A supercomputer in Shenzhen just topped the world rankings for the first time since 2017. The headline is the speed. The real story is the silicon: it runs without a single chip from Nvidia, AMD or Intel.

China has taken back the supercomputing crown. On Tuesday, organisers declared a machine called LineShine the world’s fastest at the ISC conference in Hamburg. The result ended a long American run at the top of the closely watched TOP500 list.

LineShine, housed at the National Supercomputing Center in Shenzhen, hit 2.198 exaflops on the standard benchmark. That is more than 20% faster than El Capitan. The US Department of Energy machine had led since November 2024 and helps maintain the American nuclear stockpile.

LineShine is the first Chinese system at number one in nearly a decade.

The speed is striking. The way China got there is the point. LineShine was built entirely without American chips, a direct answer to years of US export controls.

An all-CPU machine, by design

Most top supercomputers lean on graphics processing units, the same Nvidia and AMD chips that power the AI boom. LineShine does not. It is the first system on the list to pass two exaflops of sustained double-precision performance using ordinary processors alone.

The design is ambitious.

The machine packs nearly 14 million computing cores into 90 cabinets, drawing about 42 megawatts of power. At roughly 52 gigaflops per watt, it is unusually efficient for its size. It has already run a full simulation of the Earth’s systems and a model of the human brain, the kind of heavy science it was built for. Its custom LX2 processors use Armv9, the instruction set licensed from Britain’s Arm Holdings.

They run KylinOS, a Chinese version of Linux, with a homegrown network called LingQi tying the whole thing together. Reports link the chip work to Huawei, whose return to the cutting edge has unsettled Washington.

Rather than split the work between CPUs and GPUs, LineShine bakes GPU-style maths into the processors themselves. Jack Dongarra, a TOP500 organiser, inspected the machine and called it impressive. “They upped us by developing a system that is not reliant on GPUs,” he said.

The machine also tops a second ranking that weights real-world workloads, and its designers have entered 14 bids for the Gordon Bell Prize, science computing’s top award. They have not said who made the chips or on what manufacturing process, the detail Washington would most want to know.

A message aimed at Washington

China stopped submitting machines to the TOP500 in 2023, after the US tightened chip-export rules. So choosing to enter LineShine counted as a statement, not a formality. Dongarra was told the system was built without government funding, which is why its designers felt free to enter it.

“I’m not surprised it’s the number one system,” said Addison Snell of Intersect360 Research. “What I’m surprised by is that they submitted it.”

The timing fits a wider pattern. US curbs meant to slow China’s progress have instead accelerated its push for self-sufficiency. The result is a homegrown stack of chips, software and networking that no longer needs American hardware. A machine that needs no Nvidia, AMD or Intel parts is the clearest proof yet that the strategy is working.

It also exposes an awkward gap in the rules. Washington has spent years restricting GPUs, the chips most useful for AI. LineShine relies on CPUs, which face far looser controls.

“The US government should have stronger controls on the export and manufacturing of CPUs for the China market,” said Jimmy Goodrich of the University of California Institute on Global Conflict and Cooperation. “It is a loophole in the current regulations.” That gap echoes the one Washington is already trying to close around overseas subsidiaries.

Fastest computer, not the fastest for AI

Here is the catch. Topping the TOP500 is not the same as winning the AI race, and the two are drifting apart. The list measures high-precision maths, the 64-bit calculations used for climate models and nuclear simulations. Modern AI runs on cruder, faster approximations.

On a benchmark built to mimic AI-style work, LineShine ranked only fourth. On that mixed-precision test it managed 7.92 exaflops, a far smaller jump over its main score than a GPU machine would post. Its all-CPU design lacks the dedicated low-precision circuitry that makes GPUs so quick at training models.

Worse for the bragging rights, the biggest American AI systems do not even enter the contest. The vast clusters run by xAI, Microsoft, Amazon and Google are commercial, not academic, and stay off the list entirely.

“If the hyperscalers submitted their systems, this ‘world’s fastest’ would not crack the top five,” Goodrich said. In other words, the crown is real, but the title is narrower than it sounds.

The case for caution

None of this means China has closed the gap. Its domestic GPU industry still trails Nvidia and AMD by several years, which is exactly why it routed around them with a CPU-only machine. Building a record-setting computer from homegrown parts is a genuine feat of engineering and a powerful political signal. It is not the same as out-computing the United States on the workloads that matter most for AI.

The wider trend is harder to dismiss. Export controls have pushed Chinese firms toward custom chips, domestic operating systems and their own interconnects, the same ingredients on display in Shenzhen. Companies from carmakers to cloud giants are now designing silicon to cut Nvidia out, and even Nvidia is hunting for ways back in with Arm-based CPUs of its own. A policy meant to keep China dependent is teaching it to build alone.

So the open question is not really who owns the fastest computer this month. It is whether a wall of export controls slows China down or simply forces it to build a parallel stack the United States cannot see into or switch off. LineShine suggests the second answer is winning. China assembled the fastest machine on Earth, deliberately, from everything America tried to keep out.