China is closing ranks: Xi around one core, industry around DeepSeek
Xi turns the Long March into a loyalty lesson
Xi wants one core. Marking the Long March's 90th anniversary on Thursday, Xi said the Red Army survived by rallying around Mao Zedong and that China now faces a "new Long March". It came four days before the fifth plenum, the Central Committee's closed-door meeting, set for 26 to 29 October.
The line aimed inward. Seoul's Kyunghyang Shinmun notes that Xi blamed "leftist dogmatism" for nearly sinking the Party before Mao took command in 1935. That passage was absent from the 80th-anniversary speech a decade ago. The paper reads it as aimed at the Party itself. My read: divided leadership nearly ended it, and one core saved it.
Huang Qifan is the context. On 13 October the Central Commission for Discipline Inspection, the Party's anti-graft body, opened a probe into Huang, 74, Chongqing's mayor from 2010 to 2016. He had outlasted the falls of his superiors Bo Xilai and Sun Zhengcai.
The count is the point. SCMP's tally, relayed by ua.news, is that 35 of the Central Committee's 205 full members have been investigated or lost their legislative mandates. On 17 October I argued that surviving one purge protects nobody from the next. A retired mayor fits that.
THE TELL: state broadcaster CGTN's English summary stresses "loyalty, truth, justice and unity" and the road to a modern socialist China. In the version I read, it never mentions leftism or a political core. SCMP and the Korean paper lead on both. No mainland outlet in today's feed covered the speech.
The AI treaty reaches the legislature. The National People's Congress Standing Committee meets 2 to 6 November and will consider ratifying the agreement creating the World AI Cooperation Organization. Twenty-nine countries signed it in Shanghai on 16 July, Russia, Brazil and Cuba among them. It takes effect once a third state deposits ratification. Beijing is building its own AI institution while Washington sanctions its AI investors.
Go deeper on this section: ClaudeChatGPT
DeepSeek's investors are industrial China
Nobody is on the record. DeepSeek has raised about 100 billion yuan (~$14 billion), Bloomberg reported, which prints $15 billion. The target rose twice, from 50 billion yuan (~$7 billion) to 80 billion (~$11 billion), then settled higher. Reuters says the round is still closing and figures could change. No report carries a comment from DeepSeek.
Suppliers and customers, not funds. CATL, the battery giant, and Tencent are among the biggest investors. Quartz, citing the Wall Street Journal, adds carmaker Geely and Lenovo-backed Legend Capital as newcomers, with a valuation near $74 billion going in. Reuters says many investors are state-backed or corporate. My read: this is a national-champion round, not a venture bet.
CATL is paying while bleeding. On 20 October I flagged that CATL's profit lagged its sales. TMTPost reports the shares then fell over 12% on 21 October, the biggest one-day drop since listing, to 227.9 yuan (~$32). Third-quarter gross margin slid to 20.69%, and nine-month operating cash flow fell 23.6%. It has also put over 10 billion yuan (~$1.4 billion) into AI-energy plays.
THE NUMBER: $1 billion. That is DeepSeek's annualised revenue by 24 September, per China Daily's account of The Information's report, more than double a few months earlier. Liang Wenfeng, the former hedge-fund manager who founded DeepSeek, told investors price rises hadn't driven customers away. On my arithmetic, the valuation is about 74 times that. Pricing power is what the money is buying.
Chinese coverage is a rewrite. Zhidx, the tech site, relays Bloomberg as China's largest AI financing and ends by saying products and an IPO must prove it. It omits the state-backed investors Reuters flagged. It does add that Moonshot AI is reportedly eyeing a Hong Kong listing in early 2027, while DeepSeek has hired CITIC Securities for a Shanghai listing.
Go deeper on this section: ClaudeChatGPT
Narrow models, hard kernels, robot faith
Two decision models in one day. On Thursday Ant Group's enterprise-technology unit released LingDT-Decision-1.0 at 4 billion and 27 billion parameters. China Telecom's AI lab open-sourced Xing4.0-29B-Intent. Both pick from a fixed list of options instead of writing an answer, targeting high-volume routing that general models handle expensively.
Self-reported, but specific. Ant says token use falls over 90%, and customer-service risk detection rises from 76% to 89% versus Qwen-3.7-flash. China Telecom scores 58 against 49 for the best of 18 general models on 6,000 public-dataset questions, at under 200 milliseconds. Ant adds ordered decisions expressed as a dependency graph.
WHY IT MATTERS: generation is the wrong tool for routing. A refund, a renewal cancellation and an invoice request arrive in one sentence, and the answer set is closed. Ant concedes that spotting dependencies is not executing them: money and permissions still need business rules.
Qwen and Kimi share one attention recipe. At AICon Shenzhen a Qwen engineer laid out FlashQLA, a GPU kernel library for linear attention. Qwen runs three linear-attention layers to each softmax layer. Kimi K3 uses the same three-to-one split with its own variant.
The catch is occupancy. The recurrence allows one thread block per head, so a 32-head model leaves most of a GPU's 100-plus multiprocessors idle. The fix is sequence parallelism with a correction matrix, written as hand-built kernels because Triton can't schedule it.
Unitree's founder says he hasn't believed enough. Wang Xingxing spent 36 minutes at Unitree's Hangzhou conference urging the room to trust AI. Since August he has cut his estimate for robotics' ChatGPT moment from two-to-ten years to one-to-three. He showed GPT-6 driving a robot untrained. GeekPark's reading: if rivals supply the brains, the winner builds bodies cheapest. He wants 10 million units a year, not tens of thousands.
Belief carries a multiple. On 20 October Unitree itself warned of a trailing price-to-earnings ratio of 388. A chairman asking for more faith is also defending that number. That is my inference.
Go deeper on this section: ClaudeChatGPT
China's helium price crashed behind a wall
Helium fell 80% at home. Domestic high-purity helium peaked near 497.5 yuan (~$70) per cubic metre in April, after the Middle East conflict hit Qatar's output. It now sells for 61 to 105 yuan (~$9 to $15), while overseas prices still climb. Beijing banned helium exports on 10 July. Jiemian gives that equal billing with record Russian imports and new domestic plants.
Who it protects. China imported about 84% of its helium last year. My read: the wall insulates buyers who can't stockpile, since liquid helium boils away within roughly 45 days.
Go deeper on this section: ClaudeChatGPT
Threads we are pulling
- The export draft is a ledger. On 20 October I read Beijing's draft as a registry more than a gate. ChinaTalk's three-day reading agrees. It counts 40 additions across AI, solar and biomedicine, 14 banned and 26 licensed. Interim registration applies even to exports already under way, which it reads as Beijing building a dataset before deciding what to license.
- AI short drama got cheaper, not richer. On 20 October I said margin moves to rights and the shelf. GeekPark's interview with Huang Haonan, founder of studio Soy Sauce Culture, adds numbers. Cost per minute fell from 1,000 yuan (~$141) to 400 (~$56). Monthly output rose from 40–50 titles to 200–300, yet revenue stayed near 20 million yuan (~$2.8 million). Taste is a third owner: capacity scales, judgement doesn't.
- Brussels is not softening. SCMP reports that Brussels and Berlin feel emboldened and are drafting fresh restrictions, while Beijing bristles at the idea it was strong-armed. On 15 October I said planners would keep talking while tightening. The tightening is holding; the talking is the fragile part.
- Memory keeps climbing, on one poster's word. A V2EX user says a pair of 32GB sticks costs nearly 6,000 yuan (~$845), and a friend paid 7,100 (~$1,000) days earlier. That is one anecdote, not data, but it fits the shipment falls I have tracked since 14 October.
Go deeper on this section: ClaudeChatGPT