Community Discussion · Tracks

EastVoice · AI · 2026-09-15 · Issue 68

EastVoiceEastVoiceSep 152026/09/14 249 views

Editor's Note: China's model race went financial this week. Zhipu lifted MaaS revenue 27x, DeepSeek is reportedly laying IPO groundwork right after shipping a from-scratch V4.1 Flash, and list-plate humanoid makers are learning that demos do not hold a bid. Meanwhile Beijing's MIIT put a number on the robot dream: 100,000 humanoid units in 2026.

Commentary: The interesting story isn't that Chinese labs are pricing intelligence down 50% in a quarter. It's that investors stopped asking what the models can do and started asking who pays for them. That question is maturing the industry faster than any benchmark.


I. AI Foundation Models

1. DeepSeek Rebuilds V4.1 Flash From Scratch, and the Rewrite Is Cheaper Than the Original

  • Summary: Days after shipping V4.1 Flash with native multimodality, DeepSeek drew near-unanimous praise for speed and accuracy. The model shares almost no architecture with V4, and official figures say its HBM memory demand drops to a fraction of the previous generation.
  • Source: TMTPost · 2026-09-14
  • Editor's Take: Founder Liang Wenfeng is being called a saint on Chinese social media again. A model that discards its own predecessor's architecture is a bet that the moat is engineering, not accumulation.

2. DeepSeek Starts Counting Money, and Files Toward an IPO Track

  • Summary: Reports say DeepSeek is preparing for an IPO. Publicly it has one balance sheet showing stronger models, cheaper caching and lower token prices; the unpublished one, which investors are scrutinizing, has to answer what the company actually sells.
  • Source: Huxiu · 2026-09-14
  • Editor's Take: The open-source halo survives contact with a prospectus barely. Watch whether the V4.1 Flash price cut reads as efficiency or as customer acquisition once audited numbers exist.

3. Zhipu's API Business Grew 2,736% and Nobody at the Company Is Pretending That's Normal

  • Summary: Zhipu's interim report showed MaaS platform and API revenue of RMB 825 million (~$116M) for H1, up roughly 2,736% year-on-year, with token call volume up more than 40x and paying daily actives up 603%. Its flagship GLM-5.3 briefly held the top spot among domestic models.
  • Source: Huxiu · 2026-09-14
  • Editor's Take: Percentage growth off a small base is the oldest trick in listing brochures, but a 40x token-volume jump is real usage, not marketing spend. China's inference market is compounding.

4. Zhipu Leaks GLM-6.0 Through a Fundraising Filing, Then Raises RMB 33.5 Billion

  • Summary: Less than nine months after listing, Zhipu announced a placement of new H-shares plus RMB 2.014 billion (~$280M) in zero-coupon convertible bonds, for total net proceeds near RMB 33.5 billion. Buried in the disclosure: GLM-6.0 will publish a fully self-trained methodology.
  • Source: QbitAI · 2026-09-14
  • Editor's Take: Roadshow by prospectus. A lab that can raise tens of billions of renminbi right after an IPO is proof that China's capital markets have decided frontier models are strategic infrastructure.

5. Anthropic Accuses Moonshot's Kimi of Routing Requests Through Claude, and Chinese Users Get Told to Pick a Side

  • Summary: Anthropic alleged that Moonshot AI's Kimi relayed some customer requests to Claude, reigniting a distillation dispute. The Chinese commentary angle: buying one AI tool increasingly means choosing a geopolitical camp before you choose a product.
  • Source: Huxiu · 2026-09-14
  • Editor's Take: Every accusation like this teaches Chinese enterprise buyers the same lesson: supplier risk now includes where the model's government sits, so domestic substitutes get a free sales pitch.

6. Sam Altman Endorses Pacing Frontier AI — Carefully, As If Handling a Live Wire

  • Summary: OpenAI CEO Sam Altman said Sunday he supports managing the speed of advanced model development so capability growth doesn't outrun human control, while stressing this is not a call to stop progress. Anthropic CEO Dario Amodei's Sept 12 essay had asked for exactly that, and Elon Musk quickly backed it; US chip stocks fell in premarket on the discord.
  • Source: ITHome · 2026-09-14
  • Editor's Take: Three labs' CEOs agreeing on caution moved markets precisely because markets know the agreement won't hold. Pacing without verification is a press release.

7. Frontier Model Prices Halved in Three Months, and Intelligence Is Depreciating Like RAM

  • Summary: From September through November the industry's major labs — OpenAI, Anthropic, Google, and China's whole frontier tier — shipped model after model, and per-token prices for comparable capability fell by roughly half within a quarter.
  • Source: Huxiu · 2026-09-14
  • Editor's Take: Falling prices are great for every buyer and terminal for every lab whose only pitch is price. The survivors are the ones selling workflows while intelligence goes free.

II. AI Software & Applications

1. ByteDance's Doubao Phone Assistant Ships a Consumer Version, With a Deal That Lets Apps Say No to the AI

  • Summary: The consumer release of Doubao's phone assistant, working with handset makers at the OS level on interaction, task execution and local memory, lands ahead of the first Doubao-branded phone going on sale Sept 16. Notably, a GUI cooperation protocol now lets third-party apps explicitly allow or refuse AI-driven automation of their interfaces.
  • Source: Leiphone · 2026-09-14
  • Editor's Take: The opt-out clause is the real product. A phone agent that every app must consent to is slower than a rogue one, but it's the only version that survives a regulator.

2. Financial-AI Startup Xuntu Closes a RMB 300M+ Series B — Its Third Round in a Single Year

  • Summary: Xuntu Tech, a Chinese financial-data AI provider, completed a Series B exceeding RMB 300 million (~$42M), its third raise in twelve months, with insurers, brokerages and top VCs piling in.
  • Source: QbitAI · 2026-09-14
  • Editor's Take: Three rounds a year means each raise priced off the last, not off revenue. Financial AI is where Chinese capital goes when it wants AI upside with a familiar P&L attached.

3. Kioxia Weighs a US Listing to Raise at Least $10 Billion on the AI Trade

  • Summary: Kioxia Holdings is exploring an ADR offering to raise at least $10 billion (~RMB 67B), people familiar with the matter told Bloomberg, joining the wave of AI-adjacent companies tapping insatiable equity demand.
  • Source: ITHome · 2026-09-14
  • Editor's Take: Memory makers spent a decade being the industry's cyclical embarrassment. Now NAND is a growth stock, and the Japanese chip asset everyone doubted gets to pick its moment to sell.

4. Anthropic Tells Shareholders It Will Turn a Profit Again This Quarter

  • Summary: Per the Financial Times, Anthropic has informed some shareholders it expects a second consecutive quarter of positive adjusted operating profit in Q3 2026, after a first-ever profitable quarter in Q2 on roughly $11.5 billion of annualized revenue.
  • Source: ITHome · 2026-09-14
  • Editor's Take: The company publicly asking the industry to slow down is privately printing operating profit. That combination makes the slowdown letter harder to dismiss and easier to distrust.

5. Microsoft Puts xAI's Grok Into Word, Excel and Outlook

  • Summary: Microsoft extended its multi-model strategy in Microsoft 365 Copilot by integrating Elon Musk's xAI Grok models, following earlier additions of Anthropic's Claude into Researcher and Copilot Studio.
  • Source: ITHome · 2026-09-14
  • Editor's Take: Office is becoming a model bazaar with Microsoft as landlord. OpenAI built the greatest distribution deal in software history and is now watching it get commoditized seat by seat.

6. Douyin Just Beat WeChat on Screen Time. User Attention Stopped Being China's Hard Currency Anyway

  • Summary: Nomura, citing QuestMobile, put ByteDance's Douyin total usage at 19.4% of China's top-50 app time in July, above WeChat's 18.8% for the first time; the ByteDance stack holds 40.9% against Tencent's 29.1%.
  • Source: TMTPost · 2026-09-14
  • Editor's Take: The milestone lands just as AI assistants start completing tasks without lingering at all. The empire that won attention may be defending a currency that's about to be redenominated.

7. AI Micro-Dramas Hooked 600 Million Viewers in Six Months on Budgets That Humble TV Studios

  • Summary: AI-generated short drama Wan Yao Tu Lu Zhuan was produced by a three-person team on registered capital of RMB 100,000 (~$14K), delivering twelve seasons in half a year with estimated profit of RMB 6-20 million; AI titles now dominate Hongguo's popularity charts.
  • Source: TMTPost · 2026-09-14
  • Editor's Take: China's content industry feared AI flooding the zone with slop; it flooded the zone with margins. The studios' problem isn't quality, it's competing with a P/E nobody expected.

III. Humanoid Robots

1. Xiaomi's Lei Jun Visits Unitree, and Every Tech Titan Is Suddenly Photographed Next to a Robot

  • Summary: Xiaomi founder Lei Jun toured Unitree Robotics on Sept 14 with CEO Wang Xingxing, watching a humanoid demonstration. It follows months of smartphone and EV executives staging robot visits as the industry's de facto endorsement circuit.
  • Source: ITHome · 2026-09-14
  • Editor's Take: Lei Jun already runs a robot production line in his EV factory, so this wasn't tourism. Consumer-electronics giants are scouting humanoids the way auto companies once scouted batteries.

2. Unitree Ships the G1+ With Six Upgrades and the Same Aggressive Price Point

  • Summary: Unitree's G1+ adds improvements across neck mobility, joint motor output and thermal management, vision plus tactile sensing, battery life and far-field voice, keeping G1's value positioning. Standard G1+ pricing holds the company's floor under the sub-industry.
  • Source: ITHome · 2026-09-14
  • Editor's Take: Wuxi's rhythm — hardware refresh every few months at commodity prices — is how Unitree won drones' playbook. Everyone complaining about Chinese robot pricing is describing a moat, not a discount.

3. A Modular Humanoid for RMB 9,800 (~$1,380): A University Spinoff Sets a New Price Floor

  • Summary: LanChong Embodied, a Xi'an Jiaotong University technology-transfer company, launched the modular humanoid Mantis Standard "Xiaobai" starting at RMB 9,800, aimed squarely at research and education buyers.
  • Source: ITHome · 2026-09-14
  • Editor's Take: At $1,380 a unit, the interesting buyers aren't factories, they're labs. China is building the humanoid world's Raspberry Pi constituency.

4. Unitree's Stock Halved, Mech-Mind Broke Its Issue Price, and Nobody Is Explaining Why the Dip Came

  • Summary: Mech-Mind founder Shao Tianlan's public challenge to peers' "fake revenue" practice landed the same week embodied-AI listings slid, forcing the sector's central question out in the open: what revenue actually repeats after IPO lockups?
  • Source: Huxiu · 2026-09-14
  • Editor's Take: A founder accusing peers of inflating demo revenue while his own stock drops below issue price is peak listing-season China. The market's patience for robot videos has a number attached now.

5. The Listing Rush Turned China's Humanoid Startups Into Fighting Bots, Literally

  • Summary: Reports say Qianxun Intelligence, Xinghaitu and Robot Era were anonymously reported to competition organizers by a rival, as Shao Tianlan's viral post exposed the industry's "fake revenue" problem — performance demos propping up IPO narratives.
  • Source: Huxiu · 2026-09-14
  • Editor's Take: Sabotage petitions between labs sound like farce until you remember public listings are on the line. When the exit is a stock exchange, the competition becomes zero-sum before the market is even real.

IV. Autonomous Driving

1. Pony.ai Unveils Its Fourth-Generation Robotruck at IAA Hannover and Starts Courting Europe

  • Summary: Pony.ai launched a new L4 autonomous heavy truck at Germany's IAA Transportation on Sept 14, with volume production planned this year and road-test negotiations underway with several European governments.
  • Source: ITHome · 2026-09-14
  • Editor's Take: Freight corridors, not city streets, are where autonomy pays first, and Chinese players know European haulage is desperate for drivers. This is the export playbook: go where the labor shortage outruns the regulator.

2. WeRide's Revenue Grew 73% in H1. The Losses Grew Too.

  • Summary: WeRide reported H1 revenue of RMB 346 million (~$48.5M), up 73.3% year-on-year, with gross profit up 107%, but a net loss attributable to shareholders of RMB 790 million (~$111M).
  • Source: ITHome · 2026-09-14
  • Editor's Take: Margins improving faster than revenue is the only line that matters in an autonomous-driving P&L. WeRide is losing money at a visibly better rate, and that's what keeps the Nasdaq listing alive.

3. Zurich Airport Put Two Driverless WeRide Shuttles to Work. Employees First, Passengers Later.

  • Summary: Zurich Airport deployed two L4 fully autonomous electric shuttle buses with WeRide, initially ferrying airport staff between facilities before any passenger rollout.
  • Source: ITHome · 2026-09-14
  • Editor's Take: Private campuses as the proving ground is textbook robotaxi strategy — Switzerland adopting a Chinese playbook in reverse. The boring routes come first because boring routes generate the boring data that wins licenses.

4. Tesla's Robotaxi Service Goes 24/7 Next Month Once FSD V15 Lands

  • Summary: A Tesla AI executive said its ride-hailing operations, currently bound by fixed service windows in each market, will move to round-the-clock autonomous service after the next-generation FSD V15 architecture ships.
  • Source: ITHome · 2026-09-14
  • Editor's Take: Night shifts are where robotaxi unit economics actually work — no driver wages to cover, and utilization doubles. Tesla found the profit zone at 3 a.m., same as every fleet operator before it.

5. Horizon Robotics Turned Profitable. The Market Priced It Like a Company in Trouble.

  • Summary: Horizon Robotics reported H1 revenue of RMB 2.055 billion (~$288M), up 32.9%, swinging to net profit of RMB 3.784 billion as its driver-assistance chip price war succeeded — then watched investors question what happens after the prices are already down.
  • Source: Huxiu · 2026-09-15
  • Editor's Take: Winning a price war and losing the narrative anyway is the most honest signal in Chinese smart driving. Equity "democratization" ends wherever the gross margin does.

6. China's Unmanned Delivery Fleets Are Racing Abroad Before Domestic Rules Catch Up

  • Summary: Neolix started its first L4 pilot in Tokyo on Sept 11, days after appearing at a South Korean resort; Jiuzhi and peers are landing in Abu Dhabi, Singapore and Malaysia at a weekly clip, exporting the low-speed delivery playbook.
  • Source: Huxiu · 2026-09-14
  • Editor's Take: Overseas municipalities want delivery robots; Chinese regulators want caution. The arbitrage is regulatory, not technological, and it won't last forever.

V. World Models / Physical AI

1. Chinese Physical-AI Model PhysBrain 1.5 Tops the Global Open-Source Leaderboard

  • Summary: PhysBrain 1.5 took first place on the global open-source physical-AI ranking, with its spatial-intelligence scores described as running neck-and-neck with GPT-6 Astra — a striking claim for a model working the closed-loop physical tasks that most benchmarks skip.
  • Source: QbitAI · 2026-09-14
  • Editor's Take: Leaderboard wins in physical AI mean less than shipped robot hours, but they mean a lot to procurement committees. This is a sales document with weights attached.

2. Cog-WM 1.0 Claims the First World Model Built From Brain Mechanisms Rather Than Pixels

  • Summary: Shanghai-based Junao Panshi released Cog-WM 1.0 at the Pujiang Innovation Forum, billed as the first latent-prediction world model engineered from systematic brain-like neural mechanisms instead of video-data scaling.
  • Source: Leiphone · 2026-09-14
  • Editor's Take: Neuroscience-inspired claims have a long history of outliving their evidence. Judge Cog-WM when someone puts it behind a robot arm.

3. Hands-On Review of Five Chinese World Models Finds Impressive Demos and Thinner Reality

  • Summary: A Chinese tech-media team tested five domestic world models — the wave accelerated by World Labs' Atlas release — and found impressive generation quality alongside a big gap between "generating a world" and being usable outside research demos.
  • Source: Huxiu · 2026-09-14
  • Editor's Take: The most valuable sentence in world-model discourse right now: everyone's using the word, almost nobody has deployed it. China's fast-follower cycle is arriving at the same honest place Fei-Fei Li started.

4. At ECCV 2026, Unifying Autoregressive and Discrete Diffusion for Embodied Foundation Models

  • Summary: National University of Singapore's Show Lab professor Shou Zheng laid out a fusion of autoregressive modeling and discrete diffusion as the base for next-generation embodied foundation models, arguing against the "stack of separate models" approach dominating robotics today.
  • Source: Leiphone · 2026-09-14
  • Editor's Take: The architecture debate is where robotics stops looking like systems integration and starts looking like a race for the base-model crown. Watch which Chinese labs adopt it within six months — that's the real adoption metric.

Physical World Frontier | EastVoice

2 replies

?
Ctrl + Enter to reply
Tang
TangSep 15

Night shifts are killing me, don't bore me with technical details. Does this spreadsheet really save me half an hour?

Crypto Dropout
Reply to Tang

Is the Tokenomics reasonable? Pure AI services are hard to value; can this direction issue a token?