China's LineShine supercomputer, based in Shenzhen, kept its No. 1 position on the November 2026 TOP500 list, announced on November 16 (local time) at SC26 in Chicago. The top 10 systems were unchanged from the June list in both rank and performance, and the organizers said the number of new entries was the lowest ever. Comparing the official data for the June and November lists, however, the combined performance of the 500 listed systems rose by about 5.2%, and roughly 95% of that increase came from systems entering and leaving the list. The TOP500 ranks systems by their measured performance on HPL (High Performance LINPACK), a benchmark that solves a large system of linear equations using 64-bit floating-point numbers. That result is listed as Rmax.

AD

System turnover drove the TOP500's performance gain

According to the official TOP500 announcement, 27 supercomputers were newly listed this time, just 5.4% of the total. That is the fewest new entries since the ranking began in 1993. However, the official data files show 28 systems making their first appearance on the November list.

Counting the rows in the November Excel data where First Appearance is list 68 gives 28 entries, or 5.6% of the total. Comparing the two lists by System ID gives the same result: 28 systems appear only in the November list, and 28 systems from the June list dropped off. In the June list, 43 systems made their debut.

There are other discrepancies in the published figures. The announcement gives the combined HPL performance of the 500 systems as 19.57 exaflops and the performance of the No. 500 system as 2.857 petaflops. By contrast, the official November ranking and highlights page gives 19.70 exaflops and 2.868 petaflops, respectively. In the Excel and XML data files, rank and System ID match for all 500 systems, and Rmax also matches at the number of digits used in this article. The reason for the difference between the announcement and the data files, and which entries the announcement's 27 new systems refer to, cannot be confirmed from public materials.

Based on the data files, about 95% of the HPL performance added between June and November can be explained by turnover in the listed systems. Summing the Rmax of the 500 systems in each Excel file and converting from teraflops to exaflops, the total rose from 18.73 exaflops in June to 19.70 exaflops in November. That is an increase of about 0.97 exaflops, or roughly 5.2%.

Subtracting the combined Rmax of the 28 systems that were on the June list but dropped off in November from that of the 28 systems newly listed in November gives about 0.92 exaflops, or roughly 95% of the total increase. Meanwhile, the combined Rmax of the 472 systems common to both lists also rose, by about 0.046 exaflops.

Of the 28 new entries, just three systems—HANGANG in South Korea and Lux and Horizon in the United States—totaled 695.39 petaflops. That is about 70% of the combined HPL performance of the new systems. The performance needed to make the list has also risen: the No. 500 system went from 2.661 petaflops in June to 2.868 petaflops in November, an increase of about 8%. The No. 50 system rose from 49.85 to 56.16 petaflops, and the No. 250 system from 6.747 to 7.68336 petaflops.

LineShine stays on top, but rankings change by benchmark

LineShine debuted in June 2026 and became the first Chinese supercomputer to top the TOP500 since 2017. In the November list, the system, which uses a proprietary 304-core LX2 CPU, recorded an HPL performance of 2.198 exaflops. In second place, the US system El Capitan, built on AMD Instinct MI300A, recorded 1.809 exaflops. Japan's Fugaku came in 9th at 442.01 petaflops. The lineup of the five systems exceeding 1 exaflop in HPL—that is, one quintillion floating-point operations per second—is also unchanged from June.

However, HPL rankings do not represent performance across all kinds of computation. For example, Fugaku ranks third on the HPCG benchmark, which is more sensitive to reading data from memory and communication between nodes. HPCG includes operations such as multiplying a sparse matrix, in which most elements are zero, by a vector, and it measures performance characteristics different from those of HPL.

In the November HPCG results, LineShine was first at 22.00 petaflops, El Capitan second at 17.41 petaflops, and Fugaku third at 16.00 petaflops. HANGANG in South Korea, the highest-ranked of the newly listed systems, uses NVIDIA GH200 and placed 12th in HPL at 342.60 petaflops and 8th in HPCG at 4.29 petaflops.

Rankings shift further on HPL-MxP, which evaluates mixed-precision computing. This benchmark runs part of the computation at low precision and secures the required solution accuracy through a separate process. El Capitan took first place at 16.7 exaflops, while LineShine was fourth at 7.92 exaflops.

As these results show, even the same supercomputer can rank differently depending on the nature of the problem being computed and the numerical precision used. In AI model training as well, processing time is affected not only by computing performance but also by the efficiency of the training software and the speed of communication between nodes.

AD

Lux and Riken's "Riku": AI supercomputers for scientific research arrive

The US system Lux, which entered the list at No. 14, recorded an HPL performance of 220.79 petaflops. It is installed at Oak Ridge National Laboratory (ORNL) in the US, the same site as the third-ranked Frontier, and was built through a collaboration among ORNL, AMD, and HPE. It uses AMD Instinct MI355X GPUs and is positioned as a computing platform for applying AI to scientific research.

ORNL announced on November 16 that it had opened Lux to scientific researchers. Of the 3.5 million node-hours of computing resources available annually, about half is allocated to the US Department of Energy's Genesis Mission, and the remainder is available to AMD. A node-hour is a unit of computing resources equal to one node used for one hour. Selected research projects include building foundation models for materials development and forecasting future river flow and water temperature.

From Japan, Riken's AI for Science supercomputer "Riku" (理究) recorded 56.16 petaflops and placed 50th. According to Riken's November 13 announcement, Riku entered full operation on October 14. It is equipped with 400 NVIDIA GB200 NVL4 nodes, for a total of 1,600 B200 GPUs. The machine is used for research projects selected under the Ministry of Education, Culture, Sports, Science and Technology's AI for Science-related programs, as well as for Riken's own research.

Supercomputers built on commercial clouds are also on the TOP500. Seventh-ranked Eagle, an NVIDIA H100-based system built on Microsoft Azure, recorded an HPL performance of 561.2 petaflops.

The TOP500 is an application-based ranking that requires submission of measured results, so when a system is listed does not necessarily match when it began operating. For this November list, the deadline for new submissions was October 24, and the deadline for system installation and updating LINPACK measurements was November 1. Whether a system is listed depends on its progress from installation through performance measurement to submission. Even top-ranked LineShine has a registered installation year of 2025, but it first appeared on the TOP500 in June 2026.

The TOP500 alone can't measure the growth of AI computing infrastructure

The TOP500 organizers noted that new entries hit a record low even as the market for advanced computing equipment is expanding, and suggested that additions to computing resources may be concentrated in large-scale AI facilities. The view is that investment is shifting from traditional supercomputers focused on scientific and technical computing to facilities that provide large amounts of computing resources for training and running AI models.

Another angle on this trend comes from estimates of computing-capacity holdings by the AI research organization Epoch AI. According to Epoch AI's analysis, as of Q4 2025, five companies—Amazon, Google, Meta, Microsoft, and Oracle—held about 71% of the world's cumulative AI computing capacity, up from about 63% in Q1 2024.

This estimate converts each AI chip's nominal peak 8-bit performance into NVIDIA H100 equivalents and then totals the results by owner. However, it is a measure of how much computing capacity is held, not a direct measurement of the speed at which AI models are actually trained.

Knowing who holds the chips also doesn't reveal where they are installed. In a study Epoch AI published in September 2026, it reported that 86 tracked facilities account for an estimated 44% of the world's AI computing capacity. That estimate totals the H100-equivalent capacity of facilities judged to be operational and compares it with a global figure calculated on the assumption that it takes three months from chip delivery to operation.

With cloud services, the operator of a data center and the user of its computing resources can be different parties, and there is also uncertainty about how much visibility there is into data centers in China. Nor can these public data show how many procurement projects for conventional HPC systems were canceled or postponed because of rising demand for AI.

Another metric for comparing the training performance of AI systems is MLPerf Training from MLCommons. MLPerf Training v6.1, published on November 16, drew results from 24 organizations and added two new benchmarks.

One is "DLRMv4-HSTU" for recommendation systems, and the other is a post-training benchmark targeting the pretrained large language model "Qwen3.5-397B." MLPerf Training measures the actual time to reach a specified quality level using designated datasets. That makes it possible to compare the training performance of combined hardware and software for systems in the same task and comparison division.

The TOP500's HPL performance is one indicator of processing power in scientific and technical computing, and the scale of AI chip deployments shows the breadth of capital spending. MLPerf Training, meanwhile, measures the time to complete a specific AI training task, so each evaluates something different.

For science-focused AI supercomputers such as Lux and Riku, the allocations and selected projects published by the operating institutions also show what the investment is achieving. At Lux, for example, new computing resources are being allocated to research such as materials development and river forecasting. If more such systems appear on future TOP500 lists, they could offer a way to track not only changes in computing performance but also which scientific fields AI computing resources are actually being directed to.