China’s LineShine supercomputer has retained its position as the world’s fastest machine in the 68th edition of the Top500 list, published at the SC26 conference in Chicago on Monday. The ranking, which tracks the most powerful computing systems globally, recorded an exceptionally low number of new entries this cycle. This trend highlights a major shift in the industry, where unprecedented demand for high-performance computing hardware is being directed toward artificial intelligence data centers rather than traditional scientific supercomputers.
LineShine Architecture and Performance
Installed at the National Supercomputer Center in Shenzhen, LineShine delivers a sustained FP64 performance of 2.198 exaflops in the High Performance Linpack benchmark. The system operates at a power consumption of 42.22 megawatts. It relies on the homegrown LingKun platform, utilizing 13,789,440 cores of 304-core Arm-based LX2 processors clocked at 1.55 GHz. The machine also uses the proprietary LingQi interconnect.
Unlike other exascale-class systems, LineShine does not use compute accelerators. It relies exclusively on CPUs. This design choice results in a lower performance gain when moving from FP64 to mixed-precision data, achieving only a 3.6x increase. This is significantly lower than systems that incorporate low-precision accelerators designed for AI workloads.
The United States holds the next three spots on the list. El Capitan at the Lawrence Livermore National Laboratory ranks second with 1.809 FP64 exaflops at 29.69 megawatts. It uses AMD Instinct MI300A accelerated processing units and the HPE Cray EX255a platform. Frontier at Oak Ridge National Laboratory ranks third with 1.353 FP64 exaflops, while Aurora at Argonne National Laboratory takes fourth place with 1.012 FP64 exaflops. All three American systems use HPE Cray platforms and Slingshot-11 interconnects.
Energy Efficiency and New Entrants
Germany’s Jupiter Booster at the Jülich Supercomputing Center remains the most energy-efficient exascale supercomputer. It delivers 1.0 FP64 exaflops at just 15.79 megawatts. The system employs Nvidia GH200 Grace Hopper Superchips and the BullSequana XH3000 platform. Neither the top 10 performers nor the top 10 most energy-efficient machines changed positions since the June list.
Only 27 new supercomputers made the November Top500 list. This represents 5.4% of all machines on the list, the lowest number of new entries since the project began in 1993. The most powerful newcomer is South Korea’s Hanggang system at the KISTI Supercomputing Center. It ranks 12th with 342.6 FP64 petaflops and uses Nvidia GH200 Grace Hopper Superchips.
The aggregate performance of the top 500 systems increased by 4.4% over six months to 19.57 FP64 exaflops. The threshold to qualify for the list rose to 2.857 petaflops, up 7.3% from June. The number of systems equipped with compute accelerators increased from 275 to 291, accounting for 58.2% of the list.
Industry priorities have shifted from building machines for complex scientific problems to deploying compute capacity for commercial AI. Estimates from Epoch AI indicate that global AI compute capacity added every quarter has increased by 2.5 times per year. The world’s installed AI compute capacity now exceeds the equivalent of 15 million Nvidia H100 GPUs. Five major American hyperscalers accounted for 71% of global AI compute resources in the fourth quarter of 2025. Most new hardware deployed for these purposes does not appear in the traditional supercomputer rankings.
Source: Tom’s Hardware

