物界前沿EastVoice · AI · 2026-09-17 · Issue #70
Editor's Note: China's AI industry spent Tuesday arguing with itself. DeepSeek's lead attention-operator author published an essay about burying his own talent and hit 100k+ reads before the company distanced itself from it; XPeng's IRON robot came off an automated final-assembly line while Tesla and Figure are still fighting over production capacity; and humanoid-robot valuations in the A-share market have halved within a month of Unitree's (宇树科技) IPO. Meanwhile money keeps moving: AI drug startup Anew Labs closed a $290M first independent round backed by Sequoia China, Hillhouse and IDG.
Commentary: The pattern this week is deceleration of hype and acceleration of shipping. Fewer keynotes, more factories — that trade is usually what a maturing industry looks like.
I. AI Foundation Models
1. DeepSeek Hunts for the Cheapest Tokens in Ulanqab
- Summary: DeepSeek has anchored its infrastructure push in Ulanqab, Inner Mongolia, recruiting senior data-center operations engineers at up to RMB 30,000/month (~$4,200) and clustering compute where power is cheap. The town has become the punchline of Chinese tech forums — "blue sky, grassland, and a starry sky for 30k a month."
- Source: Huxiu · 2026-09-16
- Editor's Take: After v4.1's price war, DeepSeek's real moat is geographic. Every frontier lab eventually becomes an energy company with a chat interface.
2. iFlytek Ships Spark-Audio-1.0-Preview, Trained Entirely on Domestic Silicon
- Summary: iFlytek (科大讯飞) launched a speech foundation model trained on fully domestically-produced compute, skipping the usual cascaded speech-to-text pipeline so the model keeps tone, emphasis and emotion that transcription discards.
- Source: ITHome · 2026-09-16
- Editor's Take: The export-control angle is the quiet headline. A frontier-class audio model trained without Nvidia proves the domestic stack can carry real workloads, not just demos.
3. Tsinghua and SteadyMind Release LimiX-2, a 400M-Parameter Model That Tops Structured-Data Benchmarks
- Summary: LimiX-2, a foundation model for tabular and structured data co-released with Tsinghua's computer science department, took first place on an international benchmark, scaling parameters to 400M.
- Source: QbitAI · 2026-09-16
- Editor's Take: Everyone chases LLMs; tables still run the global economy. A small model winning a boring category is often where the money actually is.
4. Bilibili Opens an "AI Infinity Arena" — and GPT-6 Leads the First Board
- Summary: Bilibili (哔哩哔哩) launched a public model-arena on Sept 16 where 100 top global models face off in tasks designed by the platform's creators; GPT-6 Astra topped the first leaderboard across 10 reviewer tests.
- Source: QbitAI · 2026-09-16
- Editor's Take: A video site running model evals is very on-brand for China: benchmarks are content now, and whoever owns the leaderboard owns the narrative.
5. The DeepSeek Essay That the Company Had to Walk Back
- Summary: Liu Sheng, lead author of the main attention operator behind DeepSeek v4.1, published an essay titled "I had to bury my talent in yesterday" on Sept 14, which crossed 100k+ reads, hit #1 on Zhihu's trending list, and spread on X. DeepSeek responded that the piece does not represent the company's position.
- Source: ITHome · 2026-09-16
- Editor's Take: A star researcher publicly mourning the half-life of his own skills, then his employer issuing a disclaimer — this is what burnout looks like when it becomes a corporate communications problem.
6. Huawei GTS Teaches Agents to Troubleshoot Networks by Watching Them
- Summary: Huawei's Global Technical Services team deployed an agent system that reads live network state during fault diagnosis, lifting task pass rate by 24.2% and cutting token cost by up to 45% on a dual-firewall scenario that stumped previous automation.
- Source: QbitAI · 2026-09-16
- Editor's Take: Giving the agent eyes into the running system beat giving it a bigger manual. Tool-augmented agents are quietly outperforming model-scaling in enterprise ops.
7. vivo Unveils BlueLM Heart Xiǎo V Pro: Model Plus Harness Plus Safety Architecture
- Summary: At its 2026 developer conference, vivo announced a Blue Heart Intelligence stack built on "large model + harness + agent safety architecture," re-engineering the system experience around on-device agents.
- Source: ITHome · 2026-09-16
- Editor's Take: "Harness" is now a product word in Chinese phone marketing. The industry has conceded that the model alone was never the product.
II. AI Software & Applications
1. After the AI Office Wars: Agent Landscape Shifts as Tencent, ByteDance and Alibaba Fight It Out
- Summary: This summer's Chinese mega-cap battle for AI office productivity ended with the Doubao (豆包) workspace formally launched and the TRAE and Coze teams folded into the Doubao org. Wang Hua of Sinovation Ventures framed the stakes: burning a few billion was survivable, not fighting was not.
- Source: Geekpark · 2026-09-16
- Editor's Take: The losers weren't out-competed, they were reorganized into the winner. In China's agent race, the exit for a failed product is a headcount transfer.
2. Doubao Phone Goes Official — With a 30-Day Cooling-Off Period for Buyers
- Summary: The Doubao AI phone, the ByteDance (字节跳动) and ZTE nubia collaboration that promised to operate any app on the user's behalf, began formal retail sales with an unusual 30-day cooling-off window before the purchase locks in, after last December's hype cycle met several rounds of safety controversy.
- Source: Huxiu · 2026-09-16
- Editor's Take: A return policy for autonomy. The cooling-off period is the most honest user-interface decision an agent product has shipped.
3. Huawei Opens Beta for Xiaoyi Work, Its Cross-Device AI Office Assistant
- Summary: Xiaoyi Work links HarmonyOS phones, tablets and PCs, handling office, coding and design tasks, with a Windows client (v0.6.9) already downloadable from Huawei's site.
- Source: ITHome · 2026-09-16
- Editor's Take: Huawei is betting the assistant, not the app store, becomes the glue of HarmonyOS. If it works, the moat is the account, not the hardware.
4. AI Drug Startup Anew Labs Raises $290M from Sequoia China, Hillhouse and IDG
- Summary: Anew Labs (新生实验室), a drug-discovery AI company traced back to a June spin-out from ByteDance's AI-for-science unit, completed its first independent round of $290M, joined by 5Y Capital, Gaorong Capital and several domestic pharma groups as strategic investors.
- Source: Leiphone · 2026-09-16
- Editor's Take: A giant's research division leaving the building and immediately repricing at nine figures. Spin-outs are how ByteDance's AI-for-science bet cashes out.
5. TrendForce: Global Data-Center Power Demand Hits 161GW This Year, Up 31%
- Summary: Per TrendForce's (集邦咨询, a Taipei-based research house) AI server study released Sept 16, CSPs are accelerating deployment and 2026 global data-center capacity demand reaches 161GW, with the bulk of growth from AI workloads.
- Source: ITHome · 2026-09-16
- Editor's Take: The GPU conversation always ends at the substation. A 31% single-year jump in power demand is a permit-and-grid problem wearing a capex costume.
6. TMTPost Launches TiToken Aggregation Platform and BackBy Compute Optimizer
- Summary: At the 2026 ITValue Summit, media group TMTPost (钛媒体) launched TiToken, a multi-model token relay and routing platform with an overseas arm called Opensand, plus BackBy, a model compute-optimization product.
- Source: TMTPost · 2026-09-16
- Editor's Take: A tech-media company selling token arbitrage tells you where the margin migrated. When publishers become infrastructure brokers, the gold rush is fully monetized at the middleman layer.
7. Jiyuan Lüdong and Infinigence AI Partner on High-Quality Token Supply
- Summary: On Sept 15, AI infrastructure firm TokenRhythm (基元律动) signed a strategic deal with Infinigence AI (无问芯穹), combining agentic infrastructure with multi-model routing to sell guaranteed-quality token supply.
- Source: Leiphone · 2026-09-16
- Editor's Take: "Token quality" is the new SLA. Once inference is a commodity, someone has to grade the water.
III. Humanoid Robots
1. XPeng's IRON Comes Off the Line While Tesla and Figure Chase Volume
- Summary: On Sept 8 an XPeng (小鹏) IRON general-purpose humanoid completed automated final assembly and rolled off the production line, ahead of US rivals still resolving mass-production engineering.
- Source: Geekpark · 2026-09-16
- Editor's Take: China's EV playbook, replayed in two legs: supply chain first, intelligence later. XPeng is trying to win the boring half before the exciting half is solved.
2. Halved Valuations, Open Feuds, Window Guidance: Humanoid Startups Enter Their Reckoning
- Summary: Unitree Robotics (宇树科技) listed on the STAR Market Aug 19 as A-shares' "first humanoid stock," only to watch its price and market cap slide through its first month. Sector-internal fights and regulatory window guidance have followed what Huxiu calls a "tribulation."
- Source: Huxiu · 2026-09-16
- Editor's Take: A public listing is the harshest possible demo environment. The market just repriced "humanoid" from category to company.
3. When IPO Fever Meets a Frozen Primary Market: The "Unitree Phenomenon"
- Summary: An opinion column situates Unitree's post-listing slump inside the broader A-share paradox — subscription lots oversubscribed at near-zero win rates, while the secondary market treats new listings as a flood threat.
- Source: Huxiu · 2026-09-16
- Editor's Take: Useful context for foreign LPs: in China, an IPO print is a liquidity event for the index, not for the believer.
4. Jiangsu Startup Ships a "Killer Whale" That Cleans Ship Hulls 50x Faster Than Divers
- Summary: Shihang Intelligent's (世航智能) underwater inspection robot cleaned a barnacle-loaded hull in 10 hours versus two weeks of manual labor, and owners are doing the arithmetic: a cleaner hull recovers about 2 knots of average speed per trip.
- Source: Huxiu · 2026-09-16
- Editor's Take: The capital wave chasing "the next Unitree" is heading underwater. Boring, metered ROI beats stage demos, and investors have finally noticed.
5. The Truth Behind Robots Playing Table Tennis: Borrowed Eyes, Fake Brains
- Summary: A widely-shared teardown argues that viral humanoid ball-striking demos rely on external motion capture, precomputed trajectories and heavy latency hiding, and that few systems actually perceive and plan autonomously in real time.
- Source: Huxiu · 2026-09-16
- Editor's Take: Every hype cycle needs its debunking industry. Chinese robotics now has both at scale, which is arguably a sign of health.
6. JD Bundles an AI Diet Planner, a Produce Supply Chain and a Cooking Robot
- Summary: JD.com (京东) launched "Qixian Chef," a home meal service pairing an AI health-diet plan with pre-portioned grocery delivery and a smart wok robot; pre-sales opened in Beijing, Shanghai, Guangzhou and Shenzhen.
- Source: ITHome · 2026-09-16
- Editor's Take: Nobody sells a robot; everyone sells the meal. The appliance is the least interesting part of the bundle — the recurring grocery order is the business.
IV. Autonomous Driving
1. Luxeed RX Gets L3 Road-Testing Plates; Yu Chengdong Says City Expressways Are the Proving Ground
- Summary: Harmony Intelligent Mobility's (鸿蒙智行) Luxeed RX received an L3 autonomous driving road-test license and has been spotted testing on public roads, with validation focused on urban expressways, according to Huawei's Richard Yu.
- Source: ITHome · 2026-09-16
- Editor's Take: L3 is where liability becomes legal philosophy, and Chinese OEMs are running at it. Expressways are the easiest hard place to prove the insurance model works.
2. Hesai Unveils the World's Smallest High-Resolution LiDAR at InterGEO Munich
- Summary: Hesai (禾赛) launched its MT series mapping LiDAR in Munich with 3.5mm ranging precision and a 70% smaller spot size, aimed at high-precision modeling rather than driving alone.
- Source: ITHome · 2026-09-16
- Editor's Take: A Chinese LiDAR maker headlining a German geospatial trade show is the new normal. Hesai is deliberately escaping the auto cycle into surveying and robotics.
3. New Galaxy E5 Adds LiDAR and 245kW Power, Launches Sept 23
- Summary: Geely's (吉利) updated Galaxy E5 electric crossover appeared in MIIT filings ahead of a Sept 23 launch, adding a LiDAR and raising output to 245kW, alongside new fast-charging tech.
- Source: ITHome · 2026-09-16
- Editor's Take: LiDAR in a mainstream crossover means sensor pricing finally broke. The premium features of 2024 are the spec-sheet hygiene of 2026.
V. World Models / Physical AI
1. Shanghai Startup Releases Cog-WM 1.0, the First "Brain-Inspired" Cognitive World Model
- Summary: At the 2026 Pujiang Innovation Forum's brain-inspired embodied intelligence session on Sept 14, JuNao Panshi (具脑磐石) released Cog-WM 1.0, a latent-space predictive world model that lets quadruped and wheeled-humanoid robots navigate and manipulate without prebuilt maps; the team claims navigation success up 10%+ over the BSC-Nav baseline and manipulation up to 16% over π0.5.
- Source: Jiemian News (via Baidu) · 2026-09-14
- Editor's Take: LeCun's JEPA bet is getting Chinese entrants who cite the same numbers humans run on. If map-free navigation holds in real buildings, this route leapfrogs the data-hoarding paradigm entirely.
Sector count: Foundation Models 7 · Software & Applications 7 · Humanoid Robots 6 · Autonomous Driving 3 · World Models / Physical AI 1 · Total 24.
物界前沿 | EastVoice
