Happy Sunday. 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 Listing
Western coverage this week has been about the model. Reuters called Kimi K3 the world's largest open model. The New York Times said it threatens America's lead. Fortune reached for a Pink Floyd joke and a phrase everyone else has used twice already, the second DeepSeek shock. All of that is true and all of it is the part you can already read in English.
Here is the part you cannot. This weekend, Bloomberg and Sina's market wire both reported that Moonshot AI, the company behind Kimi, has sent its shareholders a resolution backing a Hong Kong listing and told investors it wants to go public in as little as six months. The round it is now raising is being talked about at more than 30 billion dollars, up from the 20 billion it closed in May, and its president said the day after K3 launched that annual recurring revenue posted the largest single-day jump in the company's history.
Read those two facts next to each other. The model that had markets calling this a second DeepSeek shock, that on the day of its launch dragged TSMC down 7 percent and sent a listed Chinese rival down 30, is the same model Moonshot is using to walk into the Hong Kong capital market. The benchmark and the listing timeline moved in the same 48 hours. K3 took the top spot on Arena's Frontend Code leaderboard, above Claude Fable 5, and by that Friday the revenue line and the IPO clock had both jumped.
This is the pattern I keep flagging. DeepSeek reset the world's read on what a Chinese lab could ship, then raised at a stepped-up valuation. CXMT filed its 29.5 billion yuan STAR IPO the same month its memory chips drew the export-control spotlight. In both cases the technical proof point was what the company took to the underwriters. A frontier model in China now converts into a term sheet in a matter of weeks, and the venue is increasingly Hong Kong rather than a private round that quietly marks itself up.
The figure worth sitting with is the revenue. Moonshot's ARR is reported around 300 million dollars, up from 200 million in April, which is real growth and still small against a 30-billion-dollar target, on the order of a hundred times revenue. That is a bet on the slope of the line, not its level. More than 70 percent of that revenue is API, not consumer subscriptions, which matters because K3 is open-weight and anyone can download it. Moonshot president Zhang Yutong said open source is not only a race to the bottom on price, that of all the developer feedback on K3 the word that landed hardest was taste. The valuation is pricing the idea that a Chinese lab can give the weights away and still sell enough of the traffic around them for a public market to pay up.
I don't know whether the six-month timeline holds. Bloomberg's own read puts the realistic listing window in late 2026 or early 2027 once you count HKEX approval, and a shareholder resolution is not a prospectus. But the direction is not ambiguous. The strongest open model in the world right now came out of Beijing on Thursday, and by the weekend its maker was papering a public listing at a valuation half again what it raised at ten weeks ago. That sequence is the story English coverage is missing while it argues about the leaderboard.
The Briefing
The clearest confirmation of the Kimi thesis came from an American buyer, not a Chinese seller. Databricks is raising a new round led by Coatue at a 188 billion dollar valuation, up 40 percent from its December mark, with the raise reported around 3 billion dollars. Its CEO said adopting Chinese open-source models is a key part of controlling AI costs. That is the demand side of everything above. The reason a Beijing lab can price itself at a hundred times revenue is that a 188-billion-dollar American company is publicly saying it runs on the weights. The information asymmetry cuts both ways, and right now the buyers are acting on something the leaderboard debate is still arguing about.
Alibaba is about to open-source a 2.4-trillion-parameter model and is telling everyone it only loses to one system. Qwen 3.8 previewed this week with a claim that it may be the strongest model outside Fable 5, priced with daytime credits at a tenth of list and cheaper at night. Put it in the queue with the rest of the fortnight. Kimi K3 at 2.8 trillion parameters, DeepSeek V4's full version in gray release and expected imminently, and MiniMax's M3 Pro at a reported 2.5 to 3 trillion parameters and also slated to open. Four trillion-scale Chinese models in two weeks, most of them open-weight. The Chinese trade press has a word for it, 井喷, a blowout, the kind you get when a well finally comes in.
SenseTime is going to launch compute into orbit. The company told Caixin it plans to send up its first batch of compute satellites this year and to build five domestic clusters of more than ten thousand chips each, working with fifteen chip firms and aiming for a thousand-satellite space compute network with Guoxing Aerospace before 2030. Read past the science-fiction framing. This is a Chinese AI firm routing around the chip and power constraints on the ground by planning to put the compute somewhere the constraints do not apply yet, and lining up domestic silicon to do it.
A humanoid startup closed a Pre-IPO round and hit 400 million dollars raised in half a year. LimX Dynamics finished a Pre-IPO round of nearly 200 million dollars, its second raise of 2026, taking the six-month total to 400 million. The embodied-AI companies are not waiting for revenue to line up capital, they are lining up capital to reach the listing. It is the same move as the lead, one rung down the maturity ladder, and it is happening across the whole sector at once.
What I Found on Bilibili This Week
The transcription pipeline is still down, so I can't give you the full translated passages I'd prefer this week. But the titles moving through the Chinese video feed tell you where the argument sits. One creator's video is titled, roughly, Nvidia's moat is collapsing, but this time Jensen did it himself, a reaction to Nvidia's own moves rather than to a Chinese chip. Another runs the numbers on domestic AI chips reaching 41 percent share while Nvidia's China share falls from 95 percent toward 55. The framing on the Chinese side has shifted from can we catch up to how fast is the substitution, and that shift is the mood underneath every IPO filing in this issue.
Signals
CXMT's IPO lottery numbers are out. The 29.5 billion yuan STAR Market listing we covered at filing has published its allotment numbers, about 7.7 million, which puts the debut within days. The memory-supercycle thread runs underneath it, with SK Hynix's chairman saying this week that chip demand will grow at least 50 percent next year.
Xi Jinping attended the World AI Conference in person for the first time, and 29 countries signed on to a new China-led AI cooperation body. English coverage has this one well in hand, from the FT to SCMP to Foreign Affairs, so I'll only note the part they underplay, that the governance pitch and the capital push are running on the same calendar. The soft-power event and the money moves are one strategy.
Yuntian Lifei laid out a two-year inference-chip roadmap at WAIC, three chips tuned separately for the prefill, decode, and decode-FFN stages of inference, all designed for ten-thousand-chip mixed clusters to cut the cost of generating a token. The price war Kimi and Qwen are running on their pricing pages now has a hardware front too.
The Bigger Picture
The question I keep coming back to is what a public listing is actually pricing when the product is given away for free.
Moonshot is open-weight. Anyone can download K3. The thing a Hong Kong IPO would sell is not the model but the claim that a lab which gives its best work away can still capture enough of the value around it, the API traffic, the enterprise contracts, the position as the default open model that companies like Databricks build on. That is a real bet and an unproven one. Three hundred million in ARR against thirty billion in valuation is a number that only works if the line keeps bending the way it did the day K3 shipped.
What makes it plausible is the buyer signal. When an American data company raising at a 188-billion-dollar valuation says out loud that Chinese open models are how it controls costs, the Chinese labs stop being a subsidy story and start being infrastructure other companies pay to build on. The toll is the API traffic on top of weights that are free to download, and 70 percent of Moonshot's revenue already comes from exactly that.
So the week reads as one motion rather than several. A model that had markets calling it a second DeepSeek shock, a revenue line posting its best day ever, a listing timeline handed to shareholders, and a rival lab and a robotics startup papering their own rounds in the same stretch. Western coverage is treating the model as the news and the money as background. In China this month the order is reversed, and the number that will matter in six months is not the leaderboard rank but the price Hong Kong puts on a company that gives its best work away.
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.