cd /news/ai-infrastructure/top500-list-adds-just-27-new-systems… · home › topics › ai-infrastructure › article
[ARTICLE · art-171696] src=hwbusters.com ↗ pub= topic=ai-infrastructure verified=true sentiment=· neutral

TOP500 List Adds Just 27 New Systems, the Fewest Ever, as Growth Shifts to AI Factories

The 68th TOP500 list, released on 16 November at SC26 in Chicago, added only 27 new systems — 5.4 percent of the 500 entries and the lowest turnover in the ranking's history, according to the official TOP500 announcement. China's LineShine held No. 1 at 2.198 exaflops on HPL, with El Capitan (1.809 EF), Frontier (1.353 EF), Aurora (1.012 EF) and JUPITER Booster (1 EF) unchanged behind it, while combined HPL performance across all 500 systems rose 4.4 percent in six months to 19.57 exaflops. The organizers attributed the stagnation to growth shifting toward AI factories, citing Epoch AI data that global AI compute additions grow roughly 2.5 times per year and that cumulative AI compute now exceeds 15 million H100-equivalents, about 71 percent of it owned by five US hyperscalers at the end of 2025.

key takeaway

The TOP500's 27 new entries and 4.4 percent six-month gain in combined HPL performance to 19.57 exaflops understate AI capacity growth, which Epoch AI estimates at roughly 2.5x per year and over 15 million H100-equivalents cumulative, because large training clusters skip the FP64 Linpack run.

by read3 min views9 publishedNov 18, 2026
TOP500 List Adds Just 27 New Systems, the Fewest Ever, as Growth Shifts to AI Factories
Image: Hwbusters (auto-discovered)

The 68th TOP500 list, released at SC26 in Chicago, has the same top ten as June and only 27 new entries, the lowest turnover since the ranking began. The organizers’ own reading is that the growth has moved to AI factories that never run Linpack for the list.

Every June and November the TOP500 list comes out, and usually something near the top moves. This time nothing did. The November 2026 edition, published on 16 November at SC26, keeps all ten leaders exactly where they were in June, and the rest of the TOP500 list barely changed either: only 27 of the 500 systems are new, 5.4 percent of the total. According to the official announcement, that is the lowest new-system count in the list’s history, and by a wide margin.

China’s LineShine stays at No. 1 with 2.198 exaflops on HPL, out of a theoretical peak of 2.736 EF. The CPU-only machine at the National Supercomputing Centre in Shenzhen took the crown from El Capitan in June, becoming the first Chinese system on top since Sunway TaihuLight in 2017, and it also leads the HPCG benchmark at 22 petaflops. El Capitan follows at 1.809 EF, then Frontier (1.353 EF), Aurora (1.012 EF) and JUPITER Booster, which sits exactly on the 1 EF line. Eni’s HPC7, Microsoft’s Eagle, HPC6, Fugaku and Alps complete the top ten.

The middle still grows, slowly #

The list is not frozen. The entry point at No. 500 rose 7.3 percent to 2.857 petaflops, No. 100 now needs 22.74 PF and No. 50 needs 55.81 PF. Combined HPL performance across all 500 systems reached 19.57 exaflops, up 4.4 percent in six months, and 27 machines now exceed 100 PF. Accelerators show up in 291 of the systems, or 58.2 percent.

The best new arrival is Hangang at South Korea’s KISTI, an HPE Cray EX254n system built on NVIDIA GH200, which enters at No. 12 with 342.6 PF.

Those are respectable gains for a mature ranking. For a field where a single AI training cluster can hold well over 100,000 GPUs, they are modest.

Where the hardware went #

The organizers say it themselves. Citing Epoch AI data, the release notes that additions to global AI compute capacity have been growing roughly 2.5 times per year, that cumulative AI compute now exceeds 15 million H100-equivalents, and that five US hyperscalers owned about 71 percent of it at the end of 2025. Their conclusion, also picked up by Network World: “This suggests that growth has indeed shifted toward AI factories, and from capability systems to capacity systems.”

Put simply, the biggest machines being built right now are not being submitted. A cluster running training jobs around the clock has little reason to stop for a multi-hour FP64 Linpack run, and much of that silicon is tuned for FP8 and FP4 work rather than double precision anyway. Eagle, on the list since 2023, is still the only cloud system in the top ten.

Green500 and mixed precision #

The Green500 top ten is unchanged as well. KAIROS at CALMIP and the University of Toulouse leads with 73.28 GFlops/W, ahead of ROMEO-2025 (70.91) and the Levante GPU extension at DKRZ (69.43). On the HPL-MxP mixed-precision benchmark, which is closer to how AI hardware actually gets used, El Capitan leads with 16.7 EF, ahead of Aurora at 11.6 EF and Frontier at 11.4 EF.

None of this makes the TOP500 list useless. It remains the most consistent public record of scientific supercomputing, with a method that goes back to 1993. But this edition’s stillness says more about who chooses to take part than about the state of the hardware. When the largest systems on the planet are private and never benchmarked in public, a list with no changes is mostly measuring its own coverage.

── more in #ai-infrastructure 4 stories · sorted by recency
── more on @top500 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/top500-list-adds-jus…] indexed:0 read:3min 2026-11-18 · —