55 percent of time supports open R&D, 30 percent supports national strategic R&D including K-Moonshot, and 15 percent supports industrial innovation with consulting help.
Is Hangang in the world Top 10?
Not yet confirmed. The ministry expects a top-10-class placement at SC2026 based on theoretical peak and the June 2026 cutoff. Formal entry requires a verified Linpack run submitted before the Top500 deadlines.
What will Hangang be used for first?
Early priorities include K-Moonshot projects, AI drug discovery, new materials and batteries, plus continued simulation work such as crash analysis, weather forecasting, and engine design. A quantum-hybrid program for optimization problems is planned for the first half of 2027.
Why Hangang Matters Beyond the Ranking
Rankings fade. Access lasts. Hangang matters because it gives thousands of researchers who could never buy 8,496 GPUs a place to test ambitious ideas, from a one-node pilot to a grand-challenge run. The machine is commissioned, in pilot, and not yet fully open. Watch December 16, then watch the papers. Hangang offers a route into work on that scale, where 24 times the compute of the last machine will show up.
