Top500 adds just 27 new supercomputers, a record low, as AI soaks up the hardware
The November 2026 TOP500 list, released at SC26, kept China's LineShine at No. 1 with an unchanged top 10 and only 27 new systems, the fewest in the list's history, as new chips flow to AI data centers instead.
Kiln, covering chips, data centers and power
, 3 min read
27new systems on the 500-machine list, a record low
What to know
- China's LineShine stayed No. 1 at 2.198 exaflops and the whole top 10 was unchanged from June.
- Only 27 of the 500 systems are new, 5.4% of the list, the lowest turnover since the list began.
- The list's organizers point to Epoch AI estimates that five US hyperscalers held about 71% of global AI compute at the end of 2025.
Source: TOP500 November 2026 release via HPCwire; Network World
The most telling number on the new TOP500 list isn't at the top. It's 27: the count of machines that are new since June, out of 500. That is 5.4% of the list, and the organizers' release, issued at the SC26 conference in Chicago on Monday, calls it the lowest new-system turnover in the list's history "by a wide margin."
The top didn't move either. China's LineShine, at the National Supercomputing Centre in Shenzhen, held No. 1 with 2.198 exaflops on the HPL benchmark. Behind it sit the same nine machines as in June, in the same order: El Capitan at Lawrence Livermore (1.809 exaflops), Frontier at Oak Ridge (1.353), Aurora at Argonne (1.012) and Germany's JUPITER Booster, which sits right on the 1.000 exaflop line. Network World notes those five are the only exascale systems on the list.
Where the chips went instead
The organizers asked the obvious question in their own release: if so few new supercomputers were built, where is the new hardware going? Their answer leans on Epoch AI's estimates. Cumulative AI compute worldwide has passed the equivalent of 15 million Nvidia H100 GPUs, quarterly additions are growing about 2.5 times a year, and five US hyperscalers owned an estimated 71% of global AI compute at the end of 2025. The release's conclusion is that growth has moved from scientific "capability systems" to the capacity systems inside AI factories, and that component shortages and AI demand, not a slowdown in technology, explain the stall.
ITPro, using the same Epoch data, puts two AI sites, SpaceX's Colossus 2 and OpenAI's Stargate campus in Abilene, at about 2.1 million H100-equivalents combined. None of those buildings runs HPL for the list, so the TOP500 is now mostly a census of public science machines, not of the biggest computers on Earth.
The list still grew, just slowly. Total HPL performance across all 500 systems reached 19.57 exaflops, up 4.4%. Work that backward and June's total was about 18.75 exaflops (19.57 ÷ 1.044), so the whole list added roughly 0.8 exaflops in five months, less than half of El Capitan. LineShine alone is about 11% of the list's total (2.198 ÷ 19.57), and the five exascale machines together are about 38% (7.372 ÷ 19.57). The entry bar for No. 500 rose 7.3%, to 2.857 petaflops.
The power yardstick
The efficiency numbers are where the two leaders differ most. LineShine is a CPU-only design, 304-core LX2 processors with no accelerators, and the June release put it at about 42.2 megawatts and 52.07 gigaflops per watt. El Capitan, built on AMD's MI300A chips, runs at 60.94 gigaflops per watt. That implies El Capitan draws about 29.7 megawatts for its HPL run (1.809 exaflops ÷ 60.94 gigaflops per watt), around 12.5 megawatts less than LineShine while producing about 18% less performance. For scale, one 1 GW AI campus of the kind counties are now voting on would draw roughly 24 LineShines' worth of power.
Accelerators keep spreading through the list: 291 systems, 58.2%, now use them. The Green500 top 10 was unchanged, led by KAIROS at Toulouse at 73.28 gigaflops per watt. The biggest newcomer was South Korea's HANGANG at KISTI, an HPE system on Nvidia GH200 chips, at No. 12 with 342.6 petaflops.
Fugaku held No. 9, even as RIKEN works through the breach that took it offline, which we reported Tuesday. Nvidia's answer to the science market's complaint that AI chips skimp on FP64 math is Rubin HPC, due in 2028; for why most of the list now leans on accelerators, see our GPU vs CPU explainer, and for what a gigawatt campus means on the ground, why data centers need so much power and water.
June brought two new entries to the top 10, LineShine and Italy's HPC7; November brought none. The question for June 2027 is whether any new science machine lands in time to break the freeze, or whether the chips keep going to buildings that never submit a score.
Sources
- SC26: TOP500 Leadership Holds Steady as New System Turnover Falls to a Historic Low, TOP500 release via HPCwire, November 16, 2026
- China's LineShine remains world's fastest supercomputer, Network World, November 16, 2026
- The AI data center boom is sapping the strength of the Top500, ITPro, November 17, 2026
- LineShine Debuts at No. 1 as the TOP500 Enters a New Global Exascale Era, TOP500, June 23, 2026
AI chips and data centersNvidia
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.