社区讨论 · 政策

EastVoice · AI · 2026-08-11 · Issue #775

EastVoice8月11日2026/08/11 325 浏览

Editor’s Note: Unitree Robotics priced its STAR Market IPO at ¥150.80/share with a 0.018% lottery-rate, setting a new valuation anchor for China’s robotics sector. DeepSeek-V4-Flash topped OpenRouter’s global leaderboard with 8.83 trillion weekly tokens. Chengdu enacted China’s first municipal AI regulation, effective September 1.

Commentary: When Chinese models sweep the top four slots on global API leaderboards and a humanoid-robot IPO draws lottery-level demand, the narrative is shifting from “China catches up” to “China sets the pace.”

Large Models & Algorithms

DeepSeek-V4-Flash Tops Global Token Usage at 8.83 Trillion Weekly Calls

Summary: OpenRouter’s weekly leaderboard (August 3–9, 2026) shows global AI model API calls reached 69 trillion tokens, up 21.48% week-over-week. DeepSeek-V4-Flash production release claimed the #1 spot with 8.83 trillion tokens (up 570% WoW). Chinese models occupied all top four positions. Agent and coding applications accounted for 19% of apps but 81.2% of total token consumption.
Source: OpenRouter / Huashang Daily · 2026-08-11
Editor’s Take: A 570% week-over-week surge isn’t incremental growth — it’s a regime change. DeepSeek-V4-Flash’s cost of ~$0.03 per comparable task versus Claude Fable 5’s $3.15 is drawing a “DeepSeek kill line” that mid-tier models must either undercut or outperform.

Alibaba Releases Qwen3.8-Max with 2.4-Trillion-Parameter MoE Architecture

Summary: Alibaba unveiled Qwen3.8-Max, a 2.4-trillion-parameter Mixture-of-Experts model with major improvements in coding and professional productivity. It ranks in the global top tier on Arena benchmarks. The 27B and Max variants will be open-sourced. The model is deeply integrated with DingTalk, Taobao, and other Alibaba ecosystem products.
Source: DAMO Developer Matrix · 2026-08-11
Editor’s Take: 2.4 trillion parameters via MoE — Alibaba found the sweet spot between scale and efficiency. Deep integration with DingTalk and Taobao gives Qwen a moat no standalone model can replicate.

Alibaba Qwen Open Platform Launches Across Mobile, PC, and Smart Glasses

Summary: Alibaba’s Qwen open platform went live, covering logistics, real estate, local services, finance, and automotive in its first batch. Users can invoke multiple service providers through conversational AI without switching between apps.
Source: AIBase · 2026-08-10
Editor’s Take: Qwen is evolving from a chatbot into an operating system. When AI can directly call logistics, finance, and real-estate APIs, it stops being a tool and becomes the interface.

Zhipu GLM-5.2 Released, Surpassing GPT-5.5 on Select Benchmarks

Summary: Zhipu AI released GLM-5.2, which outperforms GPT-5.5 on several benchmarks. GLM-5.5 is expected in August 2026 under an MIT license. The release fills a high-end slot in China’s domestic model lineup.
Source: DAMO Developer Matrix · 2026-08-11
Editor’s Take: Five flagship models in eight weeks — Zhipu’s iteration pace is unmatched domestically. GLM-5.2 beating GPT-5.5 on specific benchmarks proves Chinese models can compete at the frontier in targeted tasks.

ByteDance Seedance 2.5: 30-Second One-Take Video Generation

Summary: ByteDance’s Seed team released Seedance 2.5, its next-generation video generation model. Single-shot duration increased from 15 to 30 seconds, with support for multi-round extension, 50+ multimodal reference inputs, and per-second local editing. API pricing: ¥42/million tokens (video-input) and ¥70/million tokens (text-to-video), roughly ¥1.51/second at 720P.
Source: Toutiao/YouXi AI / NetEase · 2026-08-10
Editor’s Take: Doubling from 15 to 30 seconds isn’t incremental — it’s the difference between stitching fragments and telling a story. At ¥1.51/second, the cost is real but the creative ceiling just jumped.

WeChat Moments Tests “AI-Assisted Writing” Feature

Summary: WeChat is grey-testing an “AI-Assisted Writing” feature in Moments. An AI assistant named “Xiaowei” auto-generates post copy from text, voice, or image input, offering three tonal styles: concise, playful, and literary.
Source: AIBase / Sina Finance · 2026-08-10
Editor’s Take: WeChat’s 1.2 billion users plus AI writing equals the world’s largest AI content-generation gateway. Tencent is finally making its most “grounded” AI move.

Chinese LLMs Accelerate: Five Flagship Releases in Eight Weeks

Summary: From July to August 2026, China’s large-model industry entered a new high-speed iteration cycle. DeepSeek-V4-Flash, Qwen3.8-Max, GLM-5.2, and Seedance 2.5 launched in rapid succession. The US-China frontier-model gap continues to narrow, with competition shifting toward agent capability, inference cost, and real-world deployment.
Source: Toutiao/Qingyan AI · 2026-08-11
Editor’s Take: Five flagships in eight weeks — a pace that’s making Silicon Valley nervous. Chinese model competition has shifted from “who’s bigger” to “who’s cheaper and more useful,” which is the right direction.

Baidu “Yijing” Named in Sullivan’s 2026 China AI Agent Top 10 Lists

Summary: Baidu’s full-scenario digital-human platform “Yijing” was included in both “Most Globally Promising Chinese Agent Top 10” and “Most Commercially Promising Agent Top 10” by Frost & Sullivan. The platform supports 75 languages and covers TikTok, Shopee, and Lazada.
Source: 36Kr · 2026-08-10
Editor’s Take: Digital humans plus cross-border e-commerce — this is Baidu’s most compelling commercialization story at the application layer. 75-language coverage reveals deeper overseas ambitions than the market recognizes.

AI Chips & Compute

China’s First Fully Domestic 100K-GPU AI Supercluster Goes Live

Summary: According to China’s NDRC, the nation’s first fully domestically produced 100,000-GPU AI supercluster has been officially commissioned, marking a new phase in China’s compute infrastructure. Multiple national compute nodes are undergoing expansion.
Source: CCTV News / ByteNet · 2026-08-10
Editor’s Take: “100K GPUs, fully domestic” — the emphasis is on “fully domestic.” From chips to interconnects to software stacks, China’s compute infrastructure completeness is improving rapidly.

Moore Threads H1 Revenue Hits ¥1.74B, Up 147% YoY

Summary: Chinese GPU leader Moore Threads reported H1 2026 revenue of ¥1.74 billion (up 147% YoY), gross profit of ¥989 million (up 104% YoY), and net-loss reduction of 95.73%. Its MTT S5000 AI cluster achieved scaled sales and was the first to pass national security certification. The MUSA developer ecosystem exceeded 800,000 users.
Source: CLS / ByteNet · 2026-08-10
Editor’s Take: Revenue doubled, losses shrank 96% — Moore Threads is proving that domestic GPUs aren’t just policy plays but viable businesses. S5000’s scaled deployment is a milestone.

TSMC July Revenue NT$467.6B, Up 44.7% YoY

Summary: TSMC reported July revenue of approximately NT$467.6 billion (up 5.6% MoM, up 44.7% YoY). January–July cumulative revenue reached NT$2.87 trillion (up 37% YoY). AI chip demand continues to drive growth.
Source: Toutiao/Financial Breakfast · 2026-08-11
Editor’s Take: 44.7% monthly YoY growth — TSMC’s results are the most direct “thermometer” for AI compute demand. In the global chip arms race, TSMC remains the highest-certainty beneficiary.

Goldman Sachs: Global AI Investment to Exceed $1 Trillion by End of 2026

Summary: Goldman Sachs projects global AI-related investment will reach $1.019 trillion by end of 2026, with the US accounting for $581 billion. Cumulative global AI investment since 2022 is expected to reach $1.8 trillion by year-end. US AI capex as a share of GDP is projected to rise from 1.8% in 2026 to 2.8% by 2028.
Source: Coze News / Toutiao/Financial Breakfast · 2026-08-10
Editor’s Take: A trillion dollars a year — AI has graduated from “sector” to “economy.” Goldman’s call that China’s AI sector has no systemic bubble, with power supply chain, hardware infrastructure, and physical AI as top picks, validates our thesis.

Goldman Sachs: China AI Valuations Reset to Healthy Levels, Three Sub-Sectors Most Attractive

Summary: Goldman Sachs research notes China’s AI sector has no systemic bubble and valuations have reset to healthy levels after recent market corrections. Foreign investors remain significantly underweight Chinese AI stocks. The three most attractive sub-sectors: power supply chain, hardware infrastructure, and physical AI.
Source: Toutiao/Financial Breakfast · 2026-08-11
Editor’s Take: “Foreign investors remain significantly underweight” — this is bullish for domestic AI investors, implying substantial room for foreign capital re-allocation. Physical AI as a top pick confirms our sector thesis.

Embodied AI & Autonomous Driving

Unitree Robotics STAR Market IPO Priced at ¥150.80/Share

Summary: Unitree Robotics disclosed its STAR Market IPO subscription results. Priced at ¥150.80/share with 40.45 million shares issued. After the clawback mechanism, the online allocation was 9.71 million shares with a winning rate of just 0.018%.
Source: Toutiao/Financial Breakfast · 2026-08-11
Editor’s Take: A 0.018% subscription rate rivals lottery odds — the market’s enthusiasm for a robotics-sector leader is unmistakable. The ¥150.80 price sets a new valuation anchor for A-share robotics.

116,000 New Humanoid-Robot Firms Registered in H1 2026

Summary: China’s State Administration for Market Regulation reports 116,000 new humanoid-robot enterprises registered in H1 2026, up 9.5% YoY. The industry is accelerating into a mass-production phase.
Source: Securities Daily / ByteNet · 2026-08-10
Editor’s Take: 116,000 new registrations in six months — the humanoid-robot sector is experiencing a “wild growth” phase reminiscent of China’s NEV industry a decade ago. A brutal shakeout inevitably follows.

Physical AI Becomes WAIC 2026 Core Theme; Automakers Pivot to Embodied Intelligence

Summary: At WAIC 2026, embodied intelligence was elevated to a co-equal pillar alongside compute infrastructure for the first time. NIO launched its “RuiDong” embodied-intelligence platform; Li Auto redefined its vision as “embodied intelligent robots entering the physical world”; XPeng declared “cars are robots.” Momenta debuted as “the first physical-AI stock,” reaching HK$70B+ market cap on its first trading day.
Source: 36Kr/Caijing Magazine · 2026-08-10
Editor’s Take: When every automaker says “we’re an AI company,” it’s either consensus or a narrative bubble. WAIC 2026’s answer may take three to five years to resolve.

China’s First Mandatory Autonomous-Driving Safety Standard: L3/L4 Vehicles Need Certification from 2027

Summary: China issued its first mandatory national safety standard for autonomous driving. L3 and L4 vehicles must pass mandatory safety certification starting in 2027 before they can operate on public roads.
Source: Toutiao/Multiple Sources · 2026-08-11
Editor’s Take: A mandatory national standard is a license to operate. For tier-one players, it raises the barrier; for the long tail, it’s an exit notice. 2027 will be the watershed for China’s advanced autonomous driving.

AI Applications & Enterprise

AI Talent War Intensifies: Nine-Figure Offers Become the Norm in Silicon Valley

Summary: The AI talent war continues to escalate. Top AI engineers are now valued above entire companies. AI firms compete with gifts, nine-figure compensation packages, and equity incentives for scarce talent.
Source: Business Insider · 2026-08-11
Editor’s Take: Talent is the most irreplaceable resource in the AI arms race. When a single engineer’s compensation exceeds a company’s market cap, the scarcity is beyond what capital markets have priced in.

2026 AI for Science Conference Set for August 21–22 in Beijing

Summary: The 2026 AI for Science Conference will be held August 21–22 at Zhongguancun, Beijing, focusing on AI’s transformation of scientific research paradigms. Three tracks cover fundamental research, tech-industry integration, and international cooperation, with 12 parallel sessions spanning biomedicine, energy, materials science, and more.
Source: The Paper · 2026-08-10
Editor’s Take: AI for Science is moving from slogan to infrastructure. Twelve parallel sessions across disciplines signal that AI’s penetration into scientific research is no longer point-in-time breakthroughs — it’s becoming systemic.

Regulation & Policy

Chengdu Enacts China’s First Municipal AI Regulation

Summary: Chengdu’s “Measures for Promoting the AI Industry” take effect September 1, 2026 — the first municipal-level AI regulation in China. It covers six chapters and 36 articles spanning innovation, industrial ecosystem, resource allocation, and governance, including an AI ethics review committee.
Source: CRI/Sichuan Daily · 2026-08-10
Editor’s Take: China’s first — Chengdu is legislating to set rules for AI. The ethics review committee is particularly noteworthy and may become a template for other cities.

WeChat Video Channels AI E-Commerce Rules Take Effect August 10: Label or Get Throttled

Summary: WeChat Video Channels’ “AI Application Governance for Short-Video E-Commerce” took effect on August 10. Four categories of AI-generated content — product visuals, digital-human spokespersons, AI posters, and other AI visuals — must be proactively labeled or face distribution throttling. Digital-human livestreams must “declare identity + have real humans on duty + ensure content accuracy.”
Source: Toutiao/Xiaohe · 2026-08-10
Editor’s Take: This rule targets the biggest grey area in AI-powered e-commerce: using AI to impersonate humans. “Label or get throttled” may become the standard playbook across all content platforms.

Nearly 200 Silicon Valley Startups Lobby Against Restrictions on Chinese Open-Source Models

Summary: Senior US officials proposed investigations and sanctions against Chinese AI companies, triggering strong pushback from Silicon Valley. Nearly 200 startups organized as the “Small Tech Association” sent an open letter to the White House. Meta, NVIDIA, and Microsoft joined the opposition.
Source: Economic Daily / Toutiao/Financial Breakfast · 2026-08-11
Editor’s Take: Silicon Valley defending Chinese open-source models — not out of goodwill, but out of self-interest. Restricting Chinese open-source would harm US startups that depend on it. It’s a textbook case of “hurt them a thousand, wound yourself eight hundred.”

EastVoice | PhysiX Frontier — China AI Insights, Delivered Daily

政策