Happy Wednesday. 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 Repricing
On Friday and Monday, Zhipu lost about 38 billion dollars.
Zhipu is the Beijing lab behind the GLM models, and in January it became the first pure-play large-model company to list anywhere, trading in Hong Kong under the tag everyone gave it, the global first LLM stock. On July 17 its shares fell 28.49 percent. On July 20 they fell another 19.56 percent. Two sessions took more than 300 billion Hong Kong dollars off its market value, which is roughly 38 billion US, and left the whole company worth about 414.6 billion HKD, down around 70 percent from where it traded in June.
The thing that did it was a free model. Moonshot released Kimi K3 in the early hours of the 17th, 2.8 trillion parameters, open weights, and it took the top spot on the Frontend Code arena above Zhipu's own GLM. That is the same K3 launch I wrote about when Moonshot used it to paper a Hong Kong IPO. Read from Zhipu's side, the same event reads as a threat rather than a milestone, and the market read it that way in about 48 hours.
Here is why a competitor's open-weight release erases 38 billion dollars, because it is not really about which model scored higher. Zhipu's valuation, like every LLM lab's, was built on two stories. One is that its annual recurring revenue would keep compounding, the figure circulating on the 17th put it near a billion dollars, up from 250 million in March. The other is scarcity, that if you were a global fund and you wanted to own Chinese frontier AI on a public exchange, Zhipu was very nearly the only ticker you could buy. K3 put a crack in both. If the best model in a category can be downloaded for free and swapped in over a weekend, then a paying enterprise customer is a flight risk, and the ARR line stops being a trend and becomes a question. And the scarcity was already thinning, Zhipu's first lockup expired on July 8, MiniMax's the next day, and Moonshot is now papering its own listing, so the pool of Chinese-AI tickers is about to go from two to three.
The number I keep coming back to is the placement. On July 13, four days before the slide, Zhipu sold 19.78 million new shares at 1,588 Hong Kong dollars each and raised 31.4 billion HKD, one of the largest equity placements a Chinese tech company has done this year. By Monday's close the stock was at 890.5. The funds that bought that placement are down about 44 percent inside of a week. That is what a repricing feels like from the inside, the money that came in at the old valuation model is holding the loss while the market writes a new one.
I want to be careful about what this is and is not. The billion-dollar ARR figure that lit the fuse was never official, and a chunk of the two-day move is lockup mechanics and profit-taking piling on top of the K3 headline, not a clean verdict on Zhipu's business. Zhipu is still worth about 53 billion dollars and it is sitting on a placement that funds it into 2027. This is not a company failing. It is the first time a public market has had to price a listed AI lab through the arrival of a free competitor, and the answer it came back with is that the moat is narrower than the valuation assumed.
That answer travels. Every Chinese lab racing to Hong Kong, Moonshot next, then the queue behind it, is about to be priced by the same logic. The question the market asked Zhipu this week is the question all of them now have to answer, which is what you are actually selling when the weights are free.
The Briefing
The same week Zhipu was being repriced downward, a company that loses money on nearly every transaction filed to go public. SiliconFlow is an independent model-serving platform, the pipe that sits between the labs and the developers, and after 34 months it submitted a Hong Kong prospectus at a valuation estimated around 7 to 8 billion yuan. Its 2025 revenue grew 653 percent to 55.3 million yuan, and its gross margin was negative 24 percent. On its public-cloud line specifically the margin was negative 119 percent, which means for every yuan of revenue it spent 2.19 to earn it. The company has about 270 million yuan of cash and burns 14.8 million a month, so on current math it has roughly 18 months. The bet is that the market for AI plumbing grows so fast that a toll booth is worth owning even if it loses money on every car for now. It is the same capital story as the lead, one layer down and running in the other direction.
A home robot company raised nearly a billion yuan, its third round in six months. Futuring Robot closed a Pre-A of close to 1 billion yuan with strategic money in the mix, including the investment arm of Inovance, a hundred-billion-yuan industrial-automation firm that supplies robot joints and motors, and Nice Group, the laundry-detergent giant whose brands reach hundreds of millions of Chinese households. What makes this one worth flagging over the dozens of humanoid rounds is that the robots are already in 500 homes and have logged more than 50,000 hours of service. Of the nearly 50 billion yuan that poured into Chinese embodied AI in the first half of this year, almost all of it went to industrial lines and humanoid demos. Very little went to machines that actually live in a house, because the home is the hardest environment there is, no standard layout and a person moving your cup while the robot works. This is a round betting on the one company with real usage data in that setting.
Galbot showed up at the World AI Conference with a robot that had a customer instead of a dance. Where most of the show floor ran timed demos, Galbot's S1 ran all day and its dual arms held a sustained 50-kilogram payload, and the company says the robots are already working on CATL production lines moving battery modules and packs. CATL is the largest battery maker in the world. A humanoid robot doing real material handling for a customer of that size is a different kind of proof than a backflip, and it is the kind that starts to justify the valuations the sector has been raising at.
Signals
A voice-AI company said its agent business tripled. Unisound, listed in Hong Kong, told the exchange it expects first-half revenue of 530 to 580 million yuan, and that the slice powered by its large models, its agent business, grew 200 to 220 percent to 515 to 550 million yuan. It is a small number against the labs in today's lead, but it is a filed one, and it is a rare hard datapoint that agents are turning into revenue in China rather than just into announcements.
Another robot maker cleared its last gate to a listing. Shenzhen-based DOBOT, which builds robotic arms and embodied-AI systems, had its STAR Market IPO approved by the listing committee, the final review before registration. It joins the line of embodied-AI companies filing onto China's exchanges this year, the same pipeline that Unitree entered earlier this month, and the same one the money in today's briefing is racing to reach.
The Bigger Picture
The tidy version of Chinese AI capital this year has been a story about raising, round after round at bigger and bigger numbers, and I have written a lot of it. This week the story turned over and showed its other face.
Both sides ran at once. Zhipu lost 38 billion dollars in two days because a free model made a public market doubt the durability of its revenue. SiliconFlow filed to go public while losing more than a yuan on every yuan it earns. Futuring Robot pulled in a billion yuan for putting machines in 500 real homes. Galbot walked into WAIC with a robot that already works for the biggest battery maker on earth. The capital did not stop. It started sorting.
What it is sorting for is the thing a free model takes away, which is a defensible reason to be paid. A lab whose product can be downloaded and self-hosted has to prove the money follows the traffic anyway, the way Moonshot argues its API revenue does. A serving platform running at negative margins has to prove that scale eventually flips the sign. A robot company has to prove the machine survives a week in a stranger's kitchen. The rounds that are still landing are the ones that can point at a customer, a production line, a house, some fact on the ground that a competitor cannot copy over a weekend.
That is the read English coverage is going to miss while it argues about whether Kimi beats Claude. The leaderboard moved this week, and it took 38 billion dollars off a public company as it did. The market did not learn that Chinese models are weaker. It learned that they are so good, and so open, that being the company who made one is no longer, by itself, worth what everyone thought.
None of this makes Western headlines. 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.