China's homegrown LingSheng supercomputer at NSC Shenzhen now ranks as the fastest machine in the world, delivering 2.2 exaflops on the High Performance Linpack test against a peak theoretical performance of 2.74 exaflops. That HPL result makes it 21.5 percent faster than the former champion, the El Capitan system at Lawrence Livermore National Laboratory, which runs on AMD MI300A compute engines.

The machine is the second time China has skipped the accelerator model entirely. Its 13.79 million cores are driven by the LingKun LX2 CPU, an Armv9-compatible server processor with 304 active cores per chip, designed by NSC Shenzhen together with Huawei, most plausibly through its HiSilicon division. The nodes link through a proprietary interconnect called LingQi LQLink, which the source assesses as likely a variation of InfiniBand, though it could be a heavily reworked Ethernet.

Why does a CPU-only design win?

The LX2 carries enough FP64 throughput through its SVE2 vector units and SME matrix math units to reach exascale without offload engines. The trade is simple: the machine's practical limits become space and power rather than silicon density. China built its earlier OceanLight system at NSC Qingdao on the same philosophy, using the SW26010-Pro Sunway processor with 41.93 million cores to reach roughly 1.5 exaflops peak.

The source states the matter plainly: China has the money and the power to go big instead of going dense with accelerated designs. The country is also widely believed to hold more undeclared exascale systems. Two are known by name: the 1.5 exaflops OceanLight at NSC Qingdao and the 2.05 exaflops Tianhe-3 at NSC Guangzhou.

What did the June 2026 Top500 list add?

The June 2026 list brought 44 new machines totaling 5.3 exaflops of aggregate new capacity. LingSheng alone accounts for 51.6 percent of that figure. Among the rest, several patterns stand out:

• HPC centers continue to prefer Nvidia's Hopper-generation H100 and H200 GPUs over the follow-on Blackwell B200 and B300, citing lower price, higher FP64 throughput and better flops per dollar.
• Eleven new machines pair Intel Xeon CPUs with Nvidia GPUs; nine more pair AMD Epyc CPUs with Nvidia GPUs. Together these hybrids represent 15.3 percent of new installed flops.
• Three clusters do use Blackwell, though the most powerful new all-Nvidia machines still run Hopper.

The largest new commercial system is HPC7 at Italian oil and gas company Eni, built on AMD's MI300A hybrid CPU-GPU device and ranked sixth. The source calls it a chip off the El Capitan block, and notes it is the largest commercial supercomputer with submitted results, a deliberate hedge. Oil majors do not brag much about whatever larger machines they may run quietly.

CPU-only clusters persist as well. Five new AMD Epyc clusters add 0.8 percent of the new flops, and four Intel Xeon clusters add 1.8 percent. They are not taking over the world. They are not going away either.

Is AMD actually closing on Nvidia?

By machine count, Nvidia dominates the accelerated segment with 237 systems against AMD's 32. But count by capacity and the picture tightens considerably:

• AMD holds 8.18 exaflops of installed peak flops in accelerated machines; Nvidia holds 11 exaflops.
• By combined CPU and GPU concurrency, AMD machines reach 35.3 million cores against Nvidia's 38.9 million.

That is close enough to call a race, and the source suggests it may portend the shape of AI computing as well. The gap between 237 systems and near-parity in flops says something about system size: AMD's wins are large, dense, flagship-class installations, while Nvidia's reach spreads across a far broader installed base.

The Top500 site lists 274 machines with some form of acceleration, though its own text says 277. The source checked the math three times and stopped there.

What about the pace of progress?

The broader trend is less cheerful. Supercomputing has fallen off the Moore's Law wagon of doubling performance every two years, and the source attributes this to budgets rather than technology. This is a budgetary thing much more than a technology thing, it concludes.

Upgrade cycles still arrive in waves. June 2024 and November 2025 were weak for new capacity; June 2026 and November 2024 were strong, and both were dominated by exascale-class installations. Entry to the list now demands 2.66 petaflops of HPL performance, a bar the source calls frankly not that much given what a modern CPU or GPU can deliver.

One caveat governs all of this. The Top500 measures only systems that submit official HPL results, not the broader and often secretive HPC market. The source expects that hidden market to reflect the official list to some degree, which is the reason to keep reading it. With LingSheng now public and more Chinese exascale machines suspected in reserve, the competitive pressure on US accelerator roadmaps has, if anything, only just resumed.