
November 2026 Top500 List Logs Record-Low 27 New Supercomputers
Only 27 new supercomputers entered the November 2026 Top500 — half the norm and the lowest churn since 1993, as sites await Rubin, Altair and faster networking.
- By
- Rebecca Stone
- Filed
- Channel
- AI & Compute
- Read
- 4 min read
Only 27 new supercomputers made it onto the November 2026 Top500 list — roughly half the normal count and the lowest churn across all 68 rankings compiled since 1993. The stall, visible in machines formally tested with the High Performance LINPACK (HPL) benchmark, marks the weakest installation cycle the list has ever recorded.
Twice a year, the Top500 lands: in June at the International Supercomputing conference in Europe, and in November at Supercomputing in the United States. Despite varied budgets and upgrade cycles at HPC centers worldwide, install volume has historically held fairly steady. Not this time.
What caused the drop?
One possible read: priorities are shifting toward renting cloud capacity rather than acquiring a new machine every four or five years and building a datacenter for ever-hotter, ever-more-expensive HPC systems.
But don't freak out just yet. The pause may simply reflect roadmap timing. Nvidia is rolling out its "Rubin" GR200 GPUs and "Vera" CV100 CPUs, and AMD is prepping its "Altair" MI400 GPUs and "Venice" Epyc 9006 processors — iron being installed now or early in 2027 that sites may not have been ready to benchmark for the November list.
Networking matters too. The industry sits stalled at 800 Gb/sec ports while the world expects 1.6 Tb/sec ports, and that transition, more than any cloud migration, likely explains the timing gap.
What the list's caretakers say
"The result does not indicate that performance has stopped growing," write the Top500's caretakers — Erich Strohmaier of Lawrence Berkeley National Lab, Jack Dongarra of the University of Tennessee and Oak Ridge National Lab, and Horst Simon of the ADIA Lab in Abu Dhabi. "Instead, it shows that a larger share of improvement is coming from additions and upgrades below the very top rather than from wholesale replacement of leading installations."
Where is Intel in all this?
Intel has no HPC-focused GPU on the market, only parts aimed at inference. Its "Diamond Rapids" Xeon 7 with 256 P-cores — reportedly debuting in the second half of 2027 on the Intel 18A-P process (1.8 nanometers), which would put it about a year late — is sampling now, per Intel's third-quarter Wall Street analyst call. The mainstream eight-memory-channel Diamond Rapids has been killed; the existing "Granite Rapids" Xeon 6+ fills the gap until a future Xeon 8 refresh.
One cloud cluster distorts everything
Across the 27 new machines, total cores — CPU cores plus GPU streaming multiprocessors — hit 71.4 million. But one system dominates: the eSurfing Cloud, a China Telecom IaaS offering in Shanghai. Built from nodes pairing two 56-core Hygon C86 CPUs (derived from a Zen1 core licensed from AMD years ago) with eight MetaX C588 GPUs, it totals 66.63 million cores and SMs. That's 93.3 percent of all cores across the 27 newcomers — yet only 38.66 petaflops peak and 18.36 petaflops sustained on HPL, or 2.9 percent of the group's aggregate Rpeak.
The most powerful new entry is the "Hangang" system at South Korea's KISTI, built from HPE Cray EX254n nodes with Nvidia Grace-Hopper superchips and Slingshot 400 interconnect, rated at 588.3 petaflops peak. It was one of five Grace-Hopper or Grace-Blackwell systems among the newcomers, which together supplied 51.8 percent of new capacity — 682.5 petaflops aggregate. Second largest is "Lux" at Oak Ridge National Laboratory, made from HPE ProLiant XD685 nodes combining "Turin" Epyc 9575F CPUs with "Antares" MI355X accelerators, rated at 323.9 petaflops. Everything else is miniscule; all 27 machines total just 1.32 exaflops of Rpeak.
How much of the list is accelerated?
Of the 500 systems, 290 carry GPUs or other accelerators. That's 58 percent of machines, 70.4 percent of total cores, and 79.9 percent of total peak exaflops — 22.64 exaflops of FP64 versus 5.7 exaflops for CPU-only machines. Nvidia holds the lion's share of systems, but AMD GPUs power many very large machines, claiming 37.7 percent of FP64 flops against 52.9 percent for Nvidia-based systems.
A caveat the list always carries: it captures only machines whose owners submit HPL results, not the world's roughly 3,000 proper HPC sites, many running multiple machines — and that figure excludes AI training and inference systems. Old iron lingers on the list, crowding out newer entries that deserve the visibility.
Expect roughly two dozen extra installs to backfill the November 2026 shortfall across the June 2027, November 2027, and June 2028 lists — and watch how AMD's MI430X, with true high FP64 performance, fares against Nvidia's software-emulated FP64 atop matrix engines as shipments ramp over the coming 18 months.
Source: The Next Platform
More from Rebecca Stone
Show full bio
Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,036 articles
Related articles
global-pc-shipments-fell-21-percent-in-q3-as-memory-prices-bit
Global PC Shipments Fell 21 Percent in Q3 2026 as Memory Prices Bit
chip-stocks-peak-every-two-years-and-fits-the-pattern
Chip Stocks Peak Every Two Years, and 2026 Fits the Pattern
hpe-launches-proliant-gen13-servers-built-around-amd-epyc-venice-1f91a485
HPE Launches ProLiant Gen13 Servers Built Around AMD Epyc 'Venice'
server-shipments-forecast-to-rise-5-5-as-arm-gains-ground
Server shipments forecast to rise 5.5% as Arm gains ground



