PHYSIX FRONTIER EASTVOICE
2026.09.20 · Sunday
Issue 042
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
2,000 tokens/s
Alibaba's Qwen 3.8 27B model generation speed
¥5,999
Price of Nubia's Doubao-powered AI agent phone
90%
Accident reduction from driverless trucking per ministry official
$670 million
Spirit AI funding raised in two years
Top Stories 3 · dual commentary ｜ Section Picks 23 · one-line takes
Alibaba's fast Qwen model, China Telecom's domestic-chip open-source release, and Zhipu's repository upload apology.
Tongyi Qianwen / Qwen's 3.8 27B model is trending among Chinese developers, pushing close to 2,000 tokens per second and producing finished web pages in seconds. Reviewers note it handles design and front-end in one pass, while back-end work still needs a human.
Editor's take｜Small, fast, open-weight models are where the developer mindshare war is actually being won. A 27B that outputs at streaming-video framerate reframes what "local model" can do.
China Telecom released Xing4.0-29B-A4B, a 29B-parameter MoE model with 4B active parameters and native 256K context expandable to 512K. The company says it is the first Chinese agent-oriented model at this scale trained end-to-end on domestic chips and a domestic framework.
Editor's take｜The export-control bet in one line: if a telecom carrier can train a usable model without NVIDIA, the containment thesis quietly loses its teeth.
DeepSeek published an API peak/off-peak schedule: makeup work weekends and national-holiday days are billed as off-peak the entire day, extending its already aggressive discount pricing into China's calendar quirks.
What it means for you｜Developers pay off-peak rates on makeup work weekends and national holidays for DeepSeek API usage.
Editor's take｜Pricing pegged to the lunar calendar is a very Chinese kind of clever. Nobody else on earth optimizes inference costs around make-up workdays.
23 AI stories worth your time, each with a one-line take.
Developers found hundreds of megabytes of encrypted packets in a hidden local folder of Zhipu's ZCode desktop client; investigation showed the app packaged and uploaded whole repositories in the background while logged in. Zhipu issued a public apology, and the episode landed in this week's industry roundup as the agent-security story of the week.
What it means for you｜Zhipu's coding client uploaded entire user repositories to servers without explicit permission, breaching privacy expectations.
One-line take｜The real casualty is trust in local-first coding agents. An assistant that reads your entire repo without asking is a backup service you never subscribed to.
After leaving Tencent in July amid rumors of a Meta move, researcher Yao Xingcheng joined Thinking Machines Lab, the startup founded by OpenAI's former CTO. His Tencent compensation package was reportedly in the tens of millions of yuan, while the broader story is a de-rating of AI talent's middle tier even as the top layer keeps climbing.
One-line take｜Cross-border talent flows keep working even while governments talk about decoupling. A Chinese lab founder's startup and a US startup keep the same hiring pool.
At Huawei Connect, executives unveiled a chip roadmap with the Ascend 960DT pulled forward by three quarters and doubled single-chip compute, while computing-strategy chief Zhu Zhaosheng declared CANN has "crossed the ecosystem inflection point." Deputy chairman Wang Tao added that building an AI compute base "takes far more than one good chip."
One-line take｜Huawei moved the argument from silicon parity to software maturity, which is exactly where CUDA's moat lives. That is either the bravest or the most honest thing a challenger can say.
Nubia's NaviX Ultra, powered by ByteDance's Doubao phone assistant consumer edition, went on sale September 16 at ¥5,999 (~$840). It is the first "AI agent phone" to move from engineering sample to mass-market product; the category question now shifts from can it work to can it keep working through OS updates and carrier pushback.
What it means for you｜Users can buy the first mass-market AI agent phone that runs an OS via voice interaction.
One-line take｜ByteDance is betting the agent runs the phone, not the other way around. If users accept an OS they talk to instead of tap, app stores are the next disruption target.
A hands-on comparison of the three big office agents, ByteDance's Doubao Work, Tencent's WorkBuddy and Alibaba's Qwen Office, found Doubao updating fastest, including multi-agent parallelism and computer-control features added September 2, WorkBuddy winning on reliability, and Qwen lagging on workflow depth.
What it means for you｜Enterprise users choose among three competing office agents with varying reliability and feature depth.
One-line take｜Three internet giants running the same race on the same track, weekly. This is the most competitive agent-application market on earth right now, and enterprise China gets to pick from the wreckage.
A screenshot circulating in recruiter groups shows an embodied-AI company offering a ¥10.5 million (~$1.5M) total package for an algorithm expert, with a 26% headhunter fee, about ¥2.73 million on one deal. Staffing firm Career International disclosed AI-related roles now dominate its headhunting revenue.
One-line take｜When brokers earn more than engineers, the market is pricing scarcity, not productivity. The robot companies paying these fees believe talent is the only moat left.
A McKinsey Quarterly study of 701 global executives (late 2025 into early 2026, across six regions) sorted companies by how deeply AI is embedded in daily work and found the biggest gains coming from reinventing existing workflows rather than buying frontier models.
What it means for you｜Companies gain more value by reinventing workflows than by purchasing frontier models directly.
One-line take｜The survey keeps landing on the boring answer: process beats procurement. That is bad news for anyone selling AI as a drop-in product.
After the Federal Reserve raised rates 25 basis points, its first hike in three years, Chinese analysts are mapping the transmission to their own sector: higher global discount rates squeeze valuations and make dollar funds cautious just as China's AI primaries are crowding in.
What it means for you｜Higher global discount rates squeeze valuations and make dollar funds cautious for Chinese AI startups.
One-line take｜Watch how fast "decoupling" rhetoric and rate sensitivity coexist. One Fed decision reprices capex plans in Shanghai the same week.
DeepSeek's first external round, over ¥50 billion (~$7B) three months ago, pushed H1-2026 China AI funding past all of 2025. Investors close to the deals are now publicly urging founders to mark valuations to revenue reality as capital concentrates in fewer, larger checks.
What it means for you｜Investors urge founders to align valuations with revenue reality as capital concentrates in fewer checks.
One-line take｜The people writing the biggest checks are the first to warn about prices. That is a late-cycle tell and an opportunity, depending on which side of the term sheet you sit.
Honor launched a mobile version of Study Space, bundling AI homework tutoring, synchronized textbooks and lesson videos on Magic8 and WIN-series devices first.
What it means for you｜Parents access AI homework tutoring and synchronized textbooks on Honor Magic8 and WIN-series devices.
One-line take｜Education is China's stickiest consumer-AI wedge, and handset makers know parents won't price-shop a good report card.
At the G9L launch on September 17 in Beijing, CEO He Xiaopeng opened the show with the IRON humanoid walking off an automated line, claiming the world's first robot-built-robot demonstration. Industry coverage frames it as automakers chasing embodied AI past the froth stage, facing survival tests on cost, reliability and demand.
One-line take｜A car company stealing its own car launch for a robot cameo tells you where management believes the decade's margin is coming from.
Unitree Robotics closed at ¥514.98 on September 18, up 2.33%, market cap ¥208.3 billion — exactly one month after listing, and after a slide to ¥470 on September 14, 57% below the day-one high of ¥1,100 that vaporized over ¥240 billion in value.
One-line take｜The first humanoid IPO repriced the category in public. The bounce-off-the-bottom matters less than the lesson: robot dreams now have a daily mark.
Mech-Mind, which listed in Hong Kong in early September, saw founder Shao Tianlan publicly attack peers as assembly-job startups that manufacture revenue through data-collection centers and leasing affiliates before going public at two or three years old, as scrutiny grows over who is building products and who is building financial statements.
One-line take｜Insiders naming names is what happens when public money arrives before revenue does. Expect a round of forensic audits disguised as due diligence.
Xuanji Dynamics launched the Hypertron-SW01, a mid-size industrial quadruped with 45° steep-slope climbing and autonomous docking for charging, pitched at inspection and patrol across power, mining and pipelines.
One-line take｜Four legs beat two for every industrial contract signed this year. The humanoid hype and the revenue live in different bodies.
With every automaker pivoting toward robotics, an industry analysis tallies 16.74 million workers in China's auto sector and notes that every wave of headlines, EVs last year, robots this year, promises "million-level" job gaps while Germany, France, the US and Japan show the transition destroys more roles than it promises in the same window.
What it means for you｜Workers face potential job gaps as automakers pivot toward robotics and autonomous systems.
One-line take｜The assembly line's replacement is running on the same balance sheet as the people it replaces. Nobody wants the accounting that follows.
Speaking at the TEDA auto-industry forum on September 19, Xu Jian of the Research Institute of Highway, Ministry of Transport, cited industry research that scaled trunk-line autonomy would cut comprehensive transport costs over 20% and eliminate more than 90% of accidents on those corridors.
What it means for you｜Trucking companies could cut costs over 20% and eliminate most accidents on trunk-line corridors.
One-line take｜A regulator quoting a 90% safety number before a rulebook exists is a policy launch, not a speech.
Zhang Xiaoyu, chief scientist at the Chongqing-based auto giant, said at the same forum that L3 driving will reach mass production in 2027 and L4 systems begin commercial operation in 2028, aligning one of China's biggest state carmakers with the provincial licensing timelines now rolling out.
What it means for you｜Suppliers spend against Changan's roadmap for L3 mass production in 2027 and L4 in 2028.
One-line take｜When a state-owned giant prints a date, suppliers start spending against it. The calendar is the industry's new order book.
Chery Jaguar Land Rover's reborn Freelander 8 began customer deliveries September 19, priced ¥309,900–399,900 before incentives and from ¥289,900 (~$40.7K) after, running Huawei's Qiankun ADS 5 driver-assist stack.
What it means for you｜Consumers buy a reborn Freelander 8 starting at ¥289,900 with Huawei's driver-assist stack.
One-line take｜A British badge now sells on Chinese silicon. The interesting export here is Huawei's software, not the Land Rover nameplate.
Avatr, the Changan-Huawei-CATL joint venture, released final imagery for the 06T Shooting Brake Knight Edition ahead of its September 20 launch, headlined by an 896-line laser radar on the roof and distributed electric drive.
What it means for you｜Avatr 06T buyers get an 896-line laser radar and distributed electric drive features.
One-line take｜Resolution wars have moved from cameras to LiDARs, and 896 lines reads like yesterday's megapixel race. Consumers win the parts-count inflation.
Reuters reports Xpeng plans to sell its E/E architecture, cockpit systems, Turing AI chips and ADAS software to overseas carmakers beyond Volkswagen, extending licensing into robotaxi and robotics as it hunts high-margin revenue outside the Chinese price war.
What it means for you｜Foreign automakers license Xpeng's full self-driving stack, chips, and cockpit systems.
One-line take｜The endgame of a Chinese EV maker is becoming a supplier to global incumbents, Huawei's playbook with the Xpeng logo on it. Cars as loss leaders, software as the product.
Chinese embodied-AI startup Spirit AI, which raised $670 million in two years at a $2.9 billion valuation, expects a breakthrough in robot "brains" as soon as mid-2027, while its founder warns household deployment is at least eight years off; tens of its Moz1 wheeled humanoids already work on CATL and JD.com production lines, with high-quality data flagged as the binding constraint.
What it means for you｜Factories deploy imperfect robots now while waiting eight years for household robot deployment.
One-line take｜Inside a factory, imperfect robots still beat absent ones. Spirit AI is collecting the training data competitors can only simulate.
Songyan Dynamics, the robot maker famous from a humanoid half-marathon and a Spring Festival gala sketch, unveiled a general embodied-intelligence technology brand, Scalabot, with the HERON-CRA and HERON-World Model series, joining Unitree and UBTECH in the rush from bodies to brains.
One-line take｜China's robot companies all discovered the same quarterly truth: the hardware demo is a commodity, and the world model is the differentiated SKU.
Tesla's robotics team landed in Ningbo on September 16 and began fresh supplier audits on September 17 for Optimus mass production, issuing early orders tied to exclusivity and consistency checks, against a plan to build around 50,000 units in 2026 for its own plants while consumer sales remain unavailable.
What it means for you｜Tesla audits Ningbo suppliers for Optimus mass production targeting 50,000 units in 2026.
One-line take｜The most important humanoid supply chain still runs through Zhejiang factories. Whatever the robot is on stage, it is Chinese actuators in the invoice.
📖 The full EastVoice issue on the Chinese forum
📈 Physix Frontier · Alpha — the investor's read
Headline teardowns · curated signals · market moves. Five minutes to find the alpha of the AI era →