The Allotment
The DeepSeek founder took the biggest private allocation in China's largest chip IPO.
Happy Friday. 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 Allotment
Nine and a half million people entered one lottery last week, and the man who won the biggest insider share of it runs China's most famous AI lab.
The lottery was for CXMT, the Beijing company that is China's largest maker of DRAM memory chips and its only real domestic answer to Samsung, SK Hynix, and Micron. CXMT is about to list on Shanghai's STAR Market, expected on July 27, in what will be one of the biggest Chinese chip IPOs since SMIC in 2020. When the subscription window opened on July 16, 9,428,800 retail households piled in, bidding for something like 817 billion shares against a tiny float. About 7.7 million of them drew a winning number.
The win-rate tells you how hot this was. It came in at 0.4714 percent, which sounds small until you learn it is the highest win-rate in STAR Market history and roughly 15 times the year's average of about three in ten thousand. The exchange had to shift 40 percent of the institutional block back to retail just to feed the demand. Nearly ten million ordinary investors wanted a piece of a domestic chipmaker badly enough to lock up around a million yuan of market value each for the chance at 500 shares.
Then there is the name in the institutional column. Liang Wenfeng, the founder of DeepSeek, took the single largest private-fund allocation in the deal. Through his two quant firms, High-Flyer and Zhejiang Jiuzhang, 194 of his funds were allotted about 175 million yuan of CXMT stock, more than any other private manager, ahead of the next two, both also quant shops. If CXMT lists and climbs to the 3-trillion-yuan valuation some brokers model, Liang's funds stand to make around 730 million yuan on the allocation alone.
Sit with that for a second. DeepSeek is the reason a lot of the world started paying attention to Chinese AI at all, and the models it ships run hot on exactly the kind of memory CXMT makes. The demand for AI compute is what has DRAM prices climbing and what has made a domestic memory champion worth a trillion yuan or more on paper. The founder who did as much as anyone to create that demand just took the biggest private slice of the company positioned to sell into it. He is early to his own story.
I want to be careful, because none of this money is real yet. The stock has not started trading, so every gain is conditional on a debut that could go either way. And the timing is genuinely awkward. Since early July, global memory stocks have sold off hard, SK Hynix and Samsung are down close to 40 percent from their highs, and A-share memory names like GigaDevice have fallen more than 50 percent. CXMT walks onto the exchange into that. The 1-to-4.25-trillion-yuan valuation range the brokers publish is a very wide guess, not a price.
What is not conditional is the appetite. A chip that was a national-security talking point two years ago is now the thing nine million retail investors and the country's sharpest quant fund are fighting to own. The scarcity story that repriced the AI labs downward this month, the one I wrote about when Zhipu lost 38 billion dollars in two days, runs the other direction here. When the model can be copied for free, the market is not sure what it is worth. When the memory it runs on is in short supply and made by one domestic champion, the market is very sure indeed.
The Briefing
A US government official publicly accused Moonshot of stealing from a US model, and offered no evidence. On July 22, Michael Kratsios, who directs the White House Office of Science and Technology Policy, posted on X that the US has "information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," and that Moonshot built "a sophisticated internal platform to conduct large scale distillation against U.S. models" while switching access methods to avoid detection. Distillation means training your model on another model's outputs, which usually violates the other lab's terms of service and is notoriously hard to prove. Kratsios disclosed no logs, no records, nothing beyond the assertion. Moonshot has not issued a company response, though one of its engineers pushed back on the math, noting that Anthropic's Fable went public on July 1 and K3 shipped on the 15th, a 15-day window he called implausibly short to have trained a 2.8-trillion-parameter model off it. This is not the first round of this fight. Anthropic accused Moonshot, DeepSeek, and MiniMax of the same thing back in February, and the 2025 accusation that DeepSeek distilled OpenAI never produced public proof either. The pattern is consistent. A US official or lab makes the charge, the Chinese company denies it, no forensic evidence appears, and the charge gets used to justify the next export control regardless. Worth reading as an allegation, not a finding.
A Chinese memory-chip maker most people have never heard of just guided to a 1,925 percent jump in profit. Puya Semiconductor, a mid-cap Shanghai-listed maker of NOR flash and small memory chips, told the market its first-half net profit will land around 825 million yuan, up about 1,925 percent from a year ago. The company pinned it directly on the AI-compute buildout tightening memory supply and lifting prices across the board. This is the supercycle showing up as an actual profit-and-loss line, not a forecast. The same shortage is stretching the tools that make the chips. The five equipment giants who hold 70 percent of the market, Applied Materials, ASML, Lam, TEL, and KLA, have seen lead-times run 1.5 to 2 times longer, so a tool that used to arrive in six months now takes a year or more, and Samsung and SK Hynix are front-loading orders to get in line.
Alibaba ran a 2.4-trillion-parameter model on its own chips, with no Nvidia in the loop. Alibaba Cloud said its Lingjun M890 superpod, built on the T-Head Zhenwu chip its in-house silicon team designed, became the first domestic system to serve a model above 2 trillion parameters, running the new Qwen3.8 with 64 chips linked at 800 gigabytes a second across 9 terabytes of memory. A few days earlier the same T-Head team said it had shipped 560,000 chips and then open-sourced its full software stack, the SAIL toolchain that sits where Nvidia's CUDA sits. Giving away the software layer is a direct move at the one moat everyone agrees Nvidia still has, the code that everybody already knows how to write for.
Signals
Beijing built a factory whose product is tokens. The city's economic-development zone said it finished the first "token factory," a compute installation with a daily capacity of 1.4 trillion tokens, and the same day four city bureaus issued the country's first provincial-level policy aimed at AI agents. The document explicitly promotes what it calls a "Token economy," hands out hundreds of millions of yuan a year in compute, model, and data vouchers, and stands up a program for one-person companies to build on top of it. Treating tokens as a metered industrial output, with a factory and a currency, is a distinctly Chinese way to frame the AI buildout.
An 18-month-old robotics startup raised 2 billion yuan before it has a product to sell. Lingchu Intelligence, founded in September 2024, closed a combined angel and Pre-A round of about 2 billion yuan, one of the largest early-stage rounds in Chinese embodied AI this year. It open-sourced a dexterous-hand training model it says beats a leading foreign benchmark. The size of a seed-stage check like this is its own signal about how much money is chasing the robotics story right now.
The Bigger Picture
Two things happened this week that look unrelated and are not. Nine million people fought over a memory-chip IPO, and a US official accused a Chinese lab of stealing an American model. Both are the outside world putting a price on how far China's AI has come.
The IPO prices it in the direction of belief. The reason CXMT can draw ten million retail bids and the sharpest quant fund in the country is that the market has decided a domestic memory champion is a scarce and valuable thing to own, precisely because the AI demand is real and the supply is short. The accusation prices it in the direction of threat. You do not accuse a lab of industrial-scale theft unless its output is good enough that theft is the explanation you reach for. Kratsios did not claim Kimi K3 was weak. He claimed it was so strong, so fast, that it must have been copied.
Those are the same fact seen from two seats. From the exchange floor in Shanghai, China's AI is a supply chain worth a trillion yuan and a stampede to buy into. From the White House, it is a capability advanced enough to be worth accusing. The models keep getting cheaper and more open, the chips that run them keep getting scarcer and more domestic, and the world keeps trying to figure out what all of it is worth, in money on one side and in suspicion on the other.
None of this makes Western headlines the way it should. All of it matters.
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.