China has taken the world’s fastest supercomputer crown from the United States. China’s CPU-only LineShine overtook El Capitan with 2.198 exaflops on the biannual TOP500 ranking released Tuesday at the ISC 2026 conference in Hamburg, a 20 percent lead and the first time a Chinese system has topped the list since Sunway TaihuLight in 2017. LineShine is installed at the National Supercomputing Centre in Shenzhen.
The system is the first to cross the two-exaflop line without using any graphics processors at all. El Capitan drops to No. 2 after holding the top spot since November 2024. US export controls on advanced chips, the policy designed to slow Chinese progress in computing, did not stop the debut. They may have shaped it.
China’s LineShine Tops the TOP500 Ranking
The TOP500 project has measured supercomputer performance twice a year since 1993, when computer scientists Erich Strohmaier and Hans Meuer first compiled statistics on the machines in preparation for a conference on the topic. The list ranks systems by how long they take to solve dense systems of linear equations on the LINPACK benchmark. The benchmark rewards raw floating-point throughput above almost everything else. LineShine delivered 2.198 exaflops on that test, about 80 percent of its 2.736-exaflop theoretical peak, according to the TOP500 project’s announcement of LineShine taking No. 1.
El Capitan, at Lawrence Livermore National Laboratory in California, fell from No. 1 to No. 2 with 1.809 exaflops. China once dominated the list, with Chinese systems taking nearly half of the 500 slots in 2019. Chinese participation then dwindled as relations between Washington and Beijing soured, a silence that ended Tuesday.
Five Exascale Systems, Reshuffled
LineShine’s debut pushed the number of systems sustaining more than one exaflop per second on the LINPACK test from four to five. For the first time, an exascale machine now sits in each of Asia, North America, and Europe at the same time. The reshuffle moved El Capitan, Frontier, and Aurora down a notch each while keeping all three US Department of Energy machines above the one-exaflop line. JUPITER Booster, at the Jülich Supercomputing Centre in Germany, holds Europe’s sole slot above the threshold.
| Rank | System | Site | Country | HPL (Exaflop/s) |
|---|---|---|---|---|
| 1 | LineShine | National Supercomputing Center, Shenzhen | China | 2.198 |
| 2 | El Capitan | Lawrence Livermore National Laboratory | United States | 1.809 |
| 3 | Frontier | Oak Ridge National Laboratory | United States | 1.353 |
| 4 | Aurora | Argonne National Laboratory | United States | 1.012 |
| 5 | JUPITER Booster | Jülich Supercomputing Centre | Germany | 1.000 |
El Capitan is built on HPE Cray EX255a hardware using AMD 4th-generation EPYC CPUs paired with AMD Instinct MI300A accelerators. Frontier, at Oak Ridge National Laboratory, runs on a similar AMD-based design with the older MI250X accelerators. Aurora, at Argonne National Laboratory, remains the only Intel-based exascale machine on the list, and El Capitan still beats it on raw throughput with 1.809 exaflops versus Aurora’s 1.012.
Italy’s Eni S.p.A. just installed a new AMD-powered machine, HPC7, which debuted at No. 6 with 571.5 petaflops. Microsoft’s Azure-based Eagle system fell to No. 7 with 561.2 petaflops, combining Intel Xeon CPUs with NVIDIA H100 accelerators. Japan’s Fugaku, the previous CPU-only leader and now the oldest design still on the top 10, holds No. 9 at 442 petaflops. Switzerland’s Alps research supercomputer rounds out the top 10 at 434.9 petaflops.
Other countries represented in the top 20 include the United Kingdom, Japan, South Korea, Italy, the Netherlands, and Switzerland. Italy’s HPC6 system maintains a slot in the top 10. Finland’s LUMI and Italy’s Leonardo, the two systems that held No. 9 and No. 10 last edition, fell just outside the new top 10 at No. 11 and No. 12.
A CPU-Only Machine Built to Dodge Chip Curbs
LineShine is built differently from every other machine at the top of the list. It runs entirely on general-purpose central processing units, the kind in any laptop, rather than on the graphics processors that power most AI work and that sit at the center of US export controls. The TOP500 list describes LineShine as the first and only system to achieve more than two exaflops using a CPU-only design. That distinction matters in a year when US export controls continued to target the most advanced AI chips, the hardware that LineShine chose not to use.
- Total cores: 13.79 million
- Processor: Custom Chinese LX2, 304 cores per chip, 1.55 GHz
- Interconnect: Proprietary LingQi
- Operating system: Kylin OS
- Power draw: 42.2 megawatts
- Efficiency: 52.07 gigaflops per watt
- Theoretical peak: 2.736 exaflops per second
The system runs on a custom Chinese processor called LX2, with 304 cores per chip and 13.79 million cores in total running at 1.55 GHz. Linked by a proprietary interconnect called LingQi, the machine draws about 42.2 megawatts for an efficiency of 52.07 gigaflops per watt, below El Capitan’s 60.94.
LineShine chief designer Lu Yutong said at the Hamburg conference that the system “abandoned the conventional CPU-GPU architecture” in favor of domestically developed processors with built-in AI acceleration, high-speed memory, and a proprietary interconnect. Huang Xiaohui, the centre’s deputy director, said in April that LineShine had achieved full-stack independence from underlying hardware to core software. The architecture is liquid-cooled to keep the high-transistor CPUs from overheating.
Why Export Controls Failed to Slow This Down
The United States has restricted sales of advanced AI chips to China for several years, with the rules tightened in successive rounds. The controls target the GPUs that train models like ChatGPT and Claude, the hardware that LineShine chose not to use. The policy debate over what US chip export controls have and haven’t curbed in Chinese AI is ongoing. Jack Dongarra, an emeritus computer science professor at the University of Tennessee and one of the TOP500 organisers, said LineShine showed China capable of holding its own in advanced computing despite US export restrictions.
Export controls may slow China’s access to certain advanced components, but they also provide a strong incentive to develop domestic alternatives.
Dongarra added that “LineShine suggests that China has responded through large-scale investment and hardware-software codesign.” He said the controls may both constrain China and accelerate its efforts to become technologically self-sufficient. Sunway TaihuLight, which held the No. 1 spot in 2016 and 2017, was a homegrown Chinese design, and LineShine extends that lineage of nationally developed silicon.
The 2026 AI Index, released in April by Stanford University’s Institute for Human-Centered AI, puts the broader US-China technology competition in hard numbers. Its twelve key findings, drawn from a year of model evaluations, report that US and Chinese models have traded the lead in performance multiple times since early 2025. As of March 2026, Anthropic’s top model leads the field by 2.7 percent. China leads in publication volume, citations, patent output, and industrial robot installations, while the US still produces more top-tier models and higher-impact patents. Private investment tells a different story: US companies spent $285.9 billion on AI in 2025, roughly 23.1 times the $12.4 billion China spent.
- US private AI investment in 2025: $285.9 billion
- China private AI investment in 2025: $12.4 billion
- US/China private AI investment ratio: 23.1x
- AI scholars moving to the US since 2017: 89% decline
- AI scholar decline in the last year alone: 80%
Addison Snell, cofounder of the consultancy Intersect360 Research, said the US still leads globally in technology, but the gap is not wide. The number of AI scholars moving to the US has dropped 89 percent since 2017, with 80 percent of that decline coming in the last year alone. Digital sovereignty has become a central topic in supercomputing and AI today, Snell added.
The Case That the Ranking Tells Half the Story
The LINPACK benchmark, used by the TOP500 since 1993, measures one specific thing: how fast a system can solve a dense system of linear equations using 64-bit floating-point math. It rewards raw floating-point throughput above almost everything else. Dongarra said the ranking assessed one benchmark and should not be viewed as a complete measure of technological leadership. Scientific application performance, energy efficiency, software maturity, reliability, ease of use, and the ability to support a broad research community are equally important, he said.
A 2015 paper from Cornell University estimated that El Capitan achieved only 22 percent of the computational performance of xAI’s Colossus supercomputing facility in Memphis, Tennessee. Colossus is not on the TOP500 list, because it is built for AI training, not for LINPACK-style dense linear algebra. The 22 percent figure is a snapshot, not a permanent measure, but it captures why the ranking misses a large slice of modern compute.
Snell made a similar point in broader terms. It is a mistake to assume AI dominance will automatically translate to science dominance, Snell said. Consumer applications like image generation, translation, or chatbots have relevance to high-end computing but are not sufficient in themselves. Policy should reflect AI for science, not AI or science. To enable AI for science, governments must invest in both halves.
The TOP500 list is made up almost entirely of government and academic systems that volunteered to participate. Corporate tech giants such as Microsoft, Amazon, Meta, and Alphabet are at the forefront of today’s advances in AI but have largely chosen not to submit their flagship systems. Snell noted that hyperscalers such as Amazon and Microsoft would be able to claim the top spots on the TOP500 list if they wanted to. They have not, and that absence shapes what the ranking can and cannot tell a reader.
The picture the TOP500 draws is therefore narrower than the full map of compute power. It captures the systems built for national-laboratory science, where governments have been willing to fund a top-of-list machine. It misses the commercial AI factories, the cloud regions, and the internal clusters that no company is required to disclose.
China’s Quiet Years on the List
The 67th edition of the TOP500, with LineShine at No. 1, also marks a return to public form for Chinese entries on the list. China once dominated the list, with Chinese systems taking nearly half of the 500 slots in 2019, before fading from the public rankings. China’s previous No. 1, Sunway TaihuLight, held the top spot in 2016 and 2017 before being overtaken by the US-built Summit system in 2018. The intervening eight years saw Chinese systems largely absent from the public rankings, a silence that ended with Tuesday’s announcement.
Since entering service, the system has been used for climate modeling, engineering simulations, drug discovery, neuroscience, and AI, according to the National Supercomputing Centre in Shenzhen. The New York Times reported that LineShine was developed without public funding, allowing its designers to submit the system for TOP500 benchmarking this year. The funding path matters because publicly funded Chinese systems were widely believed to face political constraints on submitting benchmark results during periods of tension with Washington. Snell said the re-engagement of Chinese developers with the ranking project was as noteworthy as LineShine’s raw performance numbers. The broader tech decoupling has hit other countries’ industrial plans too, including India’s battery cell projects, as detailed in the report on India’s pause on battery cell plans over Chinese tech hurdles.
Frequently Asked Questions
What is the TOP500 list?
The TOP500 is a twice-yearly ranking of the world’s most powerful supercomputers, published since 1993. It measures performance using the LINPACK benchmark, which times how long a system takes to solve a dense system of linear equations using 64-bit floating-point math. The 67th edition was released at the ISC 2026 conference in Hamburg on Tuesday.
How fast is LineShine compared to El Capitan?
LineShine now leads El Capitan by 20 percent on the LINPACK benchmark, with both systems sitting above the two-exaflop line. Frontier, Aurora, and JUPITER Booster round out the top five at 1.353, 1.012, and 1.000 exaflops respectively. The full top five are the only systems verified above one exaflop on the current list. LineShine also takes the No. 1 position on the HPCG benchmark with 22.00 HPCG-petaflops.
What makes LineShine’s CPU-only design notable?
LineShine is the first system on the TOP500 to exceed two exaflops using only general-purpose central processing units, with no graphics processors at all. Its custom Chinese LX2 processors contribute 13.79 million cores running at 1.55 GHz, linked by a proprietary interconnect called LingQi, drawing 42.2 megawatts of power on Kylin OS.
What does LineShine mean for US chip export controls?
Jack Dongarra, an emeritus computer science professor at the University of Tennessee and one of the TOP500 organisers, told Al Jazeera that the controls appeared to push Beijing toward developing its own advanced components rather than capping its progress in compute. Addison Snell of Intersect360 Research said the result should ripple through the US, Europe, and Japan as countries compete for AI dominance. Snell added that the US still leads globally in technology, but the gap is not wide. The 2026 AI Index from Stanford University found that China had effectively closed the AI model performance gap with the US. With the rapid pace of evolution, the global order could change quickly, Snell told Al Jazeera.
Why did China go quiet on the TOP500 list after 2017?
Chinese participation in the TOP500 dwindled in recent years as relations between Washington and Beijing soured, with Chinese systems taking nearly half of the 500 slots in 2019 before fading from the public rankings. China’s previous No. 1, Sunway TaihuLight, held the top spot in 2016 and 2017 before being overtaken by the US-built Summit system in 2018. The New York Times reported that LineShine was developed without public funding, which allowed its designers to submit the system for TOP500 benchmarking this year.
