PHYSIX FRONTIER EASTVOICE
2026.10.22 · Thursday
Issue 065
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
2,000
AI agents run in parallel by OpenAI and PerturbAI
¥131.85 billion
Alibaba, Tencent and Baidu quarterly capital spending
¥14
Tencent Cloud Yuanbao Search per 1,000 calls
142 million
AI office tools monthly users, Tencent leads
Top Stories 3 · dual commentary ｜ Section Picks 24 · one-line takes
OpenAI and PerturbAI's 2,000 agents, Alibaba Tencent Baidu's ¥131.85 billion quarter, and Tencent Cloud's ¥14 search API.
PerturbAI said Wednesday that with OpenAI and Scripps Research it built a reasoning framework that runs close to 2,000 AI agents in parallel to read experimental data, search the literature and propose hypotheses for the lab to test.
Editor's take｜The pitch here is throughput, not accuracy. Whether the hypotheses survive a wet lab is a question the announcement does not answer.
Source: ITHome (2026-10-21)
TMTPost puts combined capital spending at the three Chinese cloud and model giants at ¥131.85 billion in one quarter, up 114% year on year. Alibaba alone accounted for ¥67.68 billion, up 75%.
Editor's take｜Capex is now the clearest statement of intent in Chinese AI. None of the three is waiting for demand to be proven before building.
Source: TMTPost (2026-10-21)
Tencent Cloud opened Yuanbao Search to outside developers, a stack it says runs on an index in the hundred-billion range, priced from ¥14 per 1,000 calls.
Editor's take｜Metered search is how a model feature becomes an infrastructure line item. Tencent is selling the dependency, not the assistant.
Source: ITHome (2026-10-21)
24 AI stories worth your time, each with a one-line take.
Huxiu looks at how the Xiaohuanxiong app has become the most visible product SenseTime has, in a year when alumni spinouts were the company's better-known calling card.
One-line take｜A model lab's brand is only as good as its last shipped product. A consumer app with real daily users beats a research citation.
Source: Huxiu (2026-10-21)
Leiphone put MiniMax M3.1 Flash through its own tests after the model took first place on an agentic task benchmark, positioning it against the rest of the Chinese flagship field.
One-line take｜Benchmark leadership rotates monthly in China. What holds value is whether the model survives the long-horizon tasks agents actually run.
Source: Leiphone (2026-10-21)
TMTPost works through iFlytek's claims against its own numbers, after chairman Liu Qingfeng argued the market undervalues the company and its domestic-chip route.
One-line take｜A domestic-chip strategy is a long position with no hedge. The market prices that risk below the story, and it is not obviously wrong.
Source: TMTPost (2026-10-21)
At a State Council press briefing, MIIT said it is drafting policies for a new generation of smart manufacturing built around AI. A new centrally owned national data group was announced at the same time, with its chairman and general manager named.
One-line take｜Data is where the state is consolidating first. Ownership gets settled before the market is allowed to bid on it.
Source: TMTPost (2026-10-21)
A team at HKUST open-sourced what it calls a fourth kind of university, a venue where communities of agents and ordinary users build things together.
One-line take｜Open-sourcing the venue instead of the model is a cheap way to get other people to do the interesting work.
Source: AI Era (2026-10-21)
Sensor Tower data cited by Huxiu points to Catbox, not Doubao, as the app bringing in the most revenue inside ByteDance's AI portfolio. User scale and revenue scale are pulling apart.
One-line take｜Companion apps have paying users; assistants mostly have free ones. ByteDance now owns one of each and knows which is which.
Source: Huxiu (2026-10-21)
ByteDance updated the blocked-app list for the Doubao Phone Assistant, cutting off banking, trusts, payments, internet finance and credit. The change follows the screen-automation declaration the assistant published for agents.
One-line take｜An assistant that can move money is a liability, not a feature. Blocking the category is the cheapest way to stay shipped.
Source: ITHome (2026-10-21)
QuestMobile's third-quarter report puts combined monthly users for AI office products at 142 million, up 50.6% year on year. Tencent WorkBuddy leads at 32.35 million, ahead of Baidu's Dazi at 9.77 million.
One-line take｜A three-to-one lead in a category this young is Tencent's to lose, and it will be defended with distribution rather than model quality.
Source: TMTPost (2026-10-21)
TMTPost contrasts the two plans. Alibaba keeps funding data centres inside a three-year cloud and AI programme its CEO says will overshoot the ¥380 billion headline. JD has committed ¥10 billion to robotics by 2028 and 80 related sites.
One-line take｜Both are infrastructure bets. One buys accelerators, the other buys floorspace and arms.
Source: TMTPost (2026-10-21)
More than ten mainstream brokerages have connected to third-party AI platforms including Kimi, Tencent WorkBuddy, Alibaba Qwen and Volcano Engine, shifting from in-house builds to outside ecosystems.
One-line take｜Finance was the vertical everyone expected to build in-house. Renting got there first, and compliance never was the moat.
Source: Huxiu (2026-10-21)
QbitAI profiles a multi-agent medical system from a company incubated at Tsinghua AIR, in which an assistant takes patient history and symptoms before a clinician picks up the case.
One-line take｜Intake is the part of care that was already a checklist. Automating the form is less glamorous than diagnosis and far easier to deploy.
Source: QbitAI (2026-10-21)
Ahead of Double 11, Doubao added Douyin live-commerce rooms to its product recommendations so users can order without leaving the assistant, and began pushing shopping coupons to users.
What it means for you｜Users can order without leaving the assistant and receive shopping coupons.
One-line take｜The assistant is turning into a storefront with a chat box attached. That answers a monetisation question, not a product one.
Source: Huxiu (2026-10-21)
TMTPost reports that Fengzhushou signed an AI computing order in the ¥10 billion range while holding about ¥300 million in cash, with plans to buy more than ¥7 billion of AI servers. The board has also approved guarantees for a subsidiary's borrowings.
One-line take｜Order books and balance sheets have never been further apart in Chinese AI. Whoever finances this carries the credit risk, not the customer.
Source: TMTPost (2026-10-21)
Unitree unveiled the H2-X, a humanoid it markets as a personal bodyguard with a maximum runtime of about 8 hours. In the demo it lifts a car and kicks a car door in.
One-line take｜Lifting a car is a stunt. Eight hours is a shift. Buyers will ask about the second number before the first.
Source: ITHome (2026-10-21)
Moja Robotics, the robot unit under Chery, launched Mobao, its first consumer home companion humanoid, at 0.9 metres tall with 22 degrees of freedom for walking and body expression. Moja says it has sold more than 3,000 units across 100-plus use cases in 60-plus countries.
One-line take｜A carmaker selling a companion robot is a channel decision. Chery has service centres and dealers, and that is what a home robot actually needs.
Source: ITHome (2026-10-21)
Gamma Robotics, the first market-facing company incubated by China Telecom's AI research institute TeleAI, registered in Shanghai's Xuhui district. It will work on full-model self-evolving humanoid robots with TeleAI.
One-line take｜A telecom carrier incubating a robot company follows the same logic as a carmaker doing it. Distribution and deployment sites already exist.
Source: ITHome (2026-10-21)
Geek+ detailed Gravity Control, a method that lets operators correct a robot in plain language, which it says lifted out-of-distribution task success from 5% to 95%.
One-line take｜Human correction in natural language is the cheapest supervision there is. The claim is only as good as the size of the OOD test set behind it.
Source: AI Era (2026-10-21)
At the World Intelligent Connected Vehicle Conference on Wednesday, Nio CEO Li Bin said the company's driver-assistance stack, built on a world-model architecture with closed-loop reinforcement learning, is deployed on more than 950,000 vehicles, with more than 9.14 billion kilometres of assisted driving recorded.
One-line take｜Fleet size is the input to a data flywheel, not evidence of autonomy. Nio is publishing the input, which is the honest number to publish.
Source: ITHome (2026-10-21)
Hesai published details of a next-generation 6D full-colour lidar, claiming native pixel-level alignment of RGB and depth on the chip. Its framing is that autonomy and embodied AI now want high-confidence ground truth more than more scan lines.
One-line take｜Selling colour instead of line count means selling to data pipelines. That is a sensor market talking to robot training, not to car spec sheets.
Source: ITHome (2026-10-21)
At the Physical AI and World Models conference this week, Momenta CEO Cao Xudong argued world models are accelerating the autonomy version of Moore's law and predicted 2028 as the inflection point for scaled L4 deployment.
One-line take｜A 2028 date from a company selling into that market is a planning assumption, not a forecast. The useful part is which inputs it depends on.
Source: ITHome (2026-10-21)
IDC tracked AI cockpit solutions as a standalone add-on market for the first time in its China automotive cloud report. Volcano Engine took the largest share in the first half of 2026, with Alibaba second.
One-line take｜A cockpit assistant is a cloud contract with a steering wheel attached. The first-mover list here is a cloud ranking wearing a car badge.
Source: ITHome (2026-10-21)
Shengshu Technology's world-model team started a public, continuous livestream of the full training run for Motubrain 2.0 on Wednesday, scheduled to last half a month. It follows Xiaomi's MiMo team publishing reinforcement-learning progress in the open.
One-line take｜Publishing the training run is marketing that cannot be faked. It also sets a public baseline for how quickly the next model has to improve.
Source: QbitAI (2026-10-21)
QbitAI profiles Qianjue, a Tsinghua-linked company working on latent-variable identifiability in world models, which it says predates comparable LeCun work by a year. The approach leans on physical rules and cross-embodiment reuse rather than predicting video frames.
One-line take｜Skipping frame prediction to model physical structure is a narrower bet. It is also the only version that transfers across bodies.
Source: QbitAI (2026-10-21)
Leiphone surveys the case for "live training platforms" that put robot learning on real workstations. It cites a CAICT report estimating that general embodied models need tens of millions of hours of real interaction data, while globally available high-quality real-robot data sits in the millions.
One-line take｜The data gap is the whole argument for training in production. Whoever owns the workstation owns the training set.
Source: Leiphone (2026-10-21)
📖 The full EastVoice issue on the Chinese forum