
China Keeps Fastest Supercomputer Crown as AI Data Centers Absorb New Hardware
China's LineShine tops the Top500 at 2.198 ExaFLOPS, but only 27 new systems joined the list — the fewest since 1993 — as hyperscalers route hardware to AI data centers.
- By
- Tom Whitfield
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
China's LineShine holds the title of world's fastest supercomputer at 2.198 FP64 ExaFLOPS, but the bigger story sits further down the November Top500 list: only 27 new machines joined the ranking, the lowest number since the project began in 1993. The list, published Monday at SC26 in Chicago, shows an industry pouring unprecedented capital into computing hardware while almost none of it lands in traditional supercomputing.
What stayed the same at the top?
The top ten performers and the ten most energy-efficient systems did not change position since June. LineShine, installed at the National Supercomputer Center in Shenzhen, delivers its 2.198 FP64 ExaFLOPS in the High Performance Linpack benchmark while drawing 42.22 MW.
The machine stands apart from other exascale systems in one respect: it uses no compute accelerators at all. It runs on the homegrown LingKun platform with 304-core Armv9-based LX2 processors at 1.55 GHz — 13,789,440 cores in total — linked by the proprietary LingQi interconnect. Because it relies purely on CPUs, it gains only 3.6X performance when moving from FP64 to mixed-precision data, a smaller uplift than accelerator-based systems achieve.
The lesson here is simple: raw FP64 leadership and AI-friendly throughput are increasingly different races.
How do the American exascale systems compare?
El Capitan at the Lawrence Livermore National Laboratory remains the most powerful U.S. machine, delivering 1.809 FP64 ExaFLOPS at 29.69 MW on AMD Instinct MI300A APUs and the HPE Cray EX255a platform — notably more power-efficient than LineShine.
Frontier at Oak Ridge National Laboratory, built on AMD 3rd Generation EPYC processors and Instinct MI250X GPUs, reaches 1.353 FP64 ExaFLOPS at 24.61 MW. Aurora at Argonne National Laboratory, using Intel Xeon CPU Max 9470 processors and Data Center GPU Max accelerators, offers 1.012 FP64 ExaFLOPS at 38.70 MW. All three American exascale machines share HPE's Cray EX platforms and Slingshot-11 interconnect.
Germany's Jupiter Booster at the Jülich Supercomputing Center remains the most energy-efficient exascale system. It uses Nvidia GH200 Grace Hopper Superchips, the BullSequana XH3000 platform and Nvidia's quad-rail InfiniBand NDR200 interconnect, delivering 1.0 FP64 ExaFLOPS at 15.79 MW.
Why did so few systems join the list?
The 27 newcomers represent 5.4% of all machines on the list. The strongest of them is South Korea's Hanggang at the KISTI Supercomputing Center, entering at position 12 with 342.6 FP64 PFLOPS on Nvidia GH200 Superchips and HPE's Cray EX254n platform. Hanggang also ranks eighth in the HPCG benchmark at 4.29 HPCG PFLOPS.
Aggregate HPL performance across the 500 systems rose a modest 4.4% in six months, to 19.57 FP64 ExaFLOPS. The entry bar climbed: a machine now needs 2.857 PFLOPS of sustained FP64 performance to qualify, up 7.3% from 2.662 PFLOPS in June.
Other markers moved as well:
- The 50th-ranked system rose 12%, from 49.85 to 55.81 PFLOPS; the 100th rose from 21.85 to 22.74 PFLOPS.
- Systems delivering at least 100 PFLOPS now number 27, up from 25 in June.
- Systems with compute accelerators grew from 275 to 291, or 58.2% of the list.
The takeaway is that the installed base is still upgrading, but fresh flagship deployments have nearly stalled.
Where is the hardware actually going?
The answer is commercial AI. Estimates by Epoch AI, cited by the Top500 authors, indicate that AI compute capacity added globally each quarter has increased by 2.5X per year.
The world's installed AI compute capacity already exceeds the equivalent of 15 million Nvidia H100 GPUs, according to Epoch AI. As of the fourth quarter of 2025, five major American hyperscalers controlled an estimated 71% of global AI compute resources, much of it in massive clusters.
The mechanism behind the weak Top500 intake is therefore straightforward: hyperscalers buy the newest silicon for AI workloads that the Linpack ranking does not capture, while national labs upgrade on multi-year cycles. Unless that allocation shifts, expect the Top500 to keep aging even as global compute shipments hit records.
Source: Tom's Hardware
More from Tom Whitfield
Show full bio
Staff writer covering companies, deals and strategy in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,002 articles
Related articles
china-s-lineshine-tops-top500-beats-el-capitan-on-all-cpu-design-cf4086a5
China's LineShine Tops Top500, Beats El Capitan on All-CPU Design
china-s-lingsheng-takes-supercomputing-crown-at-2-74-exaflops-peak-a751151c
China's LingSheng Takes Supercomputing Crown at 2.74 Exaflops Peak
matrix-enhanced-cpus-could-end-the-gpu-era-paper-claims-bc24c8dd
Matrix-Enhanced CPUs Could End the GPU Era, Paper Claims
alibaba-pulls-zhenwu-v900-production-to-q1-claims-3x-0241b0e7
Alibaba Pulls Zhenwu V900 Production to Q1 2027, Claims 3x



