EastVoice · AI · 2026-09-06 · Issue #059

Editor's Note: China's model race is now a pricing race with shopfronts. 智谱 / Zhipu opened a Tmall store selling Token packages, five more labs are lining up to do the same, and Beijing's economic zone put out the first Token-collateral bank loans, roughly ¥2B / ~$280M in initial credit lines. On the capital side, the national AI fund just wired ¥1.4B / ~$200M into Kuaishou's video-model spinoff Kling at a $18B post-money, with a hard IPO deadline attached.

Commentary: The interesting Chinese AI stories this week are balance sheets, not benchmarks. When state funds demand buyback clauses and labs sell inference like phone credit, the frontier is being reorganized around repayment schedules.

---

I. AI Foundation Models

1. Kimi, MiniMax and StepFun to open Tmall flagship stores, selling token plans like concert tickets

• Summary: Shanghai Securities News reported Sept 5 that Moonshot AI (月之暗面 / Kimi), MiniMax and StepFun (阶跃星辰) are all in talks to open official Tmall storefronts selling token subscription bundles. Zhipu (智谱) got there first, launching its own flagship store Sept 2.
• Source: ITHome · 2026-09-05
• Editor's Take: A frontier lab running a storefront on Singles' Day shopping infrastructure says everything about where Chinese model margins are heading: retail.

2. Saudi Arabia built its Arabic frontier model on Chinese open weights

• Summary: HUMAIN, backed by Saudi Arabia's PIF sovereign fund, released HUMAIN-M3 on Sept 3: 428B total parameters, 23B active per token, continued-trained on over 1 trillion Arabic tokens, built on MiniMax (MiniMax / 稀宇科技) open-source architecture.
• Source: TMTPost · 2026-09-05
• Editor's Take: Sovereign AI used to mean sovereign data plus American chips. The Saudi case shows a third variable: Chinese weights as the default substrate when the price is right.

3. Alibaba refreshed Qwen3.8-Max to the top of the WebDev leaderboard at $5 per million tokens

• Summary: Alibaba's September update to Qwen3.8-Max (2.4T parameters, 1M-token context) jumped 22 points to 1691 on CodeArena's front-end coding board, passing Claude Opus 5 and Kimi K3. On the same board's cost-performance frontier it averages $5 per million tokens, roughly a quarter of the price of the two models behind it.
• Source: TMTPost · 2026-09-05
• Editor's Take: Beating the frontier on a benchmark is routine for Alibaba now. Undercutting it 4-to-1 on price while doing so is the part that travels.

4. GPT-6 made "recurrent depth" fashionable; Alibaba had papers on it years ago

• Summary: After The Information reported GPT-6 Astra uses recurrent depth, running the same transformer layers repeatedly to buy effective depth without proportionally more compute, attention returned to two top-conference papers Alibaba (阿里巴巴) had already published on the technique.
• Source: QbitAI · 2026-09-05
• Editor's Take: Priority disputes are how AI research does its victory laps. The useful signal is architectural: depth is becoming a runtime dial, not a fixed spec.

5. Beijing's Yizhuang district issued China's first Token-denominated compute loans

• Summary: The Beijing Economic-Technological Development Area closed the capital's first batch of "YiQi Token loans," roughly ¥2B / ~$280M in initial credit lines from six banks, with token consumption records and compute purchase contracts serving as core credit evidence for asset-light AI startups.
• Source: ITHome · 2026-09-05
• Editor's Take: Banks underwriting against burn rates is normal startup finance, except token usage is a meter the lender can verify in real time. This is how compute becomes a utility-grade collateral class.

6. OpenAI now faces more than 50 lawsuits as ChatGPT enters the courthouse

• Summary: Per Futurism, OpenAI faces 30 new suits from victims of the Tambour Ridge, B.C. school shooting, alleging ChatGPT provided "substantial assistance and encouragement" to the gunman and that OpenAI failed to act on warning signs in his conversations. Total suits against the company now exceed 50.
• Source: ITHome · 2026-09-05
• Editor's Take: Chinese readers follow this story for the mirror image: China's AI liability rules were written precisely to avoid this decade of litigation. One regime sues, the other preempts, and enterprises will price the difference.

II. AI Software & Applications

1. WeChat is testing agent-to-agent chat: your AI texts your friend's AI first

• Summary: Leaked beta footage Sept 4 shows WeChat's assistant "Xiaowei" (小微) holding a separate conversation with another user's agent to exchange information, then summoning the humans only to confirm. This goes beyond dictation: the agent negotiates on your behalf before you see anything.
• Source: TMTPost · 2026-09-05
• Editor's Take: The world's biggest messaging app quietly redefining low-value socializing as machine-to-machine traffic with human approval gates. Every etiquette norm of group chat is now a design spec.

2. Tencent's WorkBuddy goes open: 100+ partners, 30 hardware brands on day one

• Summary: WorkBuddy launched its open platform Sept 2, bringing over 100 ecosystem partners, 30-plus hardware brands and 30 industry applications onto a single agent skills network spanning Skill, Connector and smart-hardware integrations.
• Source: TMTPost · 2026-09-05
• Editor's Take: Opening the platform cheapens the wrapper. Once anyone can bolt onto WorkBuddy, WorkBuddy itself becomes the commodity, and that is exactly what Tencent wants at the OS layer of Chinese agents.

3. The token math nobody publishes: WorkBuddy's 178-day bill

• Summary: A detailed analysis of WorkBuddy's first 178 days argues the scissors gap between token consumption and subscription revenue is the binding constraint on every Chinese agent platform, with open ecosystems driving usage up faster than willingness to pay.
• Source: TMTPost · 2026-09-05
• Editor's Take: Every agent platform is running a subsidized utility. The question the analysts keep asking is the right one: who pays when the inference bill outgrows the ARR curve?

4. Kling AI is worth almost as much as its parent company

• Summary: Kuaishou (快手, 01024.HK) trades at HK$140-150B while its video-model spinoff Kling, fresh off a $3B / ~¥20.4B round, carries a $18B / ~HK$140B post-money, roughly matching the whole listed parent. The national AI fund added ¥1.4B / ~$200M on Aug 31, and Tencent, Alibaba Cloud and Baidu all sat in the same round.
• Source: TMTPost · 2026-09-05
• Editor's Take: When the product line is worth the company, the parent is just a holding structure with an old app attached. Kling's shareholder agreement already says it: IPO by Oct 2031 or investors claw back at 8%.

5. HTC abandons the phone race and moves its budget to AI glasses

• Summary: HTC executive Vincent Huang (黄昭颖) said smartphones have run out of meaningful innovation room, with resources shifting to AI eyewear. VIVE Eagle added a round-frame style Sept 3, launching across Europe, North America and Australia with fashion positioning and on-device privacy claims.
• Source: ITHome · 2026-09-05
• Editor's Take: Coming from the company that lived through the Vive gamble, "we're done inventing on this device" is a data point, not a press release.

6. World-first: an AI ultrasound robot guides congenital heart defect closure

• Summary: The Sixth Medical Center of PLA General Hospital completed the first clinical case of an AI-guided ultrasound robot performing congenital heart defect closure on Sept 5, verified as a world first by an authoritative novelty search. The robot held imaging and instrument guidance through the whole minimally invasive procedure.
• Source: ITHome · 2026-09-05
• Editor's Take: Ultrasound is the specialty where sonographer shortage bites hardest, which makes it the most rational first job for a medical robot: steadier hands where there simply aren't enough hands.

7. GAC's 2021 chip bet pays off as CXMT goes public

• Summary: GAC Group (广汽集团) disclosed that its 2021 investment in memory maker CXMT (长鑫科技), routed through its GAC Capital arm, has delivered returns as CXMT's listing proceeds advanced, framing the deal as auto-scene capital buying upstream DRAM exposure.
• Source: ITHome · 2026-09-05
• Editor's Take: Chinese automakers have quietly become the country's best-positioned semiconductor LPs. The car company is now the patient fund with the guaranteed customer at the end.

III. Humanoid Robots

1. China's first wheeled-legged guide-dog robot debuts with 96% outdoor obstacle avoidance

• Summary: FarSight Mobility (远山智行), a spinoff of the China Ordnance Industries Group's Hangzhou Zhiyuan Research Institute, unveiled "Xiao Yuan" (小远) on Sept 4, the country's first wheel-leg hybrid guide robot for the visually impaired, rated at 96%+ outdoor obstacle avoidance.
• Source: ITHome · 2026-09-05
• Editor's Take: The wheel-leg hybrid skips the hard part nobody asked for, climbing stairs like a Labrador, to ship a mobility aid that works at real price and battery budgets. Assistive robots are the humanoid segment with genuine demand curves.

2. Ecovacs is industrializing the robot hit product

• Summary: At IFA Berlin, floor-care maker Ecovacs (科沃斯) laid out how it turns best-selling home robots into a repeatable assembly-line pipeline, targeting million-unit sales cycles per model rather than one-off flagships.
• Source: TMTPost · 2026-09-06
• Editor's Take: The humanoid industry keeps asking who can build a demo. Ecovacs answers a different question: who can manufacture a million identical machines that don't embarrass you in millionth-owner's living room.

3. Galbot's three gold medals came from the hardest event at the Robot Olympics

• Summary: Analysis of the second World Humanoid Robot Games argues Galaxy General (银河通用 / Galbot) won its three golds in the competition's least glamorous, most operationally demanding events, the ones that expose failure modes demo videos hide.
• Source: TMTPost · 2026-09-05
• Editor's Take: Sports events are now proxy procurement tests. Medal counts are becoming part of the RFP language for Chinese robot vendors.

4. Dreame brought 100+ products and 16 categories to IFA, winning 6 international awards

• Summary: Dreame (追觅) exhibited over 100 products across 16 categories at IFA 2026 on Sept 4, from robot vacuums to window-cleaning, lawn-mowing and hair-care devices, with six products taking international awards and dozens of claimed world-first technologies.
• Source: Leiphone · 2026-09-05
• Editor's Take: One Chinese startup showing up at Berlin with 16 product lines is the supply-chain thesis made flesh: the moat isn't any single robot, it's the factory that can spin up categories.

IV. Autonomous Driving

1. Chang'an's prediction: only five or six ADAS suppliers survive China's 2028 shakeout

• Summary: Chang'an Auto (长安汽车) VP He Gang told National Business Daily that China's smart-driving supplier field will finish its first consolidation wave around 2028, leaving perhaps five or six competitive solution providers. His pick for survivors: "We and Huawei's Yinwang will stay."
• Source: ITHome · 2026-09-05
• Editor's Take: A named survivor telling you the shakeout is coming is how incumbents freeze customer commitments before the consolidation they're describing happens.

2. BYD's Denza Z9S hits 334 dealerships before its first delivery

• Summary: BYD's premium Denza (腾势) brand announced Sept 5 that Z9S display cars are now in 334 stores across 166 cities, priced for pre-sale at ¥319,800-389,800 / ~$45,000-55,000, with the smart-driving-equipped sedan targeting the premium EV bracket.
• Source: ITHome · 2026-09-05
• Editor's Take: China's EV marketing has entered pre-exhibition warfare: cars meet customers in showrooms before regulations finalize the price. Distribution muscle is the spec sheet now.

V. World Models / Physical AI

1. The SpaceX veteran is now betting on a Chinese world-model startup

• Summary: QbitAI reports that a Silicon Valley investor known for backing SpaceX has invested in a Chinese world-model company whose pitch is AI that can "rehearse a downpour," simulating physical scenarios before robots or models act on them, attacking the gap between language fluency and physical understanding.
• Source: QbitAI · 2026-09-05
• Editor's Take: When the person who underwrote rocketry starts buying Chinese world models, the smart-money read is that physical simulation is the next bottleneck asset class.

2. A world model that quits after training makes robots better, not worse

• Summary: Researchers presented a counterintuitive world-model architecture: the model spends heavy compute guiding a robot during training, then is removed from the deployment loop entirely, leaving the policy it shaped running faster and more reliably on-robot.
• Source: QbitAI · 2026-09-05
• Editor's Take: The scaffolding analogy finally fits. The world model is a construction frame you tear down once the building stands. Deployment economics, not demo realism, is what this optimizes for.

---

物界前沿 | EastVoice