EastVoice · AI · 2026-09-13 · Issue #66

Editor's Note: 月之暗面 / Moonshot AI shipped Kimi K2.8 with flagship-level claims and a $1B ARR disclosed the day before a rumor storm about its founder forced a police-report statement. 优必选 / UBTECH switched on the world's first 10,000-unit-a-year humanoid factory in Liuzhou, while 宇树 / Unitree slid below ¥200B market cap just 20 days after listing. Pony.ai put a driverless Chinese-made ride-hailing service on European public roads.

Commentary: China's AI trade is splitting into two clocks. Capital markets are downshifting fast; factories and token meters are accelerating faster than either promised. This week both clocks struck at once.

---

I. AI Foundation Models

1. Moonshot Ships Kimi K2.8, Hands Its Million-Token Context Away for Free

• Summary: 月之暗面 / Moonshot AI pushed Kimi K2.8 Preview to all users on Kimi Code and Kimi Work on September 11, claiming overall performance near flagship K3 with stronger coding and agent skills. The real move is pricing: 1M-token context is now open to every membership tier, while K3 still gates it behind upper plans. Bloomberg reported the company wants $2B ARR by year-end; its ARR crossed $1B in August, up from $300M in June.
• Source: QbitAI · 2026-09-12
• Editor's Take: Trickle-down capability is a growth hack, not generosity. Moonshot is deliberately cannibalizing its own flagship funnel two months before an IPO filing, because subscription conversion at the bottom of the ladder is worth more to a listing valuation than margin at the top.

2. A Rumor With No Source Nearly Took Down China's Most Valuable AI Startup

• Summary: A screenshot with no attribution spread through Chinese VC WeChat groups late September 11 alleging trouble around founder 杨植麟 / Yang Zhilin and staff. Moonshot's official Weibo answered September 12: the claims are fabricated and malicious, and the company has filed a police report. The rumor hit a company already sitting on a reported $50B pre-money valuation, up from $4.3B a year ago.
• Source: ITHome · 2026-09-12
• Editor's Take: A valuation up 10x in nine months means the mark-to-market lives entirely in group chats. When the rumor mill can move a pre-IPO AI company, the pricing is sentiment, not diligence.

3. DeepSeek quietly turns on voice chat

• Summary: Users spotted a speaker icon in the DeepSeek app over September 11-12 as the lab begins a gray-test of AI voice conversation with four selectable voices, first reported by app watchers and confirmed by ITHome.
• Source: ITHome · 2026-09-12
• Editor's Take: DeepSeek ships features the way it used to refuse to: late, cheap, and without a keynote. Voice is the last interface the big three Chinese consumer assistants all now have.

4. China Expects to Burn 100 Quadrillion Tokens This Year

• Summary: A China Telecom Research Institute report on AI infrastructure, cited by CCTV September 12, says token demand is exploding as the industry's center of gravity moves from model races to agent deployment, projecting 10 quintillion (100 quadrillion, 10^17) tokens consumed in China in 2026.
• Source: ITHome · 2026-09-12
• Editor's Take: Nobody in Washington or San Jose publishes a national token-consumption forecast. China is formalizing token usage as an economic statistic, which is exactly what you do when you intend to plan infrastructure around it.

5. Academician Wu Hequan: China on Track for 30% of Global Compute by 2030

• Summary: At the 2026 China Computing Power Conference on September 12, CAE academician 邬贺铨 / Wu Hequan pegged March token consumption at a daily average of 140 trillion driven by agents, and projected China holding 30% of global compute by 2030. His caveat: token burn should be judged on efficiency, not volume.
• Source: ITHome · 2026-09-12
• Editor's Take: A state-affiliated figure naming a global market-share target for compute is a planning document in disguise. The interesting half is the efficiency warning, aimed squarely at the subsidy-driven GPU buildout.

6. Agent Post-Training Hunts for the Next Scaling Law

• Summary: Panels at the Inclusion·Bund Summit in Shanghai on September 12 argued the capability metric is shifting from answering correctly to acting persistently in real environments: running tools, checking results, revising strategy. Labs are pouring effort into post-training agents that accumulate experience on the job.
• Source: Leiphone · 2026-09-12
• Editor's Take: Pre-training scaled on data, RLHF scaled on preference, and both curves are flattening. Everybody is now guessing which axis of the agent loop is cheap enough to scale, and nobody wants to admit it's a guess.

7. MiniMax Charts Its Own Course to AGI Against the Frontier Pace

• Summary: Haike Finance's long read on MiniMax notes the US cadence, from Anthropic's Claude Fable 5.1 on September 1 to Meta's Muse Spark 1.3 on September 2, and asks how a four-front Chinese lab with an open-weight line keeps its AGI thesis credible while public-market investors want quarters.
• Source: TMTPost · 2026-09-12
• Editor's Take: MiniMax is the Chinese lab most willing to say it is not chasing the frontier and also expects to survive. Either that's discipline or it's a hedge; on current cash flows, investors are paying for hope either way.

8. Alibaba's Qoder CLI Opens Custom Model Access to Individual Users

• Summary: From v1.1.50, Alibaba's AI coding tool Qoder lets personal-tier users plug in their own models, a feature previously reserved for enterprise. Qoder launched in October 2025 as the command-line arm of Alibaba's agent coding platform.
• Source: ITHome · 2026-09-12
• Editor's Take: Bring-your-own-model kills lock-in revenue but multiplies the install base. When a coding agent is commodity, the durable asset is the developer's muscle memory, and Alibaba is buying it at retail.

---

II. AI Software & Applications

9. Tokens as Collateral: China's 'Compute Loan' Program Goes National

• Summary: Per CCTV Finance at the China International Fair for Trade in Services, which opened a joint finance-tech zone this year, new instruments including compute loans, compute insurance and invest-loan linkage are scaling nationwide to fund data centers and GPU cluster purchases. Booths paired banks with robot arms, rocket models and humanoids.
• Source: ITHome · 2026-09-12
• Editor's Take: Lending against token output is securitizing inference margins, and it will spread exactly as far as model pricing holds. If API prices keep falling at this cadence, someone is going to underwrite a fleet of GPU-backed bonds that default on a discount curve.

10. Ant Group Renames GPASS to 'LingYing', Betting AI Glasses Are the Next Agent Slot

• Summary: At the Bund Summit's new-terminals forum on September 11, 蚂蚁集团 / Ant Group upgraded its trusted connection framework GPASS into LingYing (灵影), positioned as an agent-native OS and open platform for AI glasses covering natural interaction, agent runtime, multi-device coordination and security, opening to chip partners.
• Source: Leiphone · 2026-09-12
• Editor's Take: Ant keeps building payment-shaped infrastructure for hardware it doesn't make. If a Chinese payments giant thinks smart glasses need a trust layer badly enough to name one, that's the strongest demand signal the category has had since 2019.

11. Tencent Video on HarmonyOS Ships On-Device A2A: One Command Wakes Cross-App Agents

• Summary: HarmonyOS's Tencent Video became the first app to expose on-device A2A (agent-to-agent) capability, letting Huawei's Xiao Yi assistant call app services through openJiuwen A2A, per the HarmonyOS developer blog September 11.
• Source: ITHome · 2026-09-12
• Editor's Take: Agent-to-agent protocols are being standardized top-down by OS vendors, not bottom-up by startups. The A2A acronym war just arrived on 800 million devices, and Google's version doesn't ship there.

12. Alibaba's Qianwen Office Lands in Kylin's Store: All Major OSes Covered

• Summary: Alibaba's enterprise agent product 千问办公 / Qianwen Office joined the Kylin software store September 12, completing native coverage of Windows, macOS, HarmonyOS, 银河麒麟 / Galaxy Kylin and UOS.
• Source: ITHome · 2026-09-12
• Editor's Take: Kylin means government and state-enterprise desktops. An AI office agent cleared for the Xin-chuang stack is a procurement checkbox, and checkboxes are how foundation models get paid, not product hype.

13. 8.8 Yuan for 5,000 Credits: CCTV Exposes the AI-API Reseller Gray Market

• Summary: CCTV reported resale platforms flooded with bargain AI subscription credits actually routed through third-party API relays that rewrite local config, letting operators intercept users' code and office documents for resale. The exposure lands while relay services quietly power a chunk of cheap Chinese API traffic.
• Source: ITHome · 2026-09-12
• Editor's Take: This is the mirror image of the US "vibe-coding leak" panic, resolved by economics instead of litigation. Discount tokens are a supply-chain story, and CCTV just told every CISO in the country whose prompt logs are a revenue line somewhere.

14. First-Half AI Funding Hits 1.7x All of 2025; VC Says 90% of Capital Goes to AI in Five Years

• Summary: At the Bund Summit's venture meetup September 9-12, AEF NextGen Fund's 贺飞 / He Fei said global AI fundraising in H1 2026 already exceeded 1.7x full-year 2025, fielding the bubble question with a forecast that over 90% of VC capital will target AI within five years.
• Source: Huxiu · 2026-09-12
• Editor's Take: The bubble accusation and the bubble plan got stated in the same breath. A 90% allocation isn't a conviction, it's a career-risk hedge, and capital that pools that tightly prices each other's exits.

15. GPU Challenger Enflame Debuts on STAR Market: Social Security Fund Up ¥150M on Day One

• Summary: Domestic cloud AI chipmaker 燧原科技 / Enflame Technology listed on the STAR Market September 11 at ¥142.18 ($20) a share, the year's third-priciest debut. The National Social Security Fund board's day-one paper gain reached ¥150M, and insurance capital booked ¥1.9B in a single day. Its fate, argues 妙投, is tied to its anchor backer and customer: 腾讯 / Tencent.
• Source: Huxiu · 2026-09-12
• Editor's Take: Long-term money queuing for a loss-making GPU IPO is the state telling the market Chinese compute has a bid. The Tencent dependence cuts both ways, though: Enflame's order book is one strategic customer's capex decision with a ticker on top.

---

III. Humanoid Robots

16. UBTECH Switches On the First Humanoid Factory Built for 10,000 Units a Year

• Summary: 优必选 / UBTECH opened its Liuzhou super-smart factory September 12, billed as the world's first industrial humanoid plant designed for 10,000-unit capacity, using industrial simulation for line planning and a digital "smart brain" for production control.
• Source: ITHome · 2026-09-12
• Editor's Take: Capacity claims used to be the fake part of the humanoid business. A factory rated in tens of thousands means someone has signed demand sheets to amortize it; the only interesting question is who.

17. Unitree Falls Below ¥200B, and China's Robot Makers Head for the Factory Floor

• Summary: 宇树 / Unitree's market cap dropped under ¥200B (~$28B) on September 11, halving in about 20 trading days from its intraday peak of ¥444.9B on the STAR debut. The harshest effect: peers that traded on "Unitree is up, so I'm worth more" are now signing factory deployment contracts instead of demo deals.
• Source: Huxiu · 2026-09-12
• Editor's Take: A 50% drawdown in three weeks in the most hyped category of the year is the market repricing orders into revenue. The companies now working factory shifts aren't the survivors; they're the ones who found customers before the halo faded.

18. "The Hot Money for Embodied AI Isn't Hot Enough": Bund VCs Cool the Robot Trade

• Summary: Investors from CAS Star, Fengrui, BAI Capital and Granite Asia told a Bund Summit session September 11 that embodied intelligence is AI's most active venture direction yet remains underfunded relative to its hype, with judgment criteria tightening toward repeatable commercial scenarios.
• Source: Huxiu · 2026-09-12
• Editor's Take: VCs saying a category has too little money while the public market halves its leader's valuation in a month is the exact mid-cycle signature. The capital that calls the robot trade underfunded is waiting for the shakeout to buy the wreckage.

---

IV. Autonomous Driving

19. Pony.ai Puts Driverless Cars on Europe's Public Roads, No Safety Driver

• Summary: 小马智行 / Pony.ai, backed by 极狐 / Arcfox hardware, joined Rimac-backed Verne's fully driverless passenger tests in Zagreb opened September 12, the first such public-road commercial test in Europe. Riders hail via the Uber app on a 22km route linking Verne's HQ, downtown and the airport; Pony's stack drives while Verne supervises remotely.
• Source: ITHome · 2026-09-12
• Editor's Take: Waymo is still lobbying city halls; Pony.ai is already turning fares in the EU, on somebody else's brand and Uber's demand. Chinese autonomy's European beachhead isn't won with tech flags but with partners who already own the licenses.

20. Wuling x Huawei SUV Ships ADS Pro at ¥149,800, Full OTA to ADS 5 on Day After Launch

• Summary: Huajing Auto (SAIC-GM-Wuling x 华为 / Huawei) added four trims to the Huajing S September 12 at a trade-in window of ¥149.8k-196.8k (~$21k-27.5k), all with Huawei Qiankun ADS Pro enhanced standard. The company confirmed a fleet-wide OTA to ADS 5 on September 13.
• Source: ITHome · 2026-09-12
• Editor's Take: Flagship driver assistance as standard equipment at 21 grand is Huawei's version of the Apple tax, minus the tax. The Western debate over whether FSD should be subscription or standard just became academic in the world's largest car market.

---

V. World Models / Physical AI

21. Shengshu's World Model Lets Robots Train Themselves

• Summary: 生数科技 / Shengshu Tech demoed a world-model route toward Recursive Self-Improvement for robots: touch sensing, memory and egocentric (Ego) data feed a loop where a failed screwdriver grab generates the correction signal without a human labeling what went wrong.
• Source: QbitAI · 2026-09-12
• Editor's Take: The humanoid bottleneck was always the cost of being told you're wrong. If self-correction scales, data factories lose their business case, and that valuation layer is about two years ahead of repricing.

22. AI Agents Are Hunting for an 'Equivalent of Competence'

• Summary: Havenlon Labs' essay reads the recent agent-security incidents as one failure mode: teams treat sandbox escapes and permission grants as isolated bugs while agents operate without any calibrated sense of what they're allowed to do, an "capability equivalent" that evaluation never measures.
• Source: Huxiu · 2026-09-12
• Editor's Take: Written a week before OpenAI's agent DDOSed a package registry, which reads like prophecy and is really just the roadmap. Every deployer who ran the eval set and shipped anyway is the audience for this essay, and they know it.

---

物界前沿 | EastVoice