The June 2026 TOP500 list marks an important shift in the recent history of supercomputing. Two years after Stackscale’s article on the world’s 500 most powerful supercomputers and Linux’s absolute dominance in this field, the debate is no longer only about the operating system. The new picture of the sector points to exascale computing, technological sovereignty, energy efficiency, heterogeneous architectures and a competition increasingly linked to Artificial Intelligence.
The main headline from the 67th edition of the TOP500 is the direct entry of LineShine into first place. The system, installed at the National Supercomputing Centre in Shenzhen, China, displaces El Capitan and becomes the world’s most powerful supercomputer according to the High Performance Linpack (HPL) benchmark. Its sustained performance reaches 2.198 Exaflop/s, above El Capitan’s 1.809 Exaflop/s, which moves down to second place.
LineShine and China’s Return to Number One
LineShine does more than return China to the top of the TOP500 for the first time since Sunway TaihuLight in 2017. It also introduces a technically interesting nuance: it exceeds two sustained exaflops on HPL with a CPU-based architecture, without relying on GPU accelerators like those that dominate much of today’s Artificial Intelligence infrastructure debate.
The system uses the LingKun platform, 304-core LX2 processors running at 1.55 GHz, the proprietary LingQi interconnect and Kylin OS. In total, it includes 13,789,440 cores and draws approximately 42.2 MW of power. The figure is huge, but it helps explain the kind of scale involved: modern supercomputing is no longer measured only by the number of processors, but by the relationship between compute, power consumption, networking, cooling, system software and the real ability to run scientific workloads in a sustained way.
There is another important detail. LineShine also ranks first on the HPCG benchmark, with 22.00 HPCG-Petaflop/s. HPCG does not replace HPL, but it provides a view that is closer to real-world application patterns, with greater emphasis on memory access, communications and system behaviour under less ideal workloads than a pure LINPACK test.
This result does not automatically mean that LineShine is the most advanced system for every Artificial Intelligence workload. In HPL-MxP, a benchmark focused on mixed-precision performance, El Capitan remains the leader with 16.7 Exaflop/s, followed by Aurora and Frontier. LineShine ranks fourth, with 7.92 Exaflop/s. That difference matters because many modern training and inference workloads benefit from accelerators designed for mixed precision and massive matrix operations.
The TOP500 remains an essential reference, but it needs to be read carefully: it measures high-performance scientific computing capacity, not every possible kind of performance in Artificial Intelligence, cloud or private hyperscale platforms, many of which are not submitted to this ranking.
Five Exascale Systems and a More Diverse Top 10
The June 2026 edition raises the number of systems capable of exceeding one sustained exaflop on HPL to five. That threshold, which for years was a laboratory objective, is now starting to consolidate as a real category of scientific infrastructure.
| Rank | Supercomputer | Country | HPL Performance |
|---|---|---|---|
| 1 | LineShine | China | 2.198 Exaflop/s |
| 2 | El Capitan | United States | 1.809 Exaflop/s |
| 3 | Frontier | United States | 1.353 Exaflop/s |
| 4 | Aurora | United States | 1.012 Exaflop/s |
| 5 | JUPITER Booster | Germany | 1.000 Exaflop/s |
| 6 | HPC7 | Italy | 571.5 Petaflop/s |
| 7 | Eagle | United States | 561.2 Petaflop/s |
| 8 | HPC6 | Italy | 477.9 Petaflop/s |
| 9 | Fugaku | Japan | 442.01 Petaflop/s |
| 10 | Alps | Switzerland | 434.9 Petaflop/s |
The table shows more than a race for first place. The United States retains three exascale systems among the top four, Germany keeps Europe’s first system above the exaflop threshold with JUPITER Booster, Italy strengthens its position with HPC7 and HPC6, Japan keeps Fugaku in the Top 10 and Switzerland enters with Alps.
The diversity of architectures is also notable. LineShine relies on Chinese silicon and a CPU-only design. El Capitan and Frontier combine AMD EPYC processors with AMD Instinct accelerators. Aurora represents the Intel path with Xeon CPU Max and Data Center GPU Max. JUPITER Booster and Alps are based on NVIDIA Grace Hopper. Eagle, deployed on Microsoft Azure, combines Intel Xeon with NVIDIA H100. Fugaku remains a distinctive case with Fujitsu A64FX processors based on Arm.
The business reading is clear: there is no single winning architecture for every workload. Infrastructure is designed according to the type of computing, latency, networking, power consumption, cooling, available software, technological sovereignty and operating cost. That same logic, at a different scale, applies to the design of private cloud, bare-metal platforms and enterprise virtualisation environments.
An organisation running critical databases, virtual desktops, analytics platforms, private Artificial Intelligence models or latency-sensitive applications should not choose infrastructure based only on the number of vCPUs or raw capacity. Prior architecture matters: CPU, memory, storage, networking, isolation, licensing, backups, high availability, observability and day-to-day operations.
Linux, Energy Efficiency and Infrastructure Sovereignty
Stackscale’s 2024 article focused on a fact that remains central: Linux dominates supercomputing. The new list does not change that conclusion. LineShine uses Kylin OS; JUPITER Booster appears with Red Hat Enterprise Linux; HPC7 and HPC6 use RHEL; other Top 10 systems run HPE Cray OS, TOSS or environments derived and adapted for large HPC installations. The exact form varies, but the pattern is the same: open, modifiable systems prepared to be tuned to the hardware.
Linux fits supercomputing for reasons very similar to those that explain its importance in cloud, bare-metal and enterprise infrastructure: it allows teams to control the system, remove unnecessary components, tune the kernel, integrate specific drivers, automate deployments, work with high-performance networks and operate highly specialised environments without depending on a black box. It is not only a licensing cost issue; it is technical control.
Energy efficiency adds another layer to the debate. The TOP500 measures power, but the Green500 ranks systems by performance per watt. In June 2026, KAIROS, installed at CALMIP / University of Toulouse-CNRS, retains first place with 73.28 Gigaflops/Watt. It is followed by ROMEO-2025, also in France, and the Levante GPU extension in Germany. All three share a BullSequana XH3000 architecture with NVIDIA Grace Hopper and NVIDIA InfiniBand NDR200 interconnect.
Efficiency is not a secondary issue. As systems grow, the limit is no longer only the purchase of hardware. It also appears in available energy, cooling, data centre space, electrical design, heat reuse, rack density and the ability to keep the system running in a stable way. In enterprise data centres, something similar happens, although at another scale: poor architecture can end up generating more cost, more power consumption and more operational complexity than the problem it was supposed to solve.
Spain maintains a relevant presence with MareNostrum 5 ACC, the accelerated partition of the Barcelona Supercomputing Center’s supercomputer, ranked 16th in the June 2026 TOP500 with 175.30 Petaflop/s of HPL performance. Although it has left the Top 10 it occupied in previous lists, it remains one of Europe’s most important scientific infrastructures and a key asset for research, simulation, biomedicine, climate, engineering and Artificial Intelligence.
Public and scientific supercomputing is not the same as the private cloud infrastructure a company uses for its critical applications, but both share several lessons. The first is that real performance depends on the complete system, not on an isolated component. The second is that technological sovereignty is built through architectural, operational and location decisions. The third is that efficiency starts before buying hardware: it starts with design.
In enterprise environments, this means assessing when it makes sense to use public cloud, when a private cloud platform is preferable, when a workload needs dedicated bare-metal and when an open alternative such as Proxmox VE can help reduce licensing dependency, improve TCO control and avoid vendor lock-in. Not every workload needs the same infrastructure, but critical workloads do require an architecture designed before migration.
Stackscale works precisely in that middle ground between traditional enterprise infrastructure and modern requirements for performance, control and sovereignty. Private cloud, bare-metal, private networks, storage, backups, high availability and managed operations are part of a conversation that is no longer only technical. For CIOs, CTOs and financial decision-makers, the question is how to sustain critical workloads with predictable costs, operational control and data hosted in Europe.
The new TOP500 2026 list should not be read only as a ranking of giant machines. It is a signal of where infrastructure is heading: more specialisation, more importance for networking, more energy pressure, more debate around sovereignty, more Linux and a greater need to choose the right architecture. Exascale remains far from most companies, but its principles already influence much closer decisions: where to host data, how to isolate critical workloads, which platform to use, how to protect backups, how to scale without losing control and how to prevent infrastructure costs from becoming unpredictable.
Frequently Asked Questions
What is the world’s most powerful supercomputer in 2026?
According to the June 2026 TOP500 list, the world’s most powerful supercomputer is LineShine, installed at the National Supercomputing Centre in Shenzhen, China, with 2.198 Exaflop/s on the HPL benchmark.
What does it mean for a supercomputer to be exascale?
An exascale system is capable of exceeding one exaflop of performance, meaning at least one quintillion floating-point operations per second. In June 2026, five systems surpassed that threshold on HPL.
Why does Linux dominate supercomputing?
Linux makes it possible to adapt the operating system to the hardware, remove unnecessary components, tune performance, integrate high-performance networks and work with very specific architectures. That is why it remains the usual foundation for major HPC systems.
What does supercomputing have to do with enterprise private cloud?
Although the scale is different, they share core principles: architectural design, dedicated performance, efficiency, data control, low-latency networks, reliable storage and secure operations. For many companies, these ideas translate into private cloud, bare-metal and open virtualisation platforms.
Source: TOP500, June 2026 TOP500 List and Highlights
