East Voice · China AI Daily | August 8, 2026 · Issue #767

Editor’s Note: Today is August 8, 2026. Unitree Robotics has priced its IPO at ¥150.8/share with subscription opening August 10, becoming the first humanoid robot maker listed on China’s A-share market. DeepSeek makes a strategic investment in Unitree alongside a joint R&D agreement for embodied AI. Cambricon reports H1 revenue of ¥6B with profit doubling. China issues its first mandatory national standard for L3/L4 autonomous driving. AI video model startups raised nearly ¥30B in three months.

---

I. AI Large Language Models (LLM / Foundation Model)

1. DeepSeek Announces Major API Price Hike, Expected 35%-50% Increase

• Summary: DeepSeek announced on August 6 that it plans to raise API pricing across the board, with some base model calls expected to increase 35%-50%, and enterprise-grade services seeing even larger increases.
• Source: Wall Street CN / DeepSeek Official · 2026-08-06
• Editor’s Take: DeepSeek previously disrupted the market with ultra-low pricing. This increase signals that compute costs and commercialization pressure are finally reaching the pricing layer — potentially marking a turning point in China’s LLM price war.

2. DeepSeek V4 Flash Hit by “Unprecedented” Traffic at Launch, API Briefly Overwhelmed

• Summary: On August 4, DeepSeek V4 Flash API experienced performance degradation due to unprecedented traffic volume. Overseas platform OpenCode reported near-total unavailability before emergency fixes restored service.
• Source: Financial Jie / OpenCode · 2026-08-04
• Editor’s Take: From “too cheap to be trusted” to “too popular to handle” — the global open-source AI community’s demand for Chinese models is real and growing.

3. ByteDance CEO Liang Rubo: We’ll Stick with Self-Developed LLMs, Accept Short-Term Lag

• Summary: At ByteDance’s mid-year all-hands meeting on August 6, CEO Liang Rubo stated the company will persist with in-house LLM development, “accepting short-term lag while optimizing for the long term — the most important thing is not to distort our approach.”
• Source: Cailianshe · 2026-08-08
• Editor’s Take: “Accepting short-term lag” — this is the first time ByteDance has publicly acknowledged its LLM isn’t leading. But the phrase “not to distort” reveals long-term strategic discipline.

4. Moonshot AI Completes Share Reform, IPO Signal Strengthened

• Summary: On July 29, Moonshot AI (maker of Kimi) completed a corporate restructuring from limited liability company to joint-stock company — a standard pre-IPO move in China.
• Source: Keji Ayi · 2026-08-03
• Editor’s Take: Share reform is a prerequisite for IPO. With Kimi K3 topping benchmarks, Moonshot’s listing timeline appears to be accelerating.

5. Five Chinese LLMs Released in Eight Weeks: From Solo Breakthrough to Cluster Surge

• Summary: In the past 8 weeks, Moonshot AI’s Kimi K3, DeepSeek V4 Flash, Zhipu GLM-5.2, ByteDance Seedance 2.5, and Alibaba Qwen3.8-Max were released consecutively, marking China’s AI shift from “one company’s breakthrough” to “cluster-level output.”
• Source: NetEase / CSDN · 2026-08-04
• Editor’s Take: Five models in eight weeks — this density exceeds any comparable period in Silicon Valley. China’s LLM competition has upgraded from individual breakthroughs to systematic output.

6. EU AI Act Article 50 Transparency Rules Take Effect August 2, Chinese Models Face New Compliance Era

• Summary: The EU AI Act’s Article 50 transparency obligations are now enforceable. AI-generated content must carry machine-readable provenance marks. Chinese models entering the European market face new compliance requirements.
• Source: Zhike Xiaotan · 2026-08-08
• Editor’s Take: The EU AI Act doesn’t only regulate European companies. Chinese models going global must prepare compliance strategies in advance or risk market access restrictions.

---

II. AI Software (AI Application / SaaS)

1. AI Video Model Funding Surges: Four Companies Raise ~¥30B in Three Months

• Summary: Kling AI closed nearly $3B (valued at $18B), Shengshu Technology raised $500M, AIShi Tech completed ¥2.98B Series C, and Yanyu Tech raised ~$300M Series B+. Combined, these four companies raised nearly ¥30 billion in three months.
• Source: Economic Information Daily · 2026-08-07
• Editor’s Take: Tencent, Alibaba, and Baidu all investing in Kling simultaneously — internet giants are both investors and competitors. The AI video “finals” have begun.

2. Kling AI Surpasses 100M Global Users, Q1 Revenue Over ¥650M, ARR Near $500M

• Summary: Kling AI’s global user base exceeded 100M (up 67% from end-2025). Q1 2026 revenue topped ¥650M (300%+ YoY growth), with March ARR approaching $500M.
• Source: Economic Information Daily · 2026-08-07
• Editor’s Take: 100M users + ~$500M ARR — Kling AI is no longer just “China’s Sora,” it’s the most commercially proven player in the global AI video space.

3. ByteDance Releases Seedance 2.5, Single Video Generation Extended to 30 Seconds

• Summary: ByteDance released next-gen video generation model Seedance 2.5, achieving breakthroughs in four core capabilities over Seedance 2.0, with single video generation extended to 30 seconds.
• Source: Economic Information Daily · 2026-08-07
• Editor’s Take: 30-second single generation — this number means AI short videos are moving from “clip stitching” to “one-take” production.

4. Kunlun Tech AI Short Drama Platform ARR Surpasses $700M, #1 in Overseas Short Drama Market

• Summary: Kunlun Tech’s AI short drama platform hit monthly revenue over $48M, ARR exceeding $570M. By end of June, ARR broke $700M, with MAU ranking #1 in the overseas short drama market.
• Source: Economic Information Daily · 2026-08-07
• Editor’s Take: AI short drama production costs dropped from $40-70K to $4-7K per episode, with $700M ARR — this may be the most mature commercial model for AI-generated content today.

5. FreeDo Tech: Spatial Intelligence Is the “18th Birthday” of World Models

• Summary: FreeDo Tech co-founder Song Bin stated in an interview that most current world models are still at the “spatial intelligence” stage — without mature spatial intelligence, world models cannot exist. FreeDo has been working on 3D spatial data for a decade and is planning a new funding round.
• Source: NetEase Tech “Attitude” · 2026-08-07
• Editor’s Take: “Spatial intelligence is the coming-of-age ceremony for world models” — this metaphor precisely identifies the industry’s real bottleneck: not generation capability, but understanding of physical space.

---

III. AI Chips & Computing Power

1. Cambricon Reports H1 Revenue of ¥6B, Net Profit Up 122% YoY

• Summary: Cambricon released its H1 2026 financials: revenue reached ¥5.996B, net profit surged 122% YoY, delivering a strong report as China’s leading AI compute chip maker.
• Source: Quick Tech · 2026-08-07
• Editor’s Take: Revenue doubled, profit doubled — Cambricon proves domestic AI chips aren’t just “usable,” they’re profitable.

2. China’s Domestic AI Chip Market Reaches ¥300B in 2026, From “Usable” to “Good”

• Summary: CITIC Securities forecasts China’s domestic AI chip market will reach ¥300B in 2026, with three parallel technology routes: inference-specialized, full-stack self-developed, and general-purpose GPU.
• Source: Securities Star / Digital碎玉斋 · 2026-08-07
• Editor’s Take: ¥300B market, three parallel tech routes — China’s AI chip industry is shifting from “catching up with NVIDIA” to “building its own ecosystem.”

3. Samsung, Micron, SK Hynix 2027 HBM/DRAM Capacity Already Sold Out to AI Firms

• Summary: The big three memory makers’ 2027 DRAM and HBM high-end capacity has been pre-booked by AI server companies, tightening supply and pushing up hardware costs including for humanoid robots.
• Source: Toutiao · 2026-08-08
• Editor’s Take: High-end memory capacity “pre-sold” through 2027 — the AI compute hunger is visible, and it will push up costs for all HBM-dependent AI hardware including robots.

4. Inference Compute Now Accounts for 2/3+ of Total AI Computing Demand

• Summary: In 2026, inference compute demand has surpassed training, accounting for over two-thirds of total AI computing, making inference-specialized chips the fastest-growing segment.
• Source: Securities Star · 2026-08-07
• Editor’s Take: Training = GPU, inference = specialized chips — the compute landscape is shifting from “training-first” to “inference-first,” directly affecting LLM deployment and pricing logic.

---

IV. Humanoid Robots

1. Unitree Robotics IPO Priced at ¥150.8/Share, Subscription Opens August 10, Valuation ~¥61B

• Summary: Unitree Robotics has priced its STAR Market IPO at ¥150.8/share, issuing 40.4M new shares (10% of total equity), raising ~¥6.1B. Online/offline subscription opens August 10, payment due August 12.
• Source: Sohu / Yicai Global / Global Times · 2026-08-06
• Editor’s Take: China’s “first humanoid robot stock” finally lands. The ¥61B valuation provides the first transparent pricing anchor for the entire embodied intelligence sector.

2. DeepSeek Makes Strategic Investment in Unitree IPO, ¥141M Stake with Joint R&D Agreement

• Summary: DeepSeek’s parent company subscribed to 933K shares worth ¥141M in Unitree’s strategic placement (36-month lock-up), signing a joint development agreement for large model AI and embodied intelligence systems.
• Source: CGTN / Global Times / Metaverse Post · 2026-08-07
• Editor’s Take: Two of the “Hangzhou Six Tigers” joining forces — DeepSeek’s model capability + Unitree’s hardware capability = China’s strongest embodied AI combination. The strategic significance far outweighs the financial.

3. Unitree Opens First Flagship Store in Beijing’s Wangfujing

• Summary: On August 5, Unitree opened its first nationwide retail store at Beijing’s Wangfujing Yintai in88 mall, where consumers can experience humanoid and quadruped robots up close.
• Source: Visual China / Toutiao · 2026-08-05
• Editor’s Take: A robot company opening a retail store — a first in China’s consumer electronics history, suggesting Unitree is pursuing a smartphone-brand-style To C strategy.

4. Unitree H1 Revenue Estimated at ¥1.05-1.13B, Net Profit ¥258-306M

• Summary: Unitree forecasts H1 2026 revenue of ¥1.05-1.13B with net profit of ¥258-306M. Full-year 2025 revenue was ~¥1.7B with humanoid robot shipments ranking #1 globally.
• Source: Yicai Global / Global Times · 2026-08-07
• Editor’s Take: Over ¥1B in H1 revenue with real profits — Unitree proves humanoid robots aren’t just a money-burning exercise, they’re a real business.

5. Zibianliang Robot Files for Hong Kong IPO, Embodied Intelligence Competes for “Model Pricing Power”

• Summary: Zibianliang Robot filed a prospectus with the Hong Kong Stock Exchange, valued at ~¥20B, becoming the third embodied intelligence company after Unitree and Zhiyuan to pursue a public listing.
• Source: HK Stock Research · 2026-08-08
• Editor’s Take: Second IPO filing after Unitree — the embodied intelligence “listing wave” is accelerating, but valuations will depend on model capability, not just hardware shipments.

6. Jinglianwen Tech Launches EGO+UMI Dual-Paradigm Data Collection for VLA Models

• Summary: Jinglianwen Tech launched its EGO+UMI dual-paradigm data collection solution in August 2026, combining first-person global vision with fine-grained end-effector manipulation to provide “high-fidelity training data” for VLA models.
• Source: Jinglianwen Tech · 2026-08-08
• Editor’s Take: The bottleneck in embodied intelligence isn’t algorithms — it’s data. High-quality VLA training data is the key to robots going from “functional” to “reliable.”

7. Embodied Intelligence IPO Wave Accelerates: Unitree, Zhiyuan, Zibianliang All Pursue Listings

• Summary: Unitree on STAR Market, Zhiyuan heading to HKEX, and Zibianliang filing in Hong Kong — three embodied intelligence companies are simultaneously pursuing public listings in H2 2026, marking a new commercialization phase.
• Source: Global Times / HK Stock Research · 2026-08-08
• Editor’s Take: Three simultaneous IPO pursuits will bring unprecedented transparency to embodied intelligence valuations and pressure companies to shift from “storytelling” to “delivering results.”

---

V. Autonomous Driving

1. China Issues First Mandatory National Standard for L3/L4 Autonomous Driving

• Summary: China’s Ministry of Industry and Information Technology officially released the first mandatory national standard for L3 (conditional automation) and L4 (high automation) driving systems, establishing a unified safety baseline.
• Source: Toutiao · 2026-08-07
• Editor’s Take: Mandatory national standard = market entry threshold. L3/L4 is no longer “do whatever you want” — safety baseline compliance is now mandatory before hitting the road.

2. Baidu Apollo Go Enters London, Exporting China’s Autonomous Driving Blueprint

• Summary: Baidu’s Apollo Go robotaxi service officially entered the London market. China also led the development of the world’s first unified technical regulation for L3+ autonomous driving.
• Source: Qilu Yidian · 2026-08-07
• Editor’s Take: From technology export to standard export — Apollo Go entering London marks China’s autonomous driving evolution from “product output” to “rule output.”

3. Hong Kong Issues First Autonomous Driving License, Apollo Go Testing in Kwun Tong and Kowloon City

• Summary: Hong Kong’s Transport Department approved Apollo Go for autonomous driving tests, with 62 vehicles authorized for multi-district testing. The city will gradually move toward fully driverless testing this year.
• Source: Qianjiang Evening News · 2026-08-02
• Editor’s Take: Hong Kong’s right-hand-drive, left-side traffic + complex urban roads make it a critical testing ground for Apollo Go’s expansion into right-hand-drive markets.

4. Shenzhen Apollo Go Experience: Bluetooth Unlock, Private Space with Karaoke

• Summary: A hands-on test of Shenzhen’s Apollo Go robotaxi revealed Bluetooth-enabled seamless boarding, private cabin with karaoke, massage, and white noise features — tech-forward comfort at its finest.
• Source: Eastmoney / Smart Driving Scout · 2026-08-05
• Editor’s Take: Robotaxis aren’t just about “no driver” — they’re redefining the “in-car experience.” Karaoke, massage, white noise — user experience is becoming the differentiator in the robotaxi competition.

---

VI. World Models / Physical AI

1. People’s Daily: Academician Zheng Nanning Explains Embodied Intelligence as “Future Industry” in 15th Five-Year Plan

• Summary: Chinese Academy of Engineering Academician Zheng Nanning published a feature article in People’s Daily, systematically explaining embodied intelligence’s concept, capability evolution, and applications, noting the 15th Five-Year Plan has listed it as a future industry.
• Source: People’s Daily · 2026-08-08
• Editor’s Take: Academician-level explainer + national policy endorsement — embodied intelligence has officially moved from “academic concept” to “national industrial strategy.”

2. WAIC 2026: National AI Application Pilot Base and Huawei Release MoWorld-3D World Model

• Summary: Ahead of WAIC on July 16, the National AI Application Pilot Base (Embodied Intelligence) jointly with Huawei and Moxin Tech released MoWorld-3D, a fully domestically-built (Ascend NPU) interactive 3D world model, available as plug-and-play for the entire industry.
• Source: Huawei / Moxin Tech · 2026-07-16
• Editor’s Take: Domestic NPU + domestic world model = full-stack sovereign physical AI infrastructure. MoWorld-3D’s “plug-and-play” positioning lowers the industry entry barrier.

3. China’s 2026 LLM Status: Two-Year Surge from Catching Up to Running Alongside

• Summary: Multiple analyses indicate China’s LLM capability has significantly narrowed the gap with the US in 2026, entering the “single-digit gap era,” with local advantages in open-source models and application deployment.
• Source: Toutiao · 2026-08-08
• Editor’s Take: From “catching up” to “running alongside” in just two years — this pace is itself the most powerful narrative for China’s AI industry.

4. H2 2026 AI Battle Begins: DeepSeek’s Lightweight Surge, Moonshot’s Quiet Share Reform

• Summary: DeepSeek’s lightweight V4 Flash model disrupted the market with extreme cost-performance, while Moonshot AI completed share reform for IPO preparation. H2 2026 China AI competition enters a new phase.
• Source: Keji Ayi · 2026-08-03
• Editor’s Take: Lightweight surge + share reform — the keywords for H2 aren’t “bigger,” they’re “lighter” and “more commercial.”

---

---

东声 East Voice — Physix Frontier’s China AI column for global readers

东声西达，中国视角。

© 2026 Physix Frontier. All rights reserved.