PHYSIX FRONTIER EASTVOICE
2026.08.13 · Thursday
Issue 007
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
2.4T
Alibaba Qwen3.8-2.4T-A95B parameter count
$1.4 trillion
Damages sought against Meta by US states
18,000
Unitree cumulative humanoid robot production
¥4 trillion
Shanghai software industry goal by 2030
Top Stories 3 · dual commentary ｜ Section Picks 15 · one-line takes
Alibaba open-sources Qwen3.8-2.4T, DeepSeek ships V4 Pro to APIs, and Tencent earnings indicate GPU supply loosening.
Alibaba's Qwen team released the weights for Qwen3.8-2.4T-A95B — a 2.4T-parameter MoE model activating 95B per token, with native 256K context. It's the first time Alibaba has open-sourced a "Max" flagship-class model, letting anyone self-host what was previously only available via API.
What it means for you｜Anyone can now self-host what was previously only available via API for the first time with Alibaba's Max flagship model.
Editor's take｜The open-weight war just escalated. When a 2.4T frontier-class model becomes downloadable, the moat shifts to tooling, alignment, and deployment expertise — not parameters.
DeepSeek rolled out the official V4 Pro release to its API (model name unchanged, new DeepSeek-V4-Pro-0813 entry in docs). The update beefs up agent capabilities, adds Responses API support and Codex integration, with internal benchmarks approaching Anthropic's Fable 5 on several tests.
Editor's take｜DeepSeek's cadence is now "ship to API first, announce later." Pragmatic — and a reminder that in this market, distribution beats spectacle.
In its Q2 2026 earnings call, Tencent said cloud GPU resources should become notably more abundant by late this year or early next, lifting cloud service capacity and MaaS opportunities. The company is coordinating compute allocation between internal models and external cloud demand.
What it means for you｜Cheaper inference across the board is coming as China's GPU crunch bottleneck passes according to Tencent's earnings tone.
Editor's take｜Tencent's tone suggests China's GPU crunch — the bottleneck that gave us "compute nationalism" — is passing. Cheaper inference across the board is coming.
15 AI stories worth your time, each with a one-line take.
Days after four US states' attorneys general sought $1.4 trillion in damages against Meta, the company open-sourced a 30B model overnight. Zuckerberg explicitly cited DeepSeek, Qwen and Kimi as the open-weight rivals to beat, framing open source as the only way to stay relevant.
One-line take｜Meta's legal nightmare and its open-source pivot are two sides of one strategy: own the open layer, because that's where developer gravity has moved. China's open models forced the move.
Alibaba Cloud debuted the Lingjun Zhenwu M890 super-node instance in Ulanqab, claiming it's China's first cloud product that can run 2T+ parameter MoE inference. Each unit is a 64-GPU high-speed interconnect slice, theoretically scalable to 100T-parameter models, no self-built cluster required.
What it means for you｜This collapses the infrastructure barrier for Chinese startups trying to serve frontier-scale models without needing a self-built cluster.
One-line take｜"Rent a supercomputer by the hour" is now real. This collapses the infrastructure barrier for Chinese startups trying to serve frontier-scale models.
ByteDance reportedly spun up a new first-level department, "AI Data & Security", on par with Seed and Flow, led by former TikTok platform trust chief Wang Yinglei. The move signals ByteDance is betting big on data governance and safety as differentiators — even though its rumored 10T-parameter flagship hasn't shipped yet.
One-line take｜ByteDance is building the "trust layer" before the model — a striking inversion of the usual order, and a sign that data strategy is now a board-level priority at Chinese AI leaders.
Huawei Cloud's office AI agent product OfficeClaw has been upgraded and repositioned as OfficeAce, a general-purpose agent spanning personal and enterprise office work — content generation, file processing, knowledge search, and PPT building. Huawei is betting on agents as the entry point into enterprise AI.
What it means for you｜The war isn't about models anymore, it's about which assistant lives in your daily workflow as giants race for enterprise desktop agents.
One-line take｜Every Chinese cloud giant is now racing for the same doorway: the enterprise desktop agent. The war isn't about models anymore, it's about which assistant lives in your daily workflow.
Midea became the exclusive home-appliance partner on Alibaba's Qwen app, integrated through the A2A (Agent-to-Agent) protocol. Users can bind Midea's AI agent for product inquiries, home-environment monitoring, energy queries, and maintenance management — all inside Qwen.
What it means for you｜Users can bind Midea's AI agent for product inquiries, home-environment monitoring, energy queries, and maintenance management inside Qwen.
One-line take｜A2A is moving from white paper to consumer reality. When your fridge's agent can talk to your AI assistant, the "everything app" starts to mean everything.
Doubao rolled out a verified-student program: 2.5x the free usage quota, plus the ¥68/month Pro plan at ¥38/month. It's a direct shot at the China campus market as AI assistants compete to become students' default tool.
What it means for you｜Verified students get 2.5x free usage quota plus the ¥68/month Pro plan at ¥38/month in a direct shot at the campus market.
One-line take｜Pricing students cheap now = default habit later. The AI subscription playbook is copying the streaming playbook.
Unitree announced cumulative production of roughly 18,000 humanoid robots — counting only bipedal humanoids, excluding wheeled-base variants. It's a scale number no Western humanoid maker comes close to.
One-line take｜The "China can't build humanoids at scale" thesis is quietly dead. 18,000 units means supply chains, quality systems, and real-world failure data Western labs only dream of.
Liu Liyuan, a former core product lead at DJI and Anker, has launched a pet-companion robot startup, closing an angel round led by Yuanhe Capital. Pet robots are emerging as one of the few consumer hardware categories with real willingness to pay.
One-line take｜Consumer robotics' next act isn't a humanoid in the living room — it's a cheaper, cuter machine that does something people already pay for. Pets are that wedge.
Lin Junyang — Alibaba's youngest-ever P10 — announced Pragmatik Labs (Yu Yong Ke Ji, "p7k") in Shanghai, backed by Tencent. The startup targets next-generation agents spanning both the digital and physical worlds. Lin left Alibaba five months ago.
One-line take｜The "agents that cross into the physical world" pitch is becoming the standard founding story for China's top AI talent. Watch this space — Lin's team built China's first trillion-parameter open models at Alibaba.
Honor announced its robots will compete in the 2nd World Humanoid Robot Games (Aug 22–26, Beijing Ice Ribbon). Honor's self-developed humanoids previously swept the top six in a half-marathon event. The games feature 50 events: 29 competitive, 21 scenario-based.
One-line take｜Phone makers racing humanoids is no longer a curiosity — Honor treats it as brand engineering. Expect more consumer-tech giants to enter the arena.
Unitree's Shanghai IPO opened subscription with a 150.80 CNY/share price. 9.78 million retail accounts subscribed for 53.6 billion shares — an over-subscription exceeding 8,000x, triggering the clawback mechanism; the final lottery rate was 0.018%.
One-line take｜8,000x oversubscription for a robot maker tells you everything about the retail appetite for China's embodied-AI story — and about how crowded that trade already is.
BridgeDP Robotics (Qiao Jie Shu Wu), a general-purpose robot operating system developer, closed a Pre-A+++ round led by China Mobile's industry fund, with Foresight Fund and Shenzhen Capital adding. Funding goes to its RoboCraft AI motion platform and OS commercialization.
One-line take｜The "picks-and-shovels" layer of humanoids — operating systems, motion platforms — is quietly raising big rounds while consumer humanoids grab headlines.
Changan Auto and Huawei signed a strategic framework agreement in Shenzhen, covering AI products, general and automotive vertical foundation models, and scene collaboration. Both chairmen attended — a signal this is a board-level alliance, not a technical trial.
One-line take｜Huawei's model business is now a car-industry strategy weapon. Every major OEM signing with Huawei tightens the grip of its full-stack playbook.
Shanghai's 15th Five-Year plan for software and information services sets a 2030 goal of ¥4 trillion in industry scale, positioning the city as an AI-empowerment "main battlefield" and global competition bridgehead, with heavy emphasis on foundational model applications and data economy.
One-line take｜Provincial-scale industrial policy is calibrating around AI consumption, not just production. The next decade of China's AI story will be measured in deployment numbers, not demos.
Samsung confirmed it has brought Anthropic's Claude into semiconductor R&D: after ~3 months of Claude Code access for software developers, design verification cycles that took a month now reportedly finish in two days — a 15x speedup at the hardest layer of chip development.
What it means for you｜Design verification cycles that took a month now reportedly finish in two days after Samsung adopted Anthropic's Claude for semiconductor R&D.
One-line take｜AI eating into chip-design verification is the quietest, most consequential adoption story of the year. Whoever owns this workflow owns the semiconductor roadmap.
📖 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 →