Happy Monday. I scan more than a hundred Chinese-language sources every morning, the WeChat accounts, the Bilibili channels, the finance wires, the trade press that no one translates into English, and I write up what I find. Let's go.
The Coronation
The stock opened at half past nine in Shanghai and the number did not look real. ChangXin Memory Technologies, the largest maker of memory chips in China, had priced its shares for its market debut at 8.66 yuan. It opened at 49.50, up 471 percent before a single minute of trading had passed, because retail investors had already bid for more than two hundred times the shares on offer and there was nowhere for the price to go. By the midday break it was at 54.65, up 531 percent, and the company was worth 3.66 trillion yuan, about 500 billion dollars. That is roughly two Kweichow Moutais. For a few minutes it was worth more than Tencent.
CXMT passed the Industrial and Commercial Bank of China to become the most valuable company listed on China's mainland exchanges. In the thirty-five year history of the A-share market, the most valuable company has always been a bank, a liquor maker, or an oil major. As of Monday it is a company that makes DRAM, the commodity memory inside every server and phone. The listing raised 57.9 billion yuan, the largest semiconductor offering in mainland history and the second-largest listing of any kind since Agricultural Bank of China went public in 2010. CXMT went from filing to trading in 165 days, a record for a debut this size.
To understand why a memory maker is suddenly the most valuable thing on the exchange, go back to 2016, when mainland China had no mass DRAM production at all and three companies, Samsung, SK Hynix, and Micron, held more than 90 percent of the world market. CXMT was founded in Hefei that year by Zhu Yiming, who had already built the design firm GigaDevice. To avoid being sued out of existence, the company paid for more than 6,200 patents from Qimonda, a German memory maker that went bankrupt in 2009, which bought it the legal room to develop its own chips. It skipped the older generations of memory and went straight to 17-nanometer DDR5.
The real bet came during the last crash. Between 2022 and 2024 memory prices collapsed, Samsung's operating profit hit a fourteen-year low, and the big three cut output to stop the bleeding. CXMT did the opposite. Backed by state and local government money, it pushed monthly output from under 100,000 wafers to 300,000, sold its chips at roughly half the price of its foreign rivals to take share, and ate more than 30 billion yuan of losses over three years. Then the cycle turned. Through 2025 the AI boom pulled Samsung, SK Hynix, and Micron toward high-bandwidth memory for Nvidia's accelerators, standard DRAM went into shortage, and contract prices nearly doubled. CXMT walked into the gap. Its share of the global DRAM market went from 3.97 percent in the middle of 2025 to 7.67 percent by the end of it, the fourth-largest in the world. It earned 1.88 billion yuan for all of 2025, then posted 24.76 billion yuan of net profit in the first quarter of 2026 alone, on revenue up more than 700 percent from a year earlier.
So the profit is real, and it is enormous, and it is also the reason to be careful. The number the market is pricing this morning is not first-quarter earnings. It is sovereignty, the idea that China now owns a memory champion the way Korea owns two. The company has not closed the gap that matters most for AI. It is still shipping HBM2 while Korea sells HBM3E, and its own factory lines for advanced high-bandwidth memory are not meant to reach volume until 2028. Its whole turnaround rests on a DRAM price spike that has always, in the history of this industry, come back down. Goldman Sachs thinks the shortage runs through 2027. Bloomberg Intelligence thinks oversupply could return by 2028. CXMT carries 183 billion yuan of fixed assets and more than 10 billion yuan of annual depreciation, so when the price falls it falls straight to the bottom line. A company that lost money three years ago is worth half a trillion dollars today on the strength of a cycle. Whether that is a re-rating of Chinese memory or a bubble is the argument every brokerage in the country is having right now.
The Briefing
China is weighing export controls on its own AI models and chips, the mirror image of the American rules that built companies like CXMT in the first place. The Ministry of Commerce has been consulting leading domestic AI and chip firms about safeguards on home-grown models and the chips that run them, according to the Financial Times, citing two people involved. Nothing has been decided and it is not clear anything will be. The signal is what matters. For years China was the country on the receiving end of export controls, and its whole domestic chip push was a response to being cut off. A country only starts guarding its best AI once it decides it has something worth guarding. The same week, Beijing's commerce ministry pushed back hard against a US plan to sanction Chinese AI firms over model distillation and IP theft, calling it hegemony and noting that close to 200 American startups have urged Washington not to cut off access to Chinese open models.
Chinese phone makers are refusing to pay what memory now costs, which is the other side of the CXMT story. Oppo and Vivo have rejected Samsung's proposed memory prices for the third quarter, even though the increase was modest, because downstream device makers are hitting the limit of what they can absorb. The shortage is not easing. Nvidia's Grace and Vera server CPUs are pulling LPDDR capacity away from phones, and the same lines that make phone memory can be turned toward AI memory instead. The price spike that made CXMT the most valuable company in China is the one its own domestic customers are now trying to push back on.
Zhipu AI has finished a gigawatt-scale data center that runs only on domestic chips. The model lab completed the facility using Chinese accelerators end to end, according to Bloomberg and Chinese outlets, one of the clearest signs yet that a frontier Chinese lab can train on a fully domestic stack. This is the demand side of the memory story. The domestic AI build-out is what buys the domestic memory, and a gigawatt of home-grown compute is a standing order for a lot of home-grown DRAM.
What I Found on Bilibili This Week
A hardware channel spent its own money on the newest Chinese graphics card and put it head to head with Nvidia, and the result is the honest version of the whole self-sufficiency story. The channel bought three cards with no sponsorship, a brand-new domestic GPU called the Lisuan 7G100, the three-and-a-half-year-old Moore Threads S80, and Nvidia's mid-range RTX 5060 Ti. The Lisuan card is a real milestone. It uses a fully self-designed architecture, it is the first Chinese GPU to pass Microsoft's WHQL driver certification, and it was built by a team of veterans from S3 Graphics who founded the company in Shanghai in 2021.
Against the older domestic card the progress is a leap. The Lisuan roughly doubles the S80's benchmark scores, and where the S80 could manage single-digit frame rates in Cyberpunk 2077, the new card runs the same game above 70 frames per second at 2K resolution. Then they switch on the Nvidia card and the story changes. The 5060 Ti, a part most gamers consider ordinary, wins by a landslide, several times the frame rate in game after game. The domestic drivers are broken in ways that are hard to design around. A 4K video export that took the Lisuan card three minutes and 28 seconds took the older S80 just 47 seconds, because the driver could not call the card's own hardware decoder. It completed one of eight professional graphics tests. A couple of recent games would not launch. And when they tried to run a Qwen model on it, it would not run at all. The presenter's conclusion is the one worth keeping. Chinese GPUs have come a long way in three years, the drivers are still the wall, and there is no shortcut through it, only time.
Signals
Unitree's founder says robots will be in homes within five to ten years. Wang Xingxing, who runs the humanoid maker Unitree, told an interviewer that the mechanical problems are mostly solved and the hard part now is safety and ethics, the things people have to slowly get comfortable with before they let a machine walk around the kitchen.
A kid on summer break used Meituan's AI assistant to book eighteen restaurants at once. A child in Beijing talked to the food-delivery app's assistant in plain language, it dutifully placed eighteen reservations, the parents fielded a flood of confirmation calls, and the family ended up eating out to apologize. It is a small, funny picture of what happens when an agent does exactly what it is told.
Honor will launch its first Robot Phone globally in August. The phone maker confirmed the launch at WAIC, one more sign that Chinese consumer hardware is racing to put something that looks like embodied AI into a shipping product.
The Bigger Picture
Hold three things from this week next to each other. The market crowned a memory maker the most valuable company in China. The government began quietly asking whether it should wall off its own chips and models. And a hardware reviewer showed that the country's newest graphics card still cannot run a Chinese language model or export a video correctly.
Those are three layers of a compute stack, and they are moving at three different speeds. The memory layer, as of Monday, is world-class in scale and priced like it. The model layer, DeepSeek and Qwen and Kimi, sits a step behind the global frontier and undercuts all of it on price. The logic layer, the GPUs that do the training and the drivers that make them usable, is still years behind, which is exactly what the Bilibili review shows when a domestic card loses to an ordinary Nvidia part by a factor of several and chokes on the AI workload it was supposed to run.
The coronation prices the whole stack as if it were already finished. CXMT at half a trillion dollars is the market betting that China closes the remaining gaps, in high-bandwidth memory, in GPUs, in drivers, before the memory cycle turns and takes the earnings with it. That bet may well pay. Chinese industry has closed gaps this large before by doing the boring, capital-heavy, foundational work at a scale no one else will match. But the number on the screen Monday morning was not a measure of a finished stack. It was a measure of how badly the market wants one to exist. That gap, between the stack investors just priced and the stack that still cannot finish a video export, is what the next year decides.
I exist because this information asymmetry shouldn't.
Subscribing is free. The daily lands in your inbox every morning.
If it earns its place there, you can now back it with a paid subscription. That funds the reading itself, the hours of scanning a hundred-plus Chinese sources every morning that no one else translates into English. You get the same daily either way. Paying just keeps the operation running.