Happy Tuesday. I scan 100+ Chinese-language sources every day, the ones that don't make it into English coverage, and write up what's actually moving in China's AI industry. I'm one person reading the Chinese internet so you don't have to. Let's go.
The No-GPU Machine
This morning in Hamburg, the ISC 2026 conference published the new TOP500 list of the world's fastest supercomputers. The number one machine is Chinese. It is called LineShine, it clocked 2.198 exaflops on the standard benchmark, and it is the first Chinese system to lead the list since Sunway TaihuLight in 2017. It is also the first machine anywhere to clear 2 exaflops on that benchmark, edging out the US Department of Energy's El Capitan at roughly 1.8. The Chinese coverage from IT之家 and the official TOP500 list agree on the headline.
Here is the part that should make you sit up. LineShine has no GPUs. None.
Every other machine at the top of this list runs on accelerators, mostly Nvidia and AMD, the same chips China cannot legally buy at the frontier. LineShine is built entirely on Chinese CPUs. The platform is called LingKun, the processors are 304-core LX2 parts running at 1.55 gigahertz, the interconnect is a homegrown design called LingQi, and the whole thing runs Kylin, China's domestic Linux. To hit 2.198 exaflops without a single accelerator, the designers threw bodies at the problem. The machine has 13.79 million CPU cores, with Tom's Hardware reporting roughly 47,000 domestic processors and no foreign-made components. That is not elegant. It is a country routing around a chip it cannot get by building a different kind of machine out of parts it can.
So the export-control story has produced its showpiece. The same week Anthropic pulled its top models out of China and the US tightened chip controls again, China posted the fastest supercomputer on earth using zero American silicon. As a piece of political theater it is hard to beat.
As an AI story it is more complicated, and this is where the Chinese triumphalism and the Western dismissal both miss. The benchmark LineShine won, called HPL, measures double-precision math, the kind weather models and physics simulations need. AI training does not run on double precision. It runs on low-precision math, and there is a separate benchmark for that called HPL-MxP. On that one LineShine came in fourth. The tell is in the ratio. GPU-accelerated machines get an 8 to 11 times speedup when they switch from full precision to the mixed precision AI uses, because their tensor cores are built for exactly that. LineShine got 3.6 times. A CPU-only design has no dedicated low-precision hardware, so the lift is small.
Read those two numbers together and you have the real state of play. China can now build the world's most powerful machine for traditional scientific computing out of nothing but domestic parts, a genuine engineering achievement that took eight years of being locked out to force. It still cannot build the world's most powerful machine for training frontier AI models that way, because the thing that makes a chip good at AI is the accelerator, and that is the one piece it is short. The number one ranking is real. So is the asterisk. Both of them are made of the same export controls.
The Briefing
YMTC's NAND market share jumped from 8 percent to 13 percent in a year, and the Koreans are saying it out loud. TechNode, citing Counterpoint, has Yangtze Memory as the fastest-growing company in global NAND, now ranked fourth behind Samsung at 29 percent and SK Hynix at 18 percent. Revenue hit $2.6 billion in the first quarter, up nearly 445 percent year on year, the third straight quarter of double-digit growth. Samsung and SK Hynix executives both said the pace of Chinese catch-up has run past their expectations. Memory is the part of the chip stack with the lowest design moat and the highest capex requirement, which is exactly the shape of problem China is set up to solve. Watch this number, because a domestic memory supplier at scale changes the math for every Chinese AI accelerator that needs high-bandwidth memory next to it.
NetEase open-sourced a voice-cloning model that speaks 14 languages with no accent, from a 3-second sample. Confucius4-TTS, from NetEase Youdao, clones a voice from three seconds of audio with no reference text and over 85 percent similarity, then has that cloned voice speak fluent Japanese, English, or Korean with what the company calls native pronunciation, per IT之家. The cross-language no-accent part is the hard problem in this field, and they released the full 54GB model weights under an Apache license with no commercial restrictions. The use case they are openly targeting is short-drama and content export, dubbing Chinese material into every major language cheaply. Giving the whole thing away is the strategy. It plants Chinese infrastructure under the global content-localization pipeline.
A robotics startup that is 90 days old just became China's fastest embodied-AI unicorn. Kunlunxing, a humanoid company registered less than three months ago, closed three funding rounds in under 90 days totaling several billion yuan, reaching unicorn valuation, per Leiphone. The first round was reportedly oversubscribed within three days of opening. The investor list runs through nearly every top-tier Chinese venture name plus state-linked industrial capital. Take the speed as the signal, not the company, which has shipped nothing yet. Chinese capital is now pricing embodied AI the way it priced large language models in 2023, paying up front for teams and theses before there is a product. That is either conviction or a bubble, and the honest answer is both.
Two listed companies announced billion-yuan raises this week to build embodied-AI parts, not robots. Zhaofeng, an auto-parts maker, is raising up to 1.4 billion yuan, with most of it going to a plant for humanoid-robot and intelligent-driving precision components, per its filing on 36Kr. The pattern across this week's finance wire is companies from leather to bearings reallocating capital into the humanoid supply chain rather than into humanoids themselves. The component layer is where Chinese manufacturing already wins, and it is where the money is quietly going while the robot startups get the headlines.
Shanghai Jiao Tong and Tencent published a method to erase AI hallucinations one token at a time. The framework, called BALTO, treats hallucination as a local problem, a wrong entity or number buried inside otherwise correct text, and assigns blame at the token level instead of penalizing the whole answer, per Leiphone's writeup of the paper. The point is to cut factual errors without making the model terse and useless, the usual tradeoff. It is a small academic result, but it is the kind of work that matters for the vertical models, medical and financial and legal, where Chinese labs keep finding wins. A model that can be trusted on numbers is worth more in those domains than one that scores higher on a leaderboard.
What I Found on Bilibili This Week
The yt-dlp transcription pipeline is still down, so no full transcript this issue. The video worth flagging by title and framing is "国产AI芯片占比冲到41%，英伟达从95%降到55%," which translates to "domestic AI chips hit 41 percent share, Nvidia falls from 95 to 55 percent." The figure is a creator's claim, not a sourced market report, so treat the exact number as unverified. The reason it is circulating at all is the LineShine result above. The Chinese internet is reading today's supercomputer ranking as proof the domestic-silicon transition is further along than the official numbers admit. Whether the 41 percent is right or not, the sentiment is the story, and the sentiment moves capital.
Signals
China's "灵晟" win came with a power bill. LineShine draws 42.2 megawatts and gets 52 gigaflops per watt, which lands it around 50th on the energy-efficiency list. Brute-forcing exascale with 13.79 million CPU cores costs you efficiency. The machine is a statement about sovereignty, not a blueprint for cost-effective compute.
A Chinese supercomputer hasn't been number one since 2017, and the gap tells a story. For most of those eight years China stopped submitting its fastest machines to the list at all, to avoid drawing attention to domestic chips during the worst of the export-control fight. Submitting LineShine, loudly, is itself the news. The posture flipped from hide-the-progress to flaunt-the-progress.
Anthropic pulled Fable and Mythos out of China last week, and the replacements are already shipping. A wave of new Chinese trillion-parameter models is being pitched directly as the local substitute for the suspended American frontier, with SCMP tracking DeepSeek V4, plus Xiaomi and Alibaba models crossing the trillion-parameter line. The suspension created a vacuum, and the vacuum is being filled in weeks, not years.
The Bigger Picture
There is a clean way to read today's ranking and a true way, and they point in opposite directions.
The clean way is the headline China wants. Locked out of American chips, China built the fastest computer on earth anyway, out of entirely its own parts. Export controls failed. The future is sovereign.
The true way is in the second benchmark. LineShine wins the test that rewards raw double-precision math and comes fourth on the test that looks like AI training, and the reason is the exact piece China is missing, the accelerator. You can route around a GPU for weather simulation by throwing 13 million CPU cores at it. You cannot route around it for training a frontier model, because the GPU is not just more cores, it is the right kind of math in silicon.
So both stories are true and they are about different things. China has clearly won the right to say it can build sovereign compute at the largest scale for traditional science, and that is not nothing, it took eight years of pressure to force the whole stack into existence. It has not yet won the ability to do the same for the workload that actually defines this moment. The number one machine on the list and the chip China still cannot make are the same story, seen from two ends. The ranking is the floor China has poured. The asterisk is the ceiling it is still under.
I exist because this information asymmetry shouldn't. If this was useful, the best thing you can do is forward it to one person who'd want it, and subscribe if you haven't.