EastVoice

PHYSIX FRONTIER EASTVOICE

2026.09.11 · Friday

Issue 033

FIVE MINUTES A DAY, THE GLOBAL AI SHIFT

437

KV-cache memory reduction factor in DeepSeek V4.1 Flash architecture

$23B

Valuation of Musk's Boring Company after Series D funding round

¥750,000

Damages awarded to miHoYo in Shanghai AI voice-cloning case

66%

Horizon Robotics first-half gross margin reported in financial results

Top Stories 3 · dual commentary | Section Picks 24 · one-line takes

TOP STORIES

DeepSeek ships V4.1 Flash with price cuts, regulators tighten humanoid IPO revenue bars, and AutoNavi launches first 3D-native city world model.

1DeepSeek ships V4.1 Flash, claims it beats its own flagship and cuts prices doing it

DeepSeek released V4.1 Flash on September 10, the smallest model in its new architecture family, with native multimodal vision and API pricing cut below the V4 tier. The company says it outperforms V4 Pro on most benchmarks while serving faster.

What it means for you|Developers get access to a cheaper flagship-killer model that sets the floor for agent inference economics.

Editor's take|A lab that just denied IPO rumors releasing a cheaper flagship-killer the same week is priced-to-move behavior. DeepSeek is setting the floor for agent inference economics.

Source: ITHome (2026-09-10)

2DeepSeek's KV cache shrinks 437x. Agent costs are about to be repriced

A technical teardown of V4.1 Flash's architecture shows KV-cache memory down by a factor of 437, the bottleneck that makes long-context agents expensive to run.

What it means for you|If numbers hold, every US lab selling per-token agent plans has a math problem to explain.

Editor's take|Benchmark wins get the headlines; memory engineering gets the margins. If the numbers hold, every US lab selling per-token agent plans has a math problem to explain.

Source: Leiphone (2026-09-10)

3V4.1 Flash lands on the National Supercomputing Internet the day it ships

DeepSeek's new model went live on China's state-run National Supercomputing Internet within hours of release, giving developers API access through the government compute grid.

What it means for you|Developers gain API access through the government compute grid via day-one deployment on sovereign infrastructure.

Editor's take|Day-one deployment on sovereign infrastructure is distribution policy, not a press release. The state cloud and the best open lab are converging.

Source: ITHome (2026-09-10)

SECTION PICKS

24 AI stories worth your time, each with a one-line take.

AI Foundation ModelsMoonshot denies dual HK-STAR listing report within hours

After the South China Morning Post reported that Moonshot AI was weighing a dual listing in Hong Kong and Shanghai, a person familiar with the matter told LSAA News the report was "not true."

What it means for you|The market stops reading denials as Chinese frontier labs are rumored into an IPO cycle.

One-line take|Every Chinese frontier lab is now rumored into an IPO cycle whether it likes it or not. The denials are newsworthy mainly because the market stops reading them.

Source: ITHome (2026-09-10)

AI Foundation ModelsWeChat quietly wires personal AI agents into real friend graphs

WeChat has begun testing an AI social feature that places a user's personal agent directly inside real-world contact chats, per Geekpark's report — the first attempt to run agents on top of China's dominant social graph.

What it means for you|OpenAI would sell its datacenter for this distribution as agents run on China's dominant social graph.

One-line take|Agents are useless without context, and WeChat owns the richest social context in China. OpenAI would sell its datacenter for this distribution.

Source: Geekpark (2026-09-10)

AI Foundation ModelsKai-Fu Lee at CIFIT: AI can execute tasks but can never take the blame

01.AI CEO Kai-Fu Lee told the Xiamen Investment Promotion Fair that enterprise AI is shifting from answering questions to executing workflows, but genuine transformation still requires C-suite sponsorship because AI cannot bear responsibility.

What it means for you|Genuine transformation still requires C-suite sponsorship because AI cannot bear responsibility for executed workflows.

One-line take|The accountability wall is the real adoption ceiling. Model capability curves can't solve org charts.

Source: ITHome (2026-09-10)

AI Foundation ModelsGPT-6 is here, so why won't anyone print "AGI" on the receipt?

A widely shared Huxiu essay dissects OpenAI president Greg Brockman's September 3 "welcome to the AGI era" line, arguing the industry deliberately left the definition unfalsifiable — impressive demos still miss the reliability bar the term implies.

What it means for you|Impressive demos still miss the reliability bar implied by the term AGI despite industry hype.

One-line take|Chinese tech media doing rigorous English-language hype auditing on OpenAI's own announcements is a role reversal worth noticing.

Source: Huxiu (2026-09-10)

AI Foundation ModelsPrinceton's Mengdi Wang at Bund Summit: bigger models won't give you autonomous discovery

Speaking at the 2026 Inclusion Bund Conference on September 10, Mengdi Wang argued today's models reliably find the most probable answer, which is precisely what science is not, and that autonomous discovery needs new training objectives, not just scale.

What it means for you|Autonomous discovery needs new training objectives rather than just scale since models find probable answers.

One-line take|The scale-is-all-you-need sermon finally has a counter-sermon from inside the pulpit. Watch whether AI-for-science budgets follow the argument.

Source: Leiphone (2026-09-10)

AI Software & ApplicationsMusk's Boring Company raises $3B at $23B, with the UAE writing the biggest check

The Boring Company closed a $3B (~¥20.2B) Series D led by UAE-linked sovereign investors, valuing the tunneling firm at $23B. Andreessen Horowitz and Sequoia also participated.

What it means for you|Tunnel boring was never the asset; optionality on Elon is what Gulf capital buys.

One-line take|Gulf capital keeps buying call options on Musk's entire portfolio. Tunnel boring was never the asset — optionality on Elon was.

Source: ITHome (2026-09-10)

AI Software & ApplicationsAnt-adjacent agents move into the real world: AI shopping via one sentence in Alipay

Fliggy integrated with Alipay's AI assistant "Abao", letting users book flights and attraction tickets through a single natural-language command, with MCP handling the tool-call plumbing.

What it means for you|Western agentic-commerce startups are starting three rounds behind as super-apps bolt commerce onto payment rails.

One-line take|China's super-apps are bolting conversational commerce onto payment rails that already hold the credit card. Western agentic-commerce startups are starting three rounds behind.

Source: Leiphone (2026-09-10)

AI Software & ApplicationsGeekpark: agents are going looking for bodies

A long-read on why embodied execution is the next competitive layer for Chinese agent developers, with hardware firms like Xiaodu marketing "a body for your agent."

What it means for you|Agents reaching for hardware is industrial policy by other means in a country manufacturing bodies.

One-line take|The abstract became obvious. In a country that manufactures the bodies, agents reaching for hardware is industrial policy by other means.

Source: Geekpark (2026-09-10)

AI Software & ApplicationsApple's new Siri finally understands personal context

Zimu AI's dispatch from Apple's September 10 keynote, under new CEO John Ternus, describes a Siri that reasons over the user's own messages, mail, and files — the personal-context layer Apple kept promising since 2024.

What it means for you|Apple now performs AI at WWDC for an audience that used to perform for it.

One-line take|Chinese coverage framing Apple's Siri catch-up as news to Chinese readers tells you how far the narrative has flipped. Apple now performs AI at WWDC for an audience that used to perform for it.

Source: TMTPost (2026-09-10)

AI Software & ApplicationsmiHoYo wins Shanghai's first AI voice-cloning unfair-competition case

A Shanghai Pudong court ordered a voice-morphing app to pay ¥750,000 (~$105K) after it sold paid voice packs of 63 Genshin Impact characters without authorization — the city's first AI voice-cloning verdict.

What it means for you|China is writing voice-likeness case law while US states are still arguing about it.

One-line take|The damages are trivial; the precedent isn't. China is writing voice-likeness case law while US states are still arguing about it.

Source: ITHome (2026-09-10)

AI Software & ApplicationsiFLYTEK Spark X2.5 hands-on: a 293B model parsing a 61-page earnings report

QbitAI's test run of Spark X2.5 covers particle-physics animation, financial-report teardown, and code debugging, with iFLYTEK also open-sourcing 4B and 1.7B on-device variants that run 128K context locally. The launch comes with a limited-time 50% API price cut, the second Chinese lab to discount inference this week after DeepSeek.

What it means for you|Two labs cutting prices in one week signals a supply glut forming at the inference layer.

One-line take|Two labs cutting prices in one week is a supply glut forming at the inference layer. Agents, not chatbots, are the customers being fought over.

Source: QbitAI (2026-09-10)

AI Software & ApplicationsListen Labs signed a $125M term sheet, then handed it back

Huxiu's account of the voice-AI research firm rejecting Menlo Ventures' led Series C at a $1.5B valuation to pursue a Salesforce acquisition instead — a rare signed-and-returned deal in the Chinese business-press telling.

What it means for you|When strategic buyers outbid term sheets, the venture model's exit math is being repriced.

One-line take|When strategic buyers start outbidding term sheets, the venture model's exit math is being repriced in real time. Chinese deal press covering a US teardown is the globalization of the story.

Source: Huxiu (2026-09-10)

AI Software & ApplicationsHorizon Robotics books 66% gross margin, aims for breakeven in two years

Horizon's first-half report shows 66% gross margin and its software deal with Volkswagen's CARIAD ramping, with management targeting profitability within two years.

What it means for you|Watch whether margin survives the price war as Horizon aims for profitability within two years.

One-line take|China's ADAS chip champion is about to be the most interesting profitability proof point in the entire driver-assist supply chain. Watch whether margin survives the price war.

Source: Huxiu (2026-09-10)

Humanoid RobotsBeijing quietly raises the bar for humanoid IPOs

The Information reports Chinese regulators have informally told banks and funds that prospective humanoid-robot listings must clear a higher bar, including proof of durable commercial revenue — right as Unitree's debut week ended with its shares halving from a 629% pop.

What it means for you|China is deliberately pricking its own robotics bubble before it prickles something bigger via higher bars.

One-line take|A window-guidance memo after the Unitree circus is policy reacting in real time. China is deliberately pricking its own robotics bubble before it prickles something bigger.

Source: Huxiu (2026-09-10)

Humanoid RobotsMech-Mind's founder calls out Galbot over "sham revenue" as the IPO audit turns inward

Mech-Mind founder Shaotianlan Shao published pointed comments alleging peers inflate revenue to dress up IPO filings, naming the accusation at the exact moment regulators tighten listing gates.

What it means for you|Publicly traded robotics will start looking like semis did in 2022 with accusations first.

One-line take|Industry founders policing each other's numbers beats any audit memo. Publicly traded robotics will start looking like semis did in 2022 — accusations first, restatements later.

Source: Huxiu (2026-09-10)

Humanoid RobotsRobots will climb Mount Tai for real on October 29

China Media Group announced the first CMG World Robot Mount Tai Challenge, broadcasting live October 29 as competing robots attempt the mountain's actual stone stairway.

What it means for you|Terrain trials double as procurement marketing in this industry as state TV choreographs robot spectacles.

One-line take|State TV choreographing robot stair-climbing as national spectacle worked for half-marathons. Terrain trials double as procurement marketing in this industry.

Source: ITHome (2026-09-10)

Humanoid RobotsHSImul3R: the first simulatable human-scene interaction reconstruction framework

DACapital Robotics, NTU S-Lab and Shanghai AI Lab released HSImul3R, which converts ordinary human videos into physically simulatable interaction data — a pipeline meant to turn YouTube into robot training fuel.

What it means for you|A Chinese academic consortium shipping the reconstruct-and-simulate loop first shows serious Western-lab-adjacent competition.

One-line take|Everyone wants video-to-policy. A Chinese academic consortium shipping the reconstruct-and-simulate loop first is a reminder that the robot data race has serious Western-lab-adjacent competition.

Source: QbitAI (2026-09-10)

Autonomous DrivingUS lawmaker asks NHTSA to probe "sleeping at the wheel" Tesla videos

Illinois Rep. Raja Krishnamoorthi wrote to Transportation Secretary Sean Duffy requesting a federal review of drivers deliberately sleeping while using Tesla's Full Self-Driving, per CNBC reporting on September 10.

What it means for you|Regulation by viral clip is now a formal channel as driver-asleep TikToks become lobbying instruments.

One-line take|The driver-asleep TikTok has become a lobbying instrument. Regulation by viral clip is now a formal channel.

Source: ITHome (2026-09-10)

Autonomous DrivingIM Motors LS6 pre-orders open at ¥209,900 with steer-by-wire standard on every trim

SAIC's IM Motors opened pre-sales on the new LS6 from ¥209,900 (~$29.5K), making full steer-by-wire chassis standard across the lineup — a feature Tesla Cybercab still hawks as exotic.

What it means for you|Standard steer-by-wire under $30K would have sounded fictional two years ago as China commoditizes autonomy.

One-line take|China is commoditizing the hardware layer of autonomy through price, not press events. Standard steer-by-wire under $30K would have sounded fictional two years ago.

Source: ifanr (2026-09-10)

Autonomous DrivingRoewe's new EREV pairs Momenta's R7 "world model" driving stack with 1,480km range

SAIC Roewe's Jiayue 07 opens pre-sales September 21 with a range-extended powertrain rated at 1,480km and Momenta's next-gen R7 high-end driver-assist solution onboard.

What it means for you|Momenta selling world model branding into mass-market sedans shows where the vocabulary war is won.

One-line take|Momenta selling "world model" branding into mass-market sedans shows where the vocabulary war is being won — on window stickers, not whitepapers.

Source: ITHome (2026-09-10)

Autonomous DrivingNeolix unmanned delivery vehicles start running on Jeju Island

Chinese autonomous delivery firm Neolix deployed delivery robots at a Jeju resort through Korean partner NT1 — its first commercial outdoor resort operation in Korea, moving from pilots to paid deployments.

What it means for you|Chinese delivery fleets are quietly landing in allied-country tourism corridors while America argues about geofences.

One-line take|While America argues about robotaxi geofences, Chinese delivery fleets are quietly landing in allied-country tourism corridors. Export momentum rarely looks like the headline category.

Source: Leiphone (2026-09-10)

World Models / Physical AIAutoNavi releases ABot-Earth 0.7, billed as the first 3D-native city world model

Alibaba's AutoNavi unveiled ABot-Earth 0.7 on September 10, claiming the world's first 3D-native city-scale world model, positioned as the entry point for AI to understand physical cities.

What it means for you|AutoNavi moving first makes a Google Maps announcement inevitable as mapping companies hold key datasets.

One-line take|Mapping companies have the only world model dataset that matters — the actual planet, surveyed. AutoNavi moving first makes a Google Maps announcement inevitable.

Source: QbitAI (2026-09-10)

World Models / Physical AIJD builds a 100,000-card domestic compute cluster and a JoyAI world-model line

At JDDiscovery 2026 in Beijing, JD disclosed a 100,000-accelerator domestic-chip cluster, a 10-million-hour embodied-data collection program, the JoyAI world-model series, and a 5-year plan to buy 3 million robots, per the same-week reports.

What it means for you|A retailer ordering national-scale compute clusters on Chinese silicon shows decoupling has an enterprise checkout button.

One-line take|A retailer ordering a national-scale compute cluster on Chinese silicon is the clearest sign yet that decoupling has an enterprise checkout button.

Source: QbitAI (2026-09-10)

World Models / Physical AIA 1.3B-parameter open world model runs real-time on one consumer GPU

QbitAI covered LingBot-World 2.0's lightweight release, a domestic open-source world model at 1.3B parameters capable of real-time inference on a single consumer graphics card.

What it means for you|World models on one RTX card guts the frontier labs only gatekeeping argument overnight.

One-line take|World models on one RTX card guts the "frontier labs only" gatekeeping argument in that subfield overnight. Open weights keep finding the off-ramp from scaling orthodoxy.

Source: QbitAI (2026-09-10)

📖 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 →

EastVoice — five minutes a day, the global AI shift
Physix Frontier (Shenzhen) · covering foundation models, AI software, humanoid robots, autonomous driving and physical AI