EastVoice · AI · 2026-11-09 · Issue #123

Editor's Note: China's robot builders spent the week shipping units instead of slides. Qiyuan's 10,000th personal robot left the line 14 days after its 5,000th, CNPC took 60% of a robot joint venture with Unitree, and GAC put 2029 on a 10,000-unit capacity target. The model crowd argued in public about where the moat sits, and the world-model camp got two more sceptical essays in a single day.

Commentary: A registry, a joint venture, a capacity target: Beijing is laying the industrial plumbing for robots before anyone has proven there is a market. That is how you end up with a supply chain that outlives the hype.

---

I. Humanoid Robots

1. CNPC takes 60% of a robot joint venture with Unitree

• Summary: China National Petroleum's Kunlun Manufacturing and Kunlun Capital signed with Unitree Robotics and Beijing Fanjie Intelligent to set up a joint venture for industrial embodied-AI robots. The split is 60/20/20, with the oil group holding control, and the stated reasoning is that acceptance testing for oilfield robots should not sit with an outside vendor.
• Source: TMTPost · 2026-11-08
• Editor's Take: An oil company buying control of a robot JV is a procurement story dressed as a technology bet. Whoever writes the acceptance test owns the specification, and in industrial robots the specification is the moat.

2. Qiyuan's 10,000th personal robot came 14 days after its 5,000th

• Summary: Shangwei New Materials' Qiyuan Robotics rolled out its 10,000th personal robot on November 8, up from unit 5,000 on October 25. Sales of the Q1 and T1 models began on September 20, which puts the entire ramp inside seven weeks.
• Source: ITHome · 2026-11-08
• Editor's Take: A 14-day doubling is a line-rate figure, not a demand figure. The question nobody in the release answers is whether those 10,000 units are in homes or in channel inventory.

3. GAC's humanoid partner targets 10,000 units a year by 2029

• Summary: Huilun Technology signed a production agreement with GAC's powertrain business unit on November 5, and the two are building a line aimed at 10,000 units of annual humanoid capacity by 2029.
• Source: ITHome · 2026-11-08
• Editor's Take: A car plant is the right place to learn how to build robots at cost, and the wrong place to learn what buyers want. The 2029 target also assumes the price war starts after the line is finished.

4. At CIIE the robots are doing shifts, and the gap is measured in millimetres

• Summary: Huxiu's dispatch from the CIIE floor has robots taking orders, sorting medicine, carrying parts and loading machine tools, with several deployments already live in stores and factories. The argument is that the "ChatGPT moment" for embodied AI now depends on millimetre-level repeatability rather than on demo capability.
• Source: Huxiu · 2026-11-08
• Editor's Take: Millimetres are where the unit economics live. A robot that misses one task in fifty needs a human standing by, and that human costs more per hour than the robot saves.

5. Unitree and Sequoia China start an early-stage robotics fund

• Summary: Unitree Robotics and Sequoia China launched the Shanyu Venture Program in late October, aimed at early-stage embodied AI, robot applications and upstream components. Several other Chinese robot makers have set up similar vehicles.
• Source: TMTPost · 2026-11-08
• Editor's Take: Robot companies moving into venture capital is what happens when your suppliers are also your competitors. The fund buys optionality on a supply chain none of them can build alone.

II. Autonomous Driving

1. Musk puts a 40 kWh pack in the Cybercab

• Summary: Musk said on X on November 8 that Tesla will push Cybercab efficiency further so the robotaxi can run a 40 kWh battery pack, and that new body colours are coming. He described it as part of a further round of improvements to the two-seater.
• Source: ITHome · 2026-11-08
• Editor's Take: A 40 kWh pack is a statement about duty cycle: short trips, depot charging, no long-haul anxiety. It also makes the vehicle cheaper to build than the sedan it replaces, which is the only reason a robotaxi fleet maths works.

2. Qijing's GX7 passes 4,000 deliveries

• Summary: Qijing held a group handover in Shenzhen on November 8, with cumulative deliveries of the GX7 past 4,000 units since launch. The car ships with Huawei's Qiankun ADS 5 driver-assistance stack and HarmonyOS Cockpit 6.
• Source: ITHome · 2026-11-08
• Editor's Take: Four thousand units is small for a car and large for a Huawei-badged experiment. The number worth watching is how many buyers paid for ADS 5 as a line item instead of taking the base trim.

3. Bosch's China ADAS chief wants the coffee ordering out of the car

• Summary: At a Wuzhen auto panel on November 5, Wu Yongqiao, who runs Bosch's China driver-assistance and control systems business, argued that the car will not become a "third space" and that the cockpit should stop duplicating what the phone already does well.
• Source: ITHome · 2026-11-08
• Editor's Take: A tier-one supplier telling OEMs to ship fewer features is either honesty or a negotiating position. It is also correct: the cockpit has spent five years adding things drivers use once.

4. Sony's see-through warning system renders what the driver cannot

• Summary: Sony China showed "Semantic Smart Travel" at CIIE on November 5, a V2X warning system that draws pedestrians and cyclists hidden behind buildings or the car ahead onto the cockpit screen, replacing the generic risk icon with a located object.
• Source: ITHome · 2026-11-08
• Editor's Take: Occlusion is the one perception problem cameras cannot solve alone, and V2X has been two years away for a decade. Sony's angle is the display layer, which ships now, rather than the network, which does not.

III. World Models & Physical AI

1. Beijing moves on the training data behind embodied AI

• Summary: The National Data Administration convened a meeting of "chain leaders" on high-quality industry datasets, framing the work around demand-driven supply, use-driven construction and quality assessment. A member of the national data expert committee raised the same theme at Wuzhen this week.
• Source: Huxiu · 2026-11-08
• Editor's Take: Robot data has no web to scrape. Every trajectory is filmed by a person with a rig, which is why the cost curve here looks like manufacturing and not like software.

2. If GPT-6 can drive a robot, who still pays for a world model?

• Summary: Huxiu's piece asks whether dedicated world models still have a buyer now that general models can control hardware, and collects the cases on both sides. The sceptics' line is that a general model with enough video will learn physics for free.
• Source: Huxiu · 2026-11-08
• Editor's Take: The world-model thesis was always a bet that general models would plateau. That bet is underwater, and the labs that raised on it now have to sell a much narrower product.

3. Capital is leaving world models

• Summary: QbitAI argues that the Tian Keyu startup controversy was a symptom rather than the cause, and that venture money had already begun rotating out of world models before the story broke. The category went from the year's hottest theme to a hard fundraising market.
• Source: QbitAI · 2026-11-08
• Editor's Take: Two essays in one day asking who pays for world models is not a coincidence. When a category needs a defence, the multiple is already gone.

4. Hitch Open goes open source, and a robot plays 142 strokes of table tennis

• Summary: The Hitch Open physical-intelligence platform, built on UC Berkeley research and launched out of Silicon Valley, made its global debut at CIIE's Lingang pavilion on November 7 and released its stack openly. A demonstration match set a rally record of 142 strokes between a robot and a human player.
• Source: ITHome · 2026-11-08
• Editor's Take: Table tennis is the right demo and the wrong product. It proves latency and control; it does not prove the arm can pick up a part that shifted two millimetres overnight.

5. Shanghai Jiao Tong's AutoDNA runs molecular biology without a bench chemist

• Summary: A team from SJTU's IPADS lab, working with the university's chemistry school and the Zhangjiang DNA storage centre, published an AI-native laboratory system called AutoDNA in Nature. It plans and executes molecular biology experiments on its own.
• Source: ITHome · 2026-11-08
• Editor's Take: A self-driving lab is the most credible form of physical AI because failure is measurable. The experiment either replicates or it does not, and no one has to write a benchmark for it.

IV. Foundation Models

1. Tencent's Yao Shunyu: the moat sits in the product, not the model

• Summary: Asked by Zhang Yaqin at a Future Science Prize dialogue what actually protects a model company, Tencent's chief AI scientist Yao Shunyu said the durable advantage lives in the product layer. A model can be replicated; the workflow built on top of it cannot.
• Source: ITHome · 2026-11-08
• Editor's Take: This is the argument you make when your model is not the best one. It also happens to be right for everyone except the two labs at the frontier.

2. iFlytek's Liu Qingfeng has delivered half of his AI plan

• Summary: TMTPost measures iFlytek's promises against its results, including Liu Qingfeng's admission at the April 29 earnings call that the share price had disappointed investors. The piece reads the company's all-domestic chip strategy as the explanation it keeps offering for a valuation that has not recovered.
• Source: TMTPost · 2026-11-08
• Editor's Take: Building on domestic silicon is a strategic asset and a quarterly tax. The market is charging iFlytek the tax now and may never pay for the asset.

3. MiniMax's Yan Junjie: the last five per cent is the hard part

• Summary: On the Invest in Progress podcast, MiniMax founder Yan Junjie put the gap between Chinese and US frontier models at roughly the final 5% of capability, and said closing it is far harder than getting the first 95%.
• Source: TMTPost · 2026-11-08
• Editor's Take: The last five points are the ones a customer can feel. Everything below them is a benchmark, and benchmarks are cheap to win.

4. China Mobile Cloud signs with Zhipu

• Summary: China Mobile Cloud and Zhipu signed a cooperation agreement on November 6 covering token technology, model services, product building and industry solutions. The stated scope is domestic compute, model serving and commercialisation.
• Source: ITHome · 2026-11-08
• Editor's Take: A telco brings the accounts a model lab cannot reach and takes the margin the lab cannot keep. Every Chinese model company will end up with one of these deals, and none of them will own the customer.

5. Tesla's inference roadmap: 80x HW4, and Dojo in orbit

• Summary: Tesla published an updated AI accelerator roadmap putting its sixth-generation inference chip at roughly 80 times the compute of HW4. Dojo D3 drops the SRAM-centric design for a multi-die package with stacked memory, and Tesla lists a space deployment for it.
• Source: ITHome · 2026-11-08
• Editor's Take: This roadmap is aimed at Optimus and robotaxi inference, not at training. A custom inference chip only pays off if you own the volume, and Tesla is the only carmaker that does.

6. Qualcomm pushes mixture-of-experts agents onto the handset

• Summary: At CIIE this week Qualcomm demonstrated an on-device mixture-of-experts agent, with its China AI product lead Wan Weixing arguing that three to four years of on-device AI have reached a turning point in how the phone gets used.
• Source: QbitAI · 2026-11-08
• Editor's Take: On-device agents are a memory business before they are a model business. An expert that has to swap in and out of mobile DRAM is a battery decision wearing an AI label.

7. Liang Wenfeng's home-province dinner

• Summary: Huxiu profiles the financing network around DeepSeek and traces how Liang Wenfeng's Guangxi and Guangdong relationships supplied early capital and hires. The line the piece keeps returning to is that new money in Chinese AI still moves through old networks.
• Source: Huxiu · 2026-11-08
• Editor's Take: "New money, old networks" is the least romantic and most accurate description of Chinese AI funding. It also explains why the sector consolidates around a handful of cities rather than a hundred.

V. AI Software

1. The AI office war: 20.97 million visits and 100,000 companies

• Summary: Huxiu's long read on China's AI office market puts Tencent's WorkBuddy at 20.97 million PC visits a month three months after launch, Alibaba's consolidation of QoderWork, Wukong and MuleRun into "Qianwen Office", and ByteDance's Doubao Work signing more than 100,000 companies in its first week.
• Source: Huxiu · 2026-11-08
• Editor's Take: All three products are distribution plays wearing an AI label. The model is a commodity at this layer; the seat on the desktop is the asset.

2. Tencent's agent stack is bigger than WorkBuddy

• Summary: Leiphone maps Tencent's agent portfolio beyond its best-known product, and the argument is that the company is wiring agents into surfaces it already owns rather than shipping one flagship assistant.
• Source: Leiphone · 2026-11-08
• Editor's Take: Tencent's advantage is owning the app people open every day, and that is also the constraint. Opening a captive surface to third-party agents costs it the control that made the surface valuable.

3. Ant's Lingguang passes 20 million users

• Summary: Ant Group's assistant Lingguang said it crossed 20 million users, and that in the week since version 2.0 launched on October 27 it handled more than 1.16 million tasks and executed 3.6 million actions.
• Source: ITHome · 2026-11-08
• Editor's Take: Tasks handled is the only figure here that matters, and it is still small. Twenty million users doing one task a week is a feature, not a platform.

4. Dewu plugs into Apple Intelligence in China

• Summary: Dewu's app now supports the China edition of Apple Intelligence. Point the iPhone camera at a product and the app returns listings, price history, an AI summary and user reviews without a search query.
• Source: ITHome · 2026-11-08
• Editor's Take: Camera-first shopping is the use case Apple has been promising since Visual Look Up. The winner is whoever owns the price history, and that is the commerce app, not the phone maker.

5. ByteDance's AI pharma spinout, and where the money went

• Summary: ByteDance carved its five-year-old AI drug-discovery unit into a standalone company, Anew Labs, in June 2026, moving the team, the algorithm platform and the pipeline out in one block. Huxiu's piece tracks big-tech executives into the sector and notes that roughly 60% of funding now goes to the top few players.
• Source: Huxiu · 2026-11-08
• Editor's Take: AI pharma consolidates because the expensive part is the wet lab and the trials, not the model. ByteDance kept control of the spinout and moved the costly half off its own books.

6. Chinese SaaS abroad: the discount is the business model

• Summary: A Chinese SaaS founder selling into Europe describes a signed deal where a 50% discount was still not enough and the customer kept pushing. Huxiu's essay reads the pattern as price competition against vendors who are really selling procurement trust.
• Source: Huxiu · 2026-11-08
• Editor's Take: Price is the only lever you have when nobody in the room has heard of you. The way out is a reference customer in the same industry, not a longer feature list.

7. Red Fruit's 200 million daily users, and the AI dramas filling them

• Summary: QuestMobile data cited by Huxiu puts Red Fruit, ByteDance's short-drama app, at 200 million daily actives and 430 million monthly actives as of September 2026, ahead of Doubao. AI-generated drama now makes up the majority of its catalogue.
• Source: Huxiu · 2026-11-08
• Editor's Take: Short drama is the first consumer category where AI video is cheaper than filming and good enough to watch. That makes it a content business with content margins, not a model business with software margins.

---

Section tally: Humanoid 5 · Autonomous 4 · World Models 5 · Foundation 7 · Software 7, 28 items.

物界前沿 | EastVoice