EastVoice · AI · 2026-10-23 · Issue #106

Editor's Note: Thursday's China story was scale, in three places. DeepSeek closed a round of roughly ¥100 billion (~$15 billion) at a ¥500 billion pre-money valuation, six months after raising ¥51 billion. Ant Group shipped a decision model it says cuts token use by more than 90%, hours before iFlytek put a speech foundation model and a domestic-compute model on stage at its 1024 developer festival. On the hardware side, Unitree marked ten years with a 10,000 to 20,000 unit G1 milestone, AgiBot signed the biggest single office lease Shanghai has ever recorded, and humanoid rental rates are down about 70% from early 2025.

Commentary: The funding number is the headline. The number that decides the next two years is how much of that capital reaches production floors rather than model leaderboards.

---

I. AI Foundation Models

1. DeepSeek Raises About ¥100 Billion at a ¥500 Billion Pre-Money Valuation

• Summary: DeepSeek closed a new round of roughly ¥100 billion (~$15 billion) at a pre-money valuation near ¥500 billion, Bloomberg reported. The investor list leans industrial, with Shixi Capital, SMIC Capital, CPE, Legend Capital and Geely joining, and existing holders Tencent, CATL and IDG following on. Together with the ~¥51 billion first round delivered in June, the company has raised around ¥150 billion in six months.
• Source: ITHome · 2026-10-22
• Editor's Take: The money is the smaller signal. The cap table now reads like a supply-chain roster, and rounds of this shape get priced on what the strategic buyers can do with the model, not on API revenue. It also lands two days after Washington ordered federal agencies to strip DeepSeek from their systems.

2. Ant Group's LingDT-Decision-1.0 Claims a 90% Cut in Token Cost

• Summary: Ant Group's technology arm released LingDT-Decision-1.0, a decision model aimed at structured, high-frequency business decisions. It ships in 4B and 27B versions, and the company says it beats Microsoft and Jev decision models on accuracy in those tasks while cutting token consumption by more than 90% and returning answers faster.
• Source: ITHome · 2026-10-22
• Editor's Take: A decision model is a cost argument, not a capability flex. If the accuracy claim holds on customer data, the buyers are procurement teams comparing per-call prices, which is where Chinese models have been winning.

3. iFlytek Launches SparkAudio 1.0, a Speech Foundation Model for the Keyboard

• Summary: At its 1024 developer festival, iFlytek launched SparkAudio 1.0, a speech foundation model that switches between Chinese and English within a sentence and identifies speakers in real time. The company says it is usable straight from its input method rather than as a separate app.
• Source: ITHome · 2026-10-22
• Editor's Take: Shipping a voice model inside an input method is the distribution trick iFlytek owns. Nobody wants another voice app; everyone types.

4. Spark X2.6 Puts Code and Agent Tasks on All-Domestic Compute

• Summary: iFlytek also released Spark X2.6, an end-cloud model it says is the first to break through on coding and agent tasks with the compute stack entirely domestic. Chairman Liu Qingfeng told the same event that domestic compute still trails the global frontier by roughly two generations, and that he expects China to close it.
• Source: ITHome · 2026-10-22
• Editor's Take: "Two generations behind" from the company selling the stack is the most useful sentence of the week. It sets a floor for what claims about parity should be measured against.

5. Mihaoyo, Tsinghua and Peking University Open-Source GameScaling, 0.8B to 27B

• Summary: A group spanning Mihaoyo's Honkai AI team plus Tsinghua, Peking University, UCAS, HKU and NUS released GameScaling, an open-source family of game models covering 0.8B to 27B parameters.
• Source: ITHome · 2026-10-22
• Editor's Take: Games are the cheapest place to get long-horizon agent trajectories. Putting the weights out means the benchmark becomes shared ground rather than a marketing slide.

6. China Telecom Open-Sources an Intent Model That Handles Both Test Formats

• Summary: China Telecom's Star AI lab released Xing4.0-29B-Intent, an intent-understanding model built to handle multiple-choice and fill-in-the-blank style tasks in one model, and published the weights.
• Source: QbitAI · 2026-10-22
• Editor's Take: A telecom carrier open-sourcing an intent model is a bid to own the layer between the user and the app. The weights are the cheap part; the operator's billing channel is the point.

7. MiniMax Falls Below Its IPO Price

• Summary: MiniMax shares dropped as much as 7.3% intraday to HK$159.3, below the HK$165 IPO price, giving back every gain since its listing in January.
• Source: ITHome · 2026-10-22
• Editor's Take: The listing gave Chinese model labs a public price. It is now lower than where the first outside buyers got in, and the next lab to file will be priced against that.

8. Xiaohongshu Merges Its Dots AI Unit Into Core Infrastructure

• Summary: An internal letter dated October 16 merged Xiaohongshu's first-tier AI department Dots with the company's core infrastructure team, the second restructuring of its AI organisation in six months. The letter argues that model capability and compute scale matter less than how tightly the model is fused with the product.
• Source: TMTPost · 2026-10-22
• Editor's Take: Folding AI into infra is an admission that the model is not the product. For a company whose moat is content ranking, that is the more honest structure.

9. Volcano Engine Opens Seedance 2.5 4K to 20 Studios and Directors

• Summary: ByteDance's Volcano Engine started first-batch enterprise trials of Seedance 2.5 4K, supplying the video model to around 20 studios and directors worldwide on a per-title basis. The pitch is finer facial performance, better generation consistency and use on large screens.
• Source: ITHome · 2026-10-22
• Editor's Take: Gating a 4K model by invitation is a pricing strategy, not a capacity limit. ByteDance is testing whether film-grade output can be sold per project instead of per token.

---

II. AI Software & Applications

1. The National AI Fund Says It Has Deployed ¥60 Billion

• Summary: MIIT told a State Council briefing that the ¥60 billion national AI industry investment fund has already placed money into a first batch of projects. The same briefing put China's core AI industry at more than ¥1.2 trillion in 2025, with AI in use at over 30% of large manufacturers.
• Source: TMTPost · 2026-10-22
• Editor's Take: State money at this scale sets the floor under valuations. Private investors can now treat the fund as the anchor buyer in every serious round, which cuts both ways for pricing.

2. Huawei Takes a ¥15 Billion Offshore Syndicated Loan

• Summary: Huawei secured a ¥15 billion offshore syndicated loan from six Chinese banks, its first offshore syndicated facility since 2020. The seven-year loan is denominated in offshore yuan and is for general corporate purposes, with Bank of China, China Construction Bank and ICBC among the lenders.
• Source: ITHome · 2026-10-22
• Editor's Take: Offshore funding coming back at a seven-year tenor is a statement about how the banks read the sanctions risk. Long money does not show up where the lenders expect a refinancing wall.

3. Token Packages Reach the Double 11 Shelf, from ¥24.5

• Summary: Tmall's Double 11 sale, which opened on the evening of October 20, put token bundles on the shelf for the first time, sold as monthly, quarterly and annual cards with the cheapest tier at ¥24.5, among roughly 72 million listed products.
• Source: TMTPost · 2026-10-22
• Editor's Take: Tokens as a consumer SKU is how inference becomes retail. It also drags model pricing into the same discount calendar as detergent.

4. Domestic AI Chipmakers Post Strong Quarters, and Move the Race Off the Chip

• Summary: Huxiu, citing Economic Observer, reports that Chinese accelerator makers are posting broadly strong third-quarter results, with MetaX guiding to year-on-year revenue growth. The competitive ground has shifted from the silicon itself to software migration, system efficiency and supply assurance.
• Source: Huxiu · 2026-10-22
• Editor's Take: A good quarter does not retire the migration problem. Buyers who have to rewrite kernels to save money will ask what happens when the next generation lands.

5. A Week of AI Hardware: Qwen Glasses at ¥2,218 and a StepFun Agent Phone

• Summary: On October 13, Alibaba's Qwen launched N1 and N1 Pro AI glasses plus a Bose co-branded ear-clip headset, priced from ¥2,218 for the glasses and ¥1,099 for the headset after subsidies. The same evening, StepFun launched its agent phone in Shanghai.
• Source: TMTPost · 2026-10-22
• Editor's Take: The AI hardware wave is being priced like consumer electronics, not like developer kits. That fixes the volume expectations early, and it is why the glasses cost less than the phone they pair with.

6. Qi Junyuan's Seven-Month-Old Today AI Has Raised Over $200 Million

• Summary: Today AI, founded seven months ago by serial entrepreneur Qi Junyuan, disclosed cumulative funding of more than $200 million, the largest early cheque for a personal-agent startup in China so far.
• Source: Huxiu · 2026-10-22
• Editor's Take: Personal agents have a distribution problem, not a model problem. Funding this fast is a bet on the founder's operator history rather than on a shipped product.

7. China Has 323 Million People Over 60. Companion AI Still Has to Prove It Sells

• Summary: Huxiu works through the economics of AI companionship for the elderly. One case it cites, reported by Legal Daily in July, is a 71-year-old woman in Beijing living alone with high blood pressure whose daughter paid more than ¥20,000 for a companion robot.
• Source: Huxiu · 2026-10-22
• Editor's Take: Care is a service business with hardware attached. A device bought once by an anxious child does not create a recurring revenue line, and the industry has not answered what does.

8. Alibaba's Accio Becomes a Double 11 Toolbox for Merchants

• Summary: Alibaba's Accio, the e-commerce AI workbench sold to merchants, is being used through Double 11 for cross-platform store diagnostics, product selection and sales analysis. The company claims merchants cut their pre-sale preparation period by half a month and can retune a hit product across platforms in about 15 minutes.
• Source: Leiphone · 2026-10-22
• Editor's Take: Selling a workbench rather than a model is the cheapest way for Alibaba to monetise its own LLM work. Merchants buy outcomes, and outcomes are what Double 11 measures.

---

III. Humanoid Robots

1. Unitree Turns Ten: Three Directions, a Contest, and 10,000 to 20,000 G1 Units

• Summary: At its partner ecosystem conference in Hangzhou, Unitree founder Wang Xingxing laid out three priorities: building mass-production capability for demand in the tens of millions to hundreds of millions of units, combining in-house development with partnerships, and expanding the ecosystem. He said the G1 has shipped about 10,000 to 20,000 units since May 2024, making it the best-selling humanoid in the world, and that Unitree will spend tens of millions of yuan a year on a world Physical AI contest. His framing for the moment: physical AI is already showing signs of being broken through.
• Source: ITHome · 2026-10-22
• Editor's Take: Ten thousand units at a hardware company is a real number; the contest is the cheaper half. Subsidising a competition buys a talent pipeline and a benchmark, and neither costs what a factory line does.

2. Surgical Robot Exports Hit ¥672 Million, Up 272.9% in Eight Months

• Summary: Customs data cited by CCTV Finance puts China's surgical robot exports at ¥672 million for January to August 2026, against ¥180 million a year earlier. Intelligent bionic robots led the 13 tracked categories at ¥1.168 billion, with Guangdong and Shanghai out front.
• Source: ITHome · 2026-10-22
• Editor's Take: Surgical robots are the export category Chinese robotics actually earns hard currency in. Class 3 device approvals are the gate, and hospitals abroad are the demand signal nobody markets.

3. AgiBot Signs Shanghai's Largest Ever Single Office Lease

• Summary: AgiBot and related entities signed more than 220,000 square metres of office space across three Lujiazui Group projects in Shanghai, including Qiantan Yuan, Qiantan Shang and Yuqiao Wisdom Valley. Reports put it as the biggest single office lease in the city's history, signed by a company less than four years old.
• Source: ITHome · 2026-10-22
• Editor's Take: 220,000 square metres is a headcount plan, not a real estate story. Someone at AgiBot thinks the hiring curve keeps going up for years.

4. Humanoid Rental Rates Have Fallen About 70%

• Summary: Huxiu reports that a fully staffed humanoid rental cost ¥10,000 a day in early 2025, while by October 2026 a Unitree U2 with technicians included rents for under ¥3,000 a day. On October 9, UBTech signed a strategic agreement with Jianfa.
• Source: Huxiu · 2026-10-22
• Editor's Take: Two-thirds off in eighteen months turns rental into a marketing channel. The vendors pushing it are selling demo capacity, and demos are not paid work.

5. A Robot Is Cheap. Making It Work Is Not.

• Summary: Huxiu talks to a second-development service provider that quoted more than ¥1 million in hardware and software for a single factory palletising and depalletising deployment, covering one robot body plus integration and delivery.
• Source: Huxiu · 2026-10-22
• Editor's Take: Integration is where embodied money actually moves today. The body is a commodity with a spec sheet; the hours are not.

---

IV. Autonomous Driving

1. MIIT Says the Second Batch of L3 Pilot Approvals Is Underway

• Summary: MIIT's Guo Shougang said at the World Intelligent Connected Vehicle Conference that assisted driving has reached scale worldwide and autonomy is entering commercial rollout. The second batch of L3 market-access pilots is under review now.
• Source: ITHome · 2026-10-22
• Editor's Take: Approvals are the product in Chinese autonomy. Whoever clears the second batch gets the reference case every other OEM will cite in its own filing.

2. Changan Puts L3 Mass Production in 2027 and L4 Commercial Operation in 2028

• Summary: Changan Automobile's chairman Zhao Fei told the same conference that he expects the industry to reach L3 mass production in 2027 and commercial L4 operations in 2028.
• Source: ITHome · 2026-10-22
• Editor's Take: A 2027 L3 date is close enough to be checkable. That makes it a commitment rather than a forecast, and Changan's suppliers will be held to it.

3. Audi Is in Talks With Huawei on the Next China-Market A8

• Summary: Germany's Manager Magazin reported that Audi is negotiating with Huawei to co-develop the next-generation A8 for China, alongside details of Audi's difficulties in both combustion and electric sales there, a change of its China management and hopes of reviving the FAW joint-venture plant.
• Source: ITHome · 2026-10-22
• Editor's Take: A flagship sedan is the last place a premium brand wants to rent its brain. Doing it anyway says how much of the Chinese premium market now runs on domestic driver-assistance stacks.

4. Beijing Moves to Regulate the Blue Light on Driver-Assistance Cars

• Summary: MIIT opened public comment on October 20 for a revised standard on exterior lighting and light-signalling devices, which brings the blue marker light used by driver-assistance systems into regulation for the first time. The lights were halted in July after four years without a governing rule.
• Source: Huxiu · 2026-10-22
• Editor's Take: Standardising the indicator light settles who is visibly driving. The rule matters less for drivers than for accident reconstruction.

5. The National AI Fund's First Autonomous Driving Bet Is New Stone

• Summary: Delivery-robot company New Stone closed a Series E of nearly $400 million, co-led by Boyu Capital and the National AI Industry Investment Fund, the largest autonomous driving round in China this year.
• Source: Huxiu · 2026-10-22
• Editor's Take: The state fund picked unmanned delivery over robotaxis. Fixed routes with commercial customers already attached are the deal that can be underwritten today.

---

V. World Models / Physical AI

1. By 2027, a Frontier AI That Cannot Control Robots Will Not Count as Frontier

• Summary: Huxiu carries the argument that embodied intelligence has stopped being a closed-door exercise. If physical AI is the path to the frontier, the piece argues, OpenAI, Anthropic and Google will all try to drive across it, and DeepSeek is already positioning for it.
• Source: Huxiu · 2026-10-22
• Editor's Take: The claim inverts the usual ordering. Labs have been treating embodiment as an application of models. The version where it becomes the test of a model is a different research budget.

2. Who Supplies OpenAI With Embodied Data?

• Summary: QbitAI traces the companies feeding real-robot data into OpenAI's physical AI work and how they got onto the global supply chain for embodied datasets.
• Source: QbitAI · 2026-10-22
• Editor's Take: Data suppliers are the least visible and most defensible position in physical AI. Models change quarterly, a labelled manipulation dataset does not.

3. Embodied Memory Is Warming Up, and IROS 2026 Set a New Long-Memory SOTA

• Summary: Leiphone reviews the best-paper work from IROS 2026 and argues that robots without long-term memory will be left behind, with a new state of the art reported on long-horizon embodied memory.
• Source: Leiphone · 2026-10-22
• Editor's Take: Memory is the part of embodied learning that hardware cannot buy. It also decides whether one deployment teaches the next one anything.

---

Physix Frontier | EastVoice · 29 items: 9 foundation models, 8 software, 5 humanoid, 5 autonomous driving, 3 world models / physical AI.