Happy Wednesday. I scan more than a hundred Chinese-language sources every morning, the newsletters and trade press and market wires that shape how China talks about its own AI industry, and I write up the parts that never make it into English. Let's go.
The Convert
For three years, the story about DeepSeek was that it did not want your money. Liang Wenfeng said so directly and often. The company would not raise outside capital, would not commercialize, would not do the roadshow circuit. It would build frontier models and give the weights away, funded out of the trading profits of his quant fund, and it would spend its attention on research rather than on the capital markets. That was the whole identity. It was most of why people liked it.
That identity is now gone, and it lasted about six weeks from one close to the next. In early June, DeepSeek closed the first outside funding round in its history, more than 50 billion yuan, the largest single round any Chinese AI company has ever raised, with Liang himself putting in 20 billion yuan to anchor it. Six weeks later, per the Financial Times and Bloomberg, it is already raising a second one. The FT and Bloomberg reporting puts the new round at roughly 71 billion dollars before the money goes in, up about 37 percent from the 52 billion the company was worth in June. Reuters has a larger version, a raise of as much as 50 billion yuan at a valuation near 500 billion yuan, roughly 74 billion dollars. Either way the direction is the same, up sharply in six weeks. DeepSeek has also hired an investment bank to prepare a mainland listing, with a filing on Shanghai's STAR Market possible by the end of this year and a 2027 IPO in view, per the same reporting.
Here is the number that reframes all of it. According to the Bloomberg Billionaires Index, Liang's personal net worth rose by about 19 billion dollars in a single day, to 36 billion, which makes him the richest AI-large-model founder in the world. The diversified giants at Alibaba and Tencent are worth far more, so this is a narrow ranking, people whose fortune comes mainly from building models. On that list he now sits first, ahead of Anthropic's Dario Amodei and OpenAI's Greg Brockman. The company he controls charges nothing for its best product.
The reason for the reversal is not a change of heart about capitalism. It is compute. DeepSeek's own hiring tells the story more plainly than any press release. The company has posted a job for an "IDC design and planning engineer," and the description says it is building a gigawatt-scale data center from site selection through construction. It is hiring people to write training and inference frameworks, to squeeze every last bit of throughput out of a GPU, to keep thousands of accelerators working in sync. Reuters has separately reported that DeepSeek is quietly designing its own inference chip, talking to chip-design firms, foundries, and memory suppliers. None of that is free. The competition last year was about whose model was smartest. This year it is about who can afford the machines to keep one running, and a lab funded by a quant fund's trading profits cannot win that race on its own balance sheet. According to The Information, Liang started planning the second round before the first had even closed.
The terms tell you who the money is meant to come from. The first round was strict, RMB only, Chinese institutions only, most of them routed through a limited partnership that Liang controlled. The second round loosens all of that, according to The Information's reporting as relayed in the Chinese trade press. Investors can put money in directly this time, or through the special-purpose vehicles that are standard in venture deals, and there is a QFLP channel built to accept foreign dollars, with Middle East capital named as a target. A company that a month ago would only take patient RMB is now building the plumbing to take Gulf sovereign money. That is a very different DeepSeek from the one people fell in love with, and the thing that made it distinctive is the thing it is trading away.
The Briefing
Liang sits at the top of a billionaire list that got very long very fast. The trade outlet Zhidongxi counted the AI-model gold rush and found at least 36 people whose net worth now tops a billion dollars purely from large-model companies, holding a combined 249 billion dollars, from just 17 firms. Three are from Chinese labs. Zhipu's chairman Liu Debing sits at 24.1 billion dollars, second on the whole list behind Brockman, and its founder Tang Jie at 5.2 billion; MiniMax's Yan Junjie is on it too. Zhipu is the tell here, its Hong Kong shares are up more than 1,000 percent since it listed in January, and its market value now runs into the hundreds of billions of yuan. The public market is willing to pay these numbers, which is a large part of why DeepSeek stopped waiting.
Xi Jinping will open the World AI Conference in Shanghai on Friday, his first time attending in person. The Foreign Ministry confirmed he will deliver the keynote, a change from prior years when the premier handled the opening and Xi sent a letter. The signal is deliberate. AI has moved up to the level of subject the top leader shows up for, alongside a high-level meeting on global AI governance, 140-plus forums, and more than 300 products debuting. The timing lines up with the money. The state AI industrial fund is already a DeepSeek shareholder from the first round, and now the head of state is opening the sector's marquee event in person.
China is preparing to let its top AI firms buy a limited number of Nvidia H200 chips. The Information reports the approvals are meant to offset a domestic shortage driven by surging AI-chip demand, a small, controlled crack in a door that Beijing spent two years framing as a matter of self-reliance. It sits oddly next to the DeepSeek own-chip effort and next to the "silicon curtain" talk from last week, where China was weighing controls on its own AI exports. The country is short enough on compute that it will quietly buy the American chip and build its own at the same time.
MiniMax is building a 2.7-trillion-parameter model, the largest any Chinese lab has put on the market. Two people told The Information about the plan. The number matters less than the direction. A model that size is a compute commitment before it is a capability, the kind of bet that only makes sense if you have the money and the machines lined up, which loops directly back to the fundraising wave. When your rivals are raising tens of billions of yuan to build gigawatt data centers, you raise too or you let the compute gap harden into a capability gap.
The Bigger Picture
DeepSeek spent three years as the counterexample. It was the lab that proved you did not need to play the capital game, that a small team funded by trading profits could reach the frontier and give the results away. That was a genuinely different model of how to build an AI company, and a lot of people, including me, found it more interesting than the American approach of raising ever-larger rounds against ever-larger compute bills.
What this week shows is that the counterexample could not hold, and it could not hold for a specific and unglamorous reason. Agents run inference continuously. Long context eats memory. Every promising new direction in AI consumes more compute than the last, and compute is bought with capital. Once the competition moved from who has the best idea to who can afford the most machines, the lab that refused capital was choosing to lose. The Information reports that Liang shifted after seeing Anthropic's progress in April and grasping what scale of compute it implied.
So the honest read is not that DeepSeek sold out. It is that the free-model, no-capital posture was always subsidized by a specific moment, the moment when cleverness could substitute for scale. That moment is ending. The revenue is real but small, 400 to 500 million dollars annualized against a valuation of 71 billion, which works out to about 148 times sales, a multiple that only survives if the growth story stays intact. The state fund is in the cap table. The Middle East money is being courted. The one thing still separating DeepSeek from every other large-model lab running on state and sovereign capital is that it gives its weights away. Whether that survives an IPO and a foreign-capital cap table is the open question, and DeepSeek has not answered it.
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.