PHYSIX FRONTIER EASTVOICE
2026.10.09 · Friday
Issue 057
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
$500M
Butterfly Effect, parent of Manus, raised over $500M
HK$1.07 trillion
Zhipu's peak market cap, roughly 20x listing valuation
$1B
Nvidia discussed investing another $1B in Figure
$5B
Tencent considering offshore bond raise of up to $5B
Top Stories 3 · dual commentary ｜ Section Picks 22 · one-line takes
Manus Parent Butterfly Effect Closes Over $500M, Alibaba's Joseph Tsai Says Open Source Is Europe's Best Bet, and Pony.ai and Uber Will Test a Seventh-Gen Robotaxi in London.
A SiliconBase piece tracks Zhipu's run to a HK$1.07 trillion market cap at its peak, roughly 20x its listing valuation, before the stock slid to HK$696 by October 7. The argument is that Chinese model makers now trade on release cycles the way hardware names do.
Editor's take｜When a research lab's valuation swings with the next model launch, the market is pricing cadence, not capability.
Germany's search engine Ecosia has moved away from France's Mistral and adopted Chinese open-source models instead, according to Politico, saying Mistral's output fell short of what it needed.
Editor's take｜A European buyer choosing a Chinese model over a European one is the more telling detail. Price and quality still beat flag-waving.
At an event in Italy on October 7, Alibaba chairman Joseph Tsai argued that Europe's cleanest route to independence from both the US and China is open-source AI, and urged the continent to build its own compute base.
Editor's take｜Coming from the company that ships the weights, the advice is self-serving and still correct. Sovereignty runs on infrastructure, not gratitude.
22 AI stories worth your time, each with a one-line take.
Hu Han, former head of Tencent Hunyuan's vision-model algorithm center, is raising several tens of millions of dollars at a target valuation in the hundreds of millions, staying on multimodal models, with Yuanshi Capital as financial adviser.
One-line take｜The talent leaving China's biggest labs keeps landing on multimodality. That is where they expect the next default model to be decided.
A TMTPost survey of September's model announcements finds a wave of founders born after 1990 now leading flagship projects, including Alibaba's Liu Da Yiheng (b. 1993) on Qwen and others at Tencent and ByteDance.
One-line take｜Handing flagship models to people barely into their thirties is a bet on raw speed over pedigree. The next ranking tables will show whether it paid off.
A new paper from Kaiming He's group uses cat videos as training material to sharpen abstract reasoning, and reports gains on ARC-style challenges without task-specific supervision.
One-line take｜The result is less about cats than about what counts as a reasoning substrate. If video works, the pretraining pipeline gets a lot wider.
A Leiphone readout of this year's Yunqi Conference argues Alibaba's strategy has shifted, with the cloud event increasingly framed around selling the full stack rather than a single model.
One-line take｜At this scale, the model is a lead magnet and the cloud is the product. Alibaba has been clear about that for a while.
StepFun will unveil STEPX Neo on October 13 in Shanghai, billed as its first model-native agent phone, with the model, OS and hardware all designed around agent workflows rather than retrofitted on top.
One-line take｜Building a phone where the agent is the interface is the harder, more honest version of the pitch. Shipping it is the part nobody has managed yet.
Butterfly Effect, parent of the AI agent Manus, said it raised more than $500M in a round led by Boyu Capital and IDG Capital, with existing backers Tencent, Sequoia China and ZhenFund joining. Reports put the new valuation near $4 billion.
One-line take｜The deal that fell through was priced at roughly $2 billion. Getting funded at double that after a blocked sale is the market saying it wants the entry point more than the acquirer did.
Eight days after its 2.0 launch, Manus reopened a Beijing office and listed 17 open roles, all in the capital, as it prepares a product aimed at the domestic market.
One-line take｜Staffing up at home while the funding headline points overseas is the usual hedge. It says the domestic agent market is worth a second front.
Tian Keyu, a former ByteDance intern, has raised close to $30M at a post-money valuation of $200M for a company that still has no public name and no shipped product, with Wuyuan Capital among the backers.
One-line take｜Funding a lab before it has a name is the clearest sign of how thin the world-model talent pool still is.
Autonomous delivery startup White Rhino closed a C2 round, bringing its Series C to $100M, led by Yinshan Capital with state and Korean financial backers. The piece notes the whole C round is still under a third of what one rival raised in a single February round.
One-line take｜Raising $100M across a full Series C used to be a milestone. Against a single rival round it now reads as a slowdown.
Nvidia has discussed investing another $1B (about ¥6.71 billion) in humanoid robot maker Figure, according to The Information. Figure is seeking more funding at a pre-money valuation that has not been disclosed.
One-line take｜The chip vendor betting this hard on one humanoid company narrows the field. It also tells you which robot platform it expects to run its silicon.
Tencent is considering an offshore bond raise of up to $5B (about ¥33.6 billion), per Bloomberg, joining a growing list of tech firms turning to debt as AI demand pushes up capital needs. The notes may be priced in dollars and offshore yuan as early as this month.
One-line take｜Even the cash-rich names are borrowing for compute now. Debt is how this cycle keeps its spending from denting reported earnings.
A Tsinghua-linked company founded seven months ago has raised $400M across three rounds at a $1.42 billion post-money valuation, while its website carries a single English tagline and no team photo or product page.
One-line take｜Capital is funding the resume and the pedigree, not the artifact. In this market the artifact arrives later, if at all.
A Huxiu account from the pig-farming sector finds a 50-person equipment maker in Foshan winning customers during a downturn, with a non-technical owner rewriting a decade-old software product using AI tools alongside interns.
One-line take｜The adoption story that matters is not in Shenzhen's campuses. It is a farm-tools shop in a bad year deciding the rewrite is worth it.
UBTech signed a strategic deal with FAW-Volkswagen to develop and test embodied robots in logistics, with both sides planning demonstration sites to speed up humanoid deployment in smart manufacturing.
One-line take｜Automakers keep becoming the first paying customers for humanoids, because their logistics floors are repetitive and measurable. That is where the deployment data will come from.
Speaking at IROS 2026, Zhanji Dynamics' Feng Zongbao described work on robots that tend CNC machines themselves, handling pick, load, insert and remove on precision parts with tolerances as tight as 0.5mm.
One-line take｜A robot that runs the machine that makes more robots closes the loop everyone talks about. The tolerance is what separates a demo from a factory.
A collaboration between Galaxy General and Tsinghua's LATENT lab trained a robot to rally at tennis using imperfect human motion data, presented at IROS 2026, sidestepping the need for clean expert trajectories.
One-line take｜Learning from messy human data is the point. Perfect demonstrations do not scale, and robots that only work on them never leave the lab.
Pony.ai and Uber plan to bring a seventh-generation robotaxi to London, starting testing within weeks as part of an expanded partnership on driverless rides.
One-line take｜A Chinese stack riding a Western platform into a new market is the pattern to watch. London is a test of regulators, not of the car.
XPeng branded its robotaxi business XPENG YOYO and launched a matching ride-hailing mini-program, positioning the service for both domestic and overseas readers with an easy-to-pronounce name.
One-line take｜Naming the service before it operates is expectation-setting. The app going live tells you the launch is close.
Kalz Dynamics is scaling driverless freight in Inner Mongolia's coal and industrial corridors, betting that platooning heavy trucks on fixed routes is the near-term path to autonomy revenue.
One-line take｜Fixed routes between mines and plants are the least glamorous and most fundable corner of autonomy. Repetition is the whole business case.
At APRCE 2026, Zhenghang Innovation launched what it calls the first 24/7 physical-AI service solution for open retail scenes, built on an embodied brain and pitched as a human-robot collaboration model for convenience stores.
One-line take｜Retail is a clean place to sell physical AI because the tasks are narrow and the stores are already instrumented. Convenience stores may become the first honest scoreboard.
At its 2026 tech day, Leapmotor said it will push its LWM world-model driving system to 350,000 already-sold lidar cars for free, claiming a first-tier ADAS stack without relying on BEV perception or rule-based code.
One-line take｜Pushing a world model to every car already on the road is a fleet-wide test. Whether the result holds up is now a public question, not an internal one.
Qizhi Mobility is positioning itself not as a model trainer but as the data infrastructure behind physical AI, arguing that as world models improve, verified real-world data becomes the scarce asset.
One-line take｜Selling picks and shovels to the physical-AI rush is a familiar play, and the moat only holds if the data stays hard to replicate.
📖 The full EastVoice issue on the Chinese forum