China's consumers kept the trips and cut the tickets — and the ticket is what everyone sells
Shandong counted heads faster than money
The pattern held three days running. Shandong's 200 monitored scenic areas logged visitors up 5.2% on 1 October against revenue up 3.3%; on the 2nd, 16.3% against 13.5%; on the 3rd, 12.3% against 8.9%. Crowds beat receipts every single day, per TMTPost. The same gap ran through the first half: 78 monitored shopping districts, footfall up 5.1%, turnover up 2.8%.
Per-trip spending is now in its second year of decline — roughly 5.7% down in 2025, another 3.2% in the first half of 2026. Urban residents averaged about 1,128 yuan (~$159) a trip in 2024 and about 1,061 (~$149) in 2025.
Income is not the constraint; the balance sheet is. Per-capita disposable income grew 5.2% in the first half while consumption spending grew 3.7% — for urban households, 4.4% against 3.0%. New-home sales value fell 13.0% over January–August, and August resale prices dropped 4.9% in tier-two cities and 5.6% in tier-three. A few hundred yuan on the payslip does not register against five figures off the flat.
And the market supplied the downgrade. China had 7,188 registered museums at the end of 2025, over 91% of them free, taking 1.56bn visits; localities issued more than 10,000 cultural-tourism measures with over 2.6bn yuan (~$366m) of subsidy behind them. When the floor price of a day out approaches zero, budget stress never shows up as staying home. TMTPost's formulation: keep the frequency, cut the intensity.
The one line that went the other way is the one the state pays for. Commerce ministry data for 1–3 October has footfall up 3.4% and turnover up 5.3%, carried by 19.63bn yuan (~$2.76bn) of trade-in sales to 3.48m people — 46,000 cars, 1.51m appliances, 1.71m digital devices, via IT Home. The fourth tranche of 62.5bn yuan (~$8.8bn) has gone out, which means all 250bn yuan (~$35bn) budgeted for 2026 is now disbursed. Q4 gets nothing.
The box office tells it without subsidy. Over 700m yuan (~$99m) through four days, four consecutive days above 100m — against 1.835bn for the whole 2025 holiday window and 2.105bn in 2024, per IT Home. By my arithmetic that is about a third behind last year's daily pace.
Go deeper on this section: ClaudeChatGPT
Two Kaifeng parks, opposite margin curves
Kaifeng Millennium City Park is the pure form of the old model. The Henan site recreates a Northern Song painting of the city at full scale, and it is now seeking a Hong Kong listing. Revenue went 673.1m, 741.7m, 746.0m yuan (~$105m) across 2023–25 — flat. Profit fell from 285.3m to 212.4m yuan (~$30m). Gross margin went 67.6%, 62.3%, 51.8%.
Tickets and one water-stage show were 83.3% of 2025 revenue, with food, beverage and hotel running a 3.2% gross margin, per TMTPost. From May 2025 the park turned a single-day ticket into three days of unlimited re-entry to chase a second purchase. Chair Wang Shuang, who came out of property and has run the company since 2012, spent 23 rounds of A-share listing coaching over nearly six years before filing in Hong Kong in June.
A few miles away, the park that sells participation instead of seats grew. Wansuishan Wuxia City, a martial-arts site in the same city, took 24.52m visits in 2025, up 146.9%, on revenue of 1.27bn yuan (~$179m), up 136.5%, from roughly 80m a year in 2019–22. Its 1,500-plus actors roam with no fixed stage. Millennium City Park runs about 410 scheduled shows a day. Scheduled shows are the thing a cost-optimising visitor skips.
Go deeper on this section: ClaudeChatGPT
Huawei stopped competing on nanometres
The Mate 90 launched on 1 October with a redefined yardstick. Huawei calls it the tau law: stop shrinking transistors, shorten signal paths. The Kirin 9050 Pro stacks two active logic layers joined by 5m vertical bonds for 125TB/s of inter-die bandwidth, 31% device-level gain and critical-path latency cut over 30%, per TechNode. Leiphone's launch writeup puts HarmonyOS devices past 90m.
The bill is paid in wafers. LogicFolding needs more silicon area per chip and more process steps, which cuts yield — and Kirin draws on the same scarce domestic lines as Ascend. Consumer chief Richard Yu said days earlier that advanced capacity in China is "very limited" and that memory prices have added about $200 to average handset cost, per CGTN. Every Mate 90 is an AI accelerator not built.
Go deeper on this section: ClaudeChatGPT
Decision models commoditised in a fortnight
The category is three weeks old. TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, shipped Jev on 15 September: a model that writes nothing and only answers questions with pre-set options, returning structured choices with probabilities at $0.042 per million input tokens and free output. Vercel said 13% of its paying gateway users called it within 24 hours — twice any previous launch, per GeekPark.
Then everyone shipped one. OpenAI's Decisions API, Cloudflare's Clef, Amazon's Strands Decider, Shanghai AI Lab's Intern-Decision. On 30 September StartLux — a Shanghai company under five months old, run by Chen Danian, who co-founded the games group Shanda in 1999 and later built WiFi Master Key — released five sizes from 0.8B to 27B, built in three days through an AI-run research pipeline.
The claimed numbers, with the asterisk the Chinese press itself attached: 63.88 against Jev 1.13's 57.91 on Decision Index 0.2.1, ahead on 31 of 38 benchmarks, self-tested against a 28 September snapshot, per TMTPost. Cloudflare claimed the lead the next day. Latency is the real product: the 4B answers three questions in 26ms, and at Q4 it is 2.71GB with 98.3% decision agreement with full weights.
THE TELL: Clef, Strands Decider and APUS's version are all post-trained on Qwen. A category with a two-week moat, built on Alibaba's torso.
Go deeper on this section: ClaudeChatGPT
China Eastern called the police on its own passenger
A flight attendant was filmed kneeling in the aisle to apologise after a service trolley clipped a male passenger's elbow. In its second statement, on Saturday, China Eastern said the passenger's conduct amounted to disturbing order aboard a civil aircraft and severely infringed the employee's personal dignity, and reported him to police, per SCMP. A state carrier publicly choosing the crew over the customer is new.
Elsewhere the mood is less resolvable. A V2EX thread titled roughly "the programmer trade feels a bit dead" argues the squeeze has reached independent developers — app downloads falling for six months, income near gone, here. Sentiment, not data, but it is the second such thread this week.
And a heritage site had to deny a video that never happened. Clips showed conservators pulling pigeons and rotted leaves from the Leshan Giant Buddha's ears while a monkey stole a hard hat. The management committee says it is AI-generated, there is nothing in the ears, and there are no monkeys on that hill, per IT Home — which then lists the labelling rules and the detention penalties.
Go deeper on this section: ClaudeChatGPT
Washington puts anti-ship missiles on Yonaguni
The NMESIS launcher goes to Yonaguni this month, Japan's closest island to Taiwan, during biennial joint exercises — a first, per SCMP. Note the sequencing: the summit warmth and the missile battery are three weeks apart.
And the same summit reads as two different events. SCMP's opinion desk notes Chinese state media spared no superlative, playing up the reception Trump gave Xi. Taipei, meanwhile, is working out what it means that Trump called China a wartime American ally; the Flying Tigers were a shared US-Taiwan asset until last month, per SCMP. Warmth is not a neutral good if you are the third party.
Go deeper on this section: ClaudeChatGPT
Threads we are pulling
- The FDE job turned into a consulting job. Following Friday's note on Anthropic's $100m for 10,000 forward-deployed engineers: QbitAI interviewed six of them. Chinese postings pay 30,000–50,000 yuan (~$4,200–7,000) a month; 01.AI, Kai-Fu Lee's company, now puts five FDEs plus five back-end staff on a project and says mapping a client's ontology takes one to three months. A Stanford PhD student running dozens of solo projects found his bank client abandoned the report generator he built — the rules changed daily — so he taught staff to hand de-identified data to a general model instead, and moved to a retainer. Over 70% of the work is talking, per QbitAI.
- Half an ontology is worse than none. Yu Jinze, chief scientist for industrial intelligence at Haier's COSMOPlat platform, says in a QCon Shanghai preview that injecting entities and relations without metric definitions scored below injecting nothing — the model picks the wrong table with more confidence, per InfoQ.
- Kimi got inside OpenAI's procurement rail. Kimi K3 can now be called from OpenAI's Codex via Baseten, billed against an enterprise's existing OpenAI purchase commitment, with Amazon Bedrock adding it on a revenue-share — the first Chinese open-weight model in that channel, per TMTPost's weekly. The Oct 2 point stands and sharpened: the distribution now runs through American paperwork.
- DeepSeek Harness answered the Claude Code question. Following Friday's telemetry complaint: team member Cui Tianyi says the new Claude Code Mods compatibility layer exists to test whether what Claude Code offers plugin authors is a subset of Harness's own extension surface, and that open source was "our original intention, not forced", per IT Home. About 60% of Harness users run third-party plugins.
- Hainan widened the launch bottleneck. Following Thursday's item on Guowang's 12,992 filed satellites and 15 launched: phase two of the Wenchang commercial spaceport is substantially built and in full-system rehearsal, adding pads three and four on about 181 acres, taking site capacity past 60 launches a year, per IT Home.
Go deeper on this section: ClaudeChatGPT