LineShine Takes the TOP500 Crown, Marking China's First Number-One Supercomputer

A previously unlisted Chinese system called LineShine has claimed the top position on the TOP500 list, displacing Lawrence Livermore's El Capitan when the 67th edition was announced at ISC High Performance 2026 in Hamburg, Germany. It is the first time a Chinese submission has held the number-one rank in the list's history.

LineShine is installed in Shenzhen. Beyond that, the TOP500 announcement confirms it had not appeared on any prior edition of the list — making its debut at the summit position an unusual entry, to put it mildly. El Capitan, the AMD Instinct MI300A-based system that had held first place, now sits in second.

The broader list signals a meaningful inflection in the global exascale landscape. The TOP500 project itself has framed this edition as the opening of a "new global exascale era," language that tracks with a shift from a single dominant exascale-class system toward a competitive field of machines crossing that threshold. Exascale — a system capable of at least one exaflop, or 10¹⁸ floating-point operations per second, on the HPL benchmark — was crossed first by Frontier in 2022. That the milestone is now being reached by multiple systems across multiple geographies is the structural change the 67th list reflects.

On the efficiency side, the Green500 — which ranks systems by performance per watt rather than raw Linpack throughput — has a new leader of its own. KAIROS, powered by NVIDIA's Grace Hopper Superchip, ranked first on the Green500 as of June 2026. Grace Hopper pairs an Arm-based Grace CPU with a Hopper GPU on a unified memory fabric, a design explicitly intended to reduce the data-movement overhead that dominates power budgets in large-scale inference and simulation workloads. KAIROS taking the efficiency crown on that architecture is consistent with where the industry has been heading since NVLink-connected CPU-GPU systems began eroding the dominance of conventional x86-plus-discrete-GPU configurations.

Worth stepping back from the raw rankings for a moment. The TOP500 uses HPL — the High-Performance Linpack benchmark, a dense linear algebra solve — as its sorting metric. HPL is a well-understood, reproducible measure, but it is deliberately narrow. It does not capture memory bandwidth-bound workloads, sparse computation, or the mixed-precision arithmetic that dominates modern AI training and inference at scale. A system optimized for HPL rank and a system optimized for LLM pre-training at trillion-parameter scale are not the same machine, and the list does not claim otherwise. Readers who work in HPC procurement or large-scale ML infrastructure know this already, but it is worth stating clearly when a headline number changes hands.

What the ranking does capture is the sustained national and institutional investment required to build and operate a machine of this class. China's path to a number-one system has been interrupted repeatedly by export controls on high-end semiconductors — successive rounds of US restrictions have blocked sales of NVIDIA A100s, H100s, and their successors to Chinese entities. LineShine's arrival at the top of the list, on hardware whose provenance has not yet been detailed publicly, will prompt fresh scrutiny of what domestic Chinese chip development — from the likes of Biren, Cambricon, and CXMT on the memory side — has actually delivered, and whether any foreign components were involved through third-party channels. Those questions are unlikely to be answered quickly.

The Green500 result points in a different direction. KAIROS running Grace Hopper to the top of the efficiency list reinforces that the near-term frontier of performance-per-watt in HPC sits with tightly integrated CPU-GPU architectures rather than discrete, PCIe-attached accelerator configurations. As data center power constraints tighten across every major geography, that matters as much to hyperscale operators as it does to national laboratories.

The next TOP500 update will come at SC26 in November. Between now and then, the practical question for the HPC community is what workloads LineShine is actually running, and for whom — details the public announcement does not yet provide.