PHYSIX FRONTIER EASTVOICE
2026.08.28 · Friday
Issue 022
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
$900M+
XPeng's robot unit landed a record round
283%
MiniMax posted revenue up in its first half
100% faster
Alibaba pushed Qwen generation speed down again
HK$80B
Alibaba raised through a share placement for AI
Top Stories 3 · dual commentary ｜ Section Picks 18 · one-line takes
Zhipu open-sourced GLM-5.3-Flash, MiniMax reported 283% revenue growth, and XPeng Robotics secured a $900M+ funding round.
The mystery model "Ox-Alpha" (nicknamed Niu Lai by the Chinese community) that topped OpenCode and OpenRouter this week turned out to be Zhipu's GLM-5.3-Flash, released and open-sourced on Aug 26. It's the first natively multimodal model in the GLM-5 line: 320B total params with 18B active, scoring 57 on the Artificial Analysis Intelligence Index and beating GLM-5.2. SenseTime's SenseCore platform is handling deployment on domestic accelerators.
Editor's take｜Open-sourcing the frontier and running it on Chinese silicon is a statement. The real test is whether overseas devs actually run it, because popularity is not adoption.
MiniMax (0100.HK) posted $116.7M revenue for H1 2026, up 283% year over year, already ahead of its $79M full-year 2025 number. Enterprise revenue grew 7x and now out-earns the consumer side, with ARR up roughly 500% and token volume about 2,000%.
Editor's take｜Half a year ago MiniMax was a consumer chatbot story. Now the B2B line pays the bills. The consumer race ended faster than anyone expected.
AI infrastructure startup TokenRhythm closed a new round of tens of millions of dollars, led by Honghui Capital with Juhe Capital and others joining. The company is building a unified model-routing gateway, positioned as China's answer to OpenRouter.
Editor's take｜When models commoditize, the money moves to the plumbing. Routing is a thin-margin business, but in China it's also the entry ticket to the AI app layer.
18 AI stories worth your time, each with a one-line take.
Alibaba's Qwen Office is the first product to ship Qwen3.8-Flash, with a new Standard mode aimed at everyday tasks. Single-task generation is about 100% faster and token consumption drops 75% on average.
What it means for you｜Single-task generation is about 100% faster and token consumption drops 75% on average in Alibaba's office suite.
One-line take｜Same model, half the tokens, twice the speed. That's how Alibaba fights for the office market without starting a price war.
Alibaba Cloud reduced the per-token price for Qwen3.8-Flash on its Bailian model platform, effective noon Beijing time Aug 27. The cut follows the model's debut in the company's office products a day earlier.
One-line take｜Qwen pricing moves in one direction only. The message to competitors: keep up with our cost curve or get out of the way.
DeepSeek raised peak-hour output pricing on V4-Pro from ¥6 to ¥27 per million tokens (+350%), effective Aug 17. The move came weeks after V4 Flash's $0.027-per-task cost drew a widely shared "kill line" on Artificial Analysis, with roughly 70% of global models sitting above it on price-performance. Analysts now debate who survives the redraw.
One-line take｜The killer feature was always the price. A 350% hike is the industry's quiet admission that free inference was never sustainable.
Alibaba launched a new Qoder, repositioning it from an AI coding tool into an agent workbench. You describe a goal in natural language; Qoder plans, calls tools, executes and verifies, and you can watch or take over at any step. Non-developers get the front seat.
What it means for you｜Non-developers get the front seat to describe goals in natural language while Qoder plans, calls tools, executes and verifies.
One-line take｜The interesting part isn't the agent, it's the framing: a user who "doesn't start from code." The IDE is no longer the front door to software.
Infant-formula maker Feihe, working with Volcano Engine, has deployed 436 agents on its production lines. The company says the question is no longer whether to adopt AI, but which scenarios deserve scale, with 10+ smart factories and over 90% digital equipment as the substrate.
One-line take｜A dairy company running 400-plus agents in production is the counterexample every "AI is only for tech" skeptic needs. The boring industries are where agent economics get proven.
Beijing's Chaoyang district market regulator fined an employee ¥100,000 in the city's first administrative penalty involving AIGC trade secrets. The worker copied the company's AI software and took it to a new employer.
One-line take｜The fine is pocket change; the signal isn't. AI code is now officially a trade secret worth litigating in China.
NetEase Youdao opened pre-orders for OpenPods, billing it as the first agent earbuds designed for iPhone users, at a subsidized ¥1,499. General sales start Sep 10 in black and white.
One-line take｜Earbuds are becoming the wearable agent interface. Whether Apple notices the invasion of its accessory moat is the fun part.
Hubei-based Xingchen Technology raised close to ¥5 billion in a Series A, a record for China's advanced packaging sector, with plans for 100k wafers per month and a ¥10B+ revenue target. Phase 2 of its 3D packaging pilot line goes live in September as Huawei's Tau law picks up industry traction.
One-line take｜Packaging is where China's chip strategy now has a credible shot at leadership, and a ¥5B round says the money agrees.
A founder born in 2003 turned his cross-border e-commerce hustle into a kids' product: an AI magnifying glass that recognizes objects and speaks their names. Point it at an ant, it says "that's an ant," and a toddler laughs.
What it means for you｜Point it at an ant, it says "that's an ant," and a toddler laughs.
One-line take｜The most human story on today's list. Sometimes the right AI product is just a toy that says "that's an ant."
TMTPost lays out the bill. Nvidia teamed with six major financial institutions to build a $500B+ financing platform for AI infrastructure, then guaranteed up to $105 billion for OpenAI's Ohio data center project. Alibaba raised HK$80 billion through a share placement on Aug 23, with proceeds earmarked for chips and AI.
One-line take｜The AI buildout has stopped being a tech story and become a capital-markets one. Chip vendors are now bankers, and the bill keeps landing on someone else's page.
XPeng's robot business closed its first private round at more than $900 million, with a post-money valuation above $6.3 billion (~¥43B). It's the largest single private round in China's embodied-intelligence industry.
One-line take｜Investors are now pricing robot companies like model labs, on software potential rather than unit shipments. That's both a validation and a warning.
A Leiphone field report from WRC argues the gap between robot demos and production lines isn't model quality but the unglamorous work: integration, tooling, fault handling, and talking to factory IT systems older than the iPhone.
One-line take｜WRC booths sell the dream; this article sells the spreadsheet. The winners will be the companies that hate on-site debugging the least.
TMTPost reports Tesla is still installing its Optimus Gen line in a Fremont plant that was emptied of Model S/X production just 46 days ago. Figure, meanwhile, has already delivered 350 humanoid units.
One-line take｜The gap between keynote promises and delivered units keeps widening. 350 shipped robots beat any amount of unveiled product.
Pollen Robotics, Hugging Face's robotics partner, opened pre-orders for Microduck, a $399 one-eyed duck robot under 10 inches tall that can pick up objects and has its own voice. It ships before Christmas in cream, graphite, or lavender.
One-line take｜$399 won't buy a workforce. It buys curiosity about open-source robots, which could be worth more in the long run.
At its Physical AI sharing event, XPeng said its Gen-2 VLA now covers L2 through L4 and its Robotaxi has won Guangzhou's remote-driving test qualification, moving into key driverless-front road validation.
One-line take｜"Driverless-front" validation with a remote-testing permit is China's answer to Waymo's on-road dominance. The regulatory door is opening, slowly and region by region.
XPeng chairman He Xiaopeng said the company hopes to secure European regulatory approval in 2027 and begin deliveries of the second-generation VLA to overseas customers, the first public date on XPeng's international autonomous-driving push.
One-line take｜China's smart-driving stack is eyeing Europe before some European brands have their own L2+ story straight. Ambitious, and European regulators will have opinions.
Humanoid maker Leju released a vertical embodied model for industry, reporting about 2x efficiency gains in pilot production lines and first place on two benchmark leaderboards, built on a pretraining-first approach.
One-line take｜Vertical world models are the monetizable slice of physical AI. General-purpose world models stay a science project; factory workflows pay rent today.
Ahead of its Physical AI day, XPeng said its third-generation Turing AI chip has been powered on, part of a push to put a "super agent on wheels" through VLA+VLM cross-domain fusion.
One-line take｜A carmaker designing its third AI chip in-house means the compute stack is now the product. Watch what that does to the hardware margin story.
📖 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 →