China's National Supercomputing Center in Shenzhen, Guangdong unveiled LineShine, an all-CPU exascale supercomputer built entirely from domestic Chinese chips, interconnects, and storage. The machine targets 2 EFLOPS using Armv9-based LX2 processors, the proprietary LingQi fabric, and Chinese storage. Framed by Shenzhen officials as proof of self-reliance "across the entire stack," the system was notable for containing no GPU accelerators anywhere in its design.

At the time of that May announcement, the 2 EFLOPS figure remained a claim requiring independent validation, particularly significant because China had not submitted a Linpack result to the TOP500 since 2019. The 67th TOP500 list, published at the ISC High Performance 2026 conference in Hamburg, Germany, confirmed the claim and exceeded it. LineShine enters at No. 1 with 2.198 EFLOPS on the High Performance Linpack benchmark—around 80% of its 2.736 EFLOPS theoretical peak—making it the first machine ever to sustain more than 2 EFLOPS of double-precision performance using CPUs alone. The achievement pushes the US Department of Energy's El Capitan into second place at 1.809 EFLOPS and marks the first time a Chinese system has led the list since 2017.

The official entry credits 13.79 million cores built from 304-core LX2 chips running at 1.55 GHz, linked by the LingQi interconnect and running the domestic Kylin operating system on a platform called "LingKun." LineShine also takes the top spot on HPCG, the benchmark that better reflects the data-intensive patterns of real applications.

On HPL-MxP—the mixed-precision benchmark that tracks the kind of lower-precision math AI training actually relies on—LineShine lands only in fourth place, at 7.92 EFLOPS. Its mixed-precision score is just 3.6 times its standard Linpack result, while accelerated American machines clear nine times. This small multiplier is the signature of a system with no dedicated low-precision hardware: outstanding at traditional 64-bit science, comparatively ordinary at AI.

LineShine draws approximately 42 MW to reach the top, for an efficiency of about 52 GFLOPS per watt; El Capitan delivers its slightly lower score at nearly 61 GFLOPS per watt. To contextualize: 42 MW is the entire installed capacity of RWE's Serra Giannina wind farm in southern Italy—six turbines built to power some 39,000 homes. Because the supercomputer runs around the clock while a wind farm only generates when the wind blows, LineShine at full tilt would consume more electricity over a year than that entire wind farm produces. China reached No. 1 with scale and electricity rather than efficiency.

LineShine is now the world's fastest supercomputer on the industry's official list, built entirely from Chinese parts: the processors, the networking, and the operating system. The whole point of US export controls was to stop China reaching this level. China reached it by using enormous numbers of ordinary processors and significant electricity consumption—a triumph of scale and determination rather than cutting-edge efficiency. However, on the benchmark that reflects AI training—the capability the export controls were genuinely targeting—LineShine sits in fourth place, well behind American systems. So China has built a magnificent machine for traditional science: drug discovery, climate and weather modeling, and engineering simulation. The export controls have not stopped China from building a supercomputer, but so far they have succeeded in stopping the building of an AI supercomputer.