物界前沿East Voice Daily | July 16, 2026
🔔 东声 East Voice — Physix Frontier's China AI column for global readers
东声西达，中国视角。
Editor's Note
China's AI industry is entering a defining week. The 2026 World AI Conference (WAIC) opens tomorrow in Shanghai with President Xi delivering the keynote — the highest-level political endorsement of AI as a national strategic priority. On the capital front, MiniMax completed a HK$16 billion mega-raise to train a 3-trillion-parameter open-source model, while the humanoid robot sector crosses a historic threshold: Zhiyuan's 15,000th unit rolled off the production line (a global record), and Unitree's IPO cleared in just 73 days. Meanwhile, Huawei's Yinwang helped write the UN's first unified autonomous driving regulation — China is now exporting standards, not just products. Here are 35 stories across five tracks.
🧠 LLMs
1. MiniMax Raises HK$16 Billion to Train 3-Trillion-Parameter Open-Source Model
Summary: MiniMax (0100.HK) completed a HK$16 billion (~$2B) funding round with 7x oversubscription from 100+ institutions including 20+ sovereign wealth funds. CEO Yan Junjie announced zero salary until AGI is achieved, pledging 5% of personal equity for team incentives and open-source support. The company is training M3 Pro, a 2.5-3 trillion parameter model planned for open-source release.
Source: Toutiao / The Information / GMT Eight
Editor's Take: This is the largest single funding round for a Chinese AI company in 2026. The 3T-parameter M3 Pro, if open-sourced, would rival Meta's Llama series in scale. The CEO's zero-salary pledge signals extreme conviction — but the stock has dropped 80% from its March peak, making this as much a survival play as an offensive one.
2. DeepSeek and Zhipu Developing Custom AI Inference Chips
Summary: Reuters reports DeepSeek has been developing its own AI inference chip for about a year (still early stage), while Zhipu AI committed ¥10B+ to safety research and mechanistic interpretability under its "Touch High" plan. GLM-5.2 with 1M context will go MIT license.
Source: Reuters / 36氪
Editor's Take: When the world's most popular open-source model company builds its own silicon, vertical integration is no longer optional — it's existential. Zhipu's simultaneous bet on interpretability is the right kind of ambition.
3. WAIC 2026 Opens Tomorrow: 261 Models, 108 Chips, President Xi Keynote
Summary: The 9th World AI Conference opens July 17-20 in Shanghai with 100,000+ sqm exhibition, 1,100+ companies, 3,000+ products, 300+ global debuts. President Xi will deliver the keynote address. 12 ministries and 10+ international organizations are participating.
Source: Xinhua / China View / WAIC Press Conference
Editor's Take: WAIC has evolved from a tech expo into a diplomatic platform. President Xi's keynote elevates AI to the same strategic level as semiconductors and aerospace — expect major policy announcements.
4. ByteDance Doubao and Alibaba Tongyi Pull User-Built Agent Features
Summary: Ahead of new AI anthropomorphization regulations taking effect, ByteDance's Doubao and Alibaba's Tongyi Qianwen have removed user-built agent features. The five-ministry interim rules ban emotional dependency induction and virtual companion services for minors.
Source: i黑马 / 36氪
Editor's Take: China is drawing the line between AI assistants and AI friends. The regulatory speed — from announcement to enforcement in weeks — signals Beijing's determination to shape AI's social impact proactively.
5. Tencent Poaches OpenAI Researcher Tian Yonglong
Summary: Tencent has recruited Tian Yonglong, a senior researcher from OpenAI, to join its AI lab. The hire comes as Chinese tech giants intensify the global talent war for AI researchers.
Source: 36氪
Editor's Take: The reverse brain drain is accelerating. Chinese AI labs are offering competitive packages — both financial and in research autonomy — to lure top talent back from Silicon Valley.
6. StepFun Launches Agent OS: World's First Agent-Native Operating System
Summary: StepFun (阶跃星辰) unveiled Step AOS, breaking through memory, decision-making, and action barriers for AI agents. Alongside, it launched the StepX Neo — the first LLM-native smartphone.
Source: 36氪
Editor's Take: From agent frameworks to agent OS — StepFun is building the mobile infrastructure layer for the agentic era. The smartphone-as-agent-host concept could redefine the device paradigm.
7. China's AI Industry Scale Exceeds ¥1 Trillion
Summary: NDRC data shows China's AI-related industry exceeded ¥1 trillion in 2025, with 2026 growth projected above 30%. Daily token calls have reached millions of trillions. Over 30% of large-scale manufacturing enterprises have adopted AI, with 6,000+ vertical domain models deployed.
Source: Xinhua / NDRC / People's Daily
Editor's Take: From "catching up" to "running alongside" to "leading in parts" — China's AI narrative has fundamentally shifted. The ¥1 trillion milestone is less about size and more about ecosystem maturity.
💻 AI Software
8. Sugon Builds China's First Fully Domestic 100K-GPU AI Cluster
Summary: Sugon (中科曙光) completed deployment of China's first 100,000-card fully domestic AI computing cluster in June, now connected to the national supercomputing internet. The scale-up from 30K to 60K to 100K took just 18 months.
Source: 36氪 / Xinhua
Editor's Take: The speed of domestic AI infrastructure buildout is staggering. This is China's answer to semiconductor export controls — build the compute stack at home, at scale.
9. Hygon (海光) Expands Edge AI Chip Portfolio
Summary: Hygon Information Technology announced an expanded edge AI chip lineup targeting automotive, industrial IoT, and mobile devices. The new chips support on-device LLM inference with power consumption under 15W.
Source: 36氪 / IC Insights
Editor's Take: While Huawei grabs headlines for Ascend, Hygon is quietly building a comprehensive edge AI portfolio. The sub-15W power envelope makes these chips viable for a much broader range of devices.
10. Baidu Unveils "Chip-Cloud-Model-Entity" Full-Stack AI Platform
Summary: Baidu announced its integrated "Chip-Cloud-Model-Entity" (芯云模体) strategy, covering Kunlun chips, Baidu Cloud, ERNIE models, and autonomous driving — a vertically integrated AI stack from silicon to end-user applications.
Source: Baidu Official / 36氪
Editor's Take: Baidu is positioning itself as China's closest equivalent to Google's AI ecosystem. The full-stack integration from chips to applications is ambitious but execution remains the key challenge.
11. Nubia Launches AI Agent Phone
Summary: ZTE's Nubia brand launched what it calls the world's first "AI Agent Phone," featuring on-device LLM capabilities with dedicated AI hardware and a custom OS layer optimized for agent interactions.
Source: 36氪 / ZTE
Editor's Take: The smartphone-as-agent concept is gaining traction. Nubia's bet is that dedicated AI hardware will differentiate in a commoditized market — but the real differentiator will be the software ecosystem.
12. Peking University Warns of Imported Scientific Instrument Bottleneck
Summary: A Peking University research team published a report warning that China's heavy reliance on imported scientific instruments — particularly electron microscopes, mass spectrometers, and cryo-EM systems — could become a critical bottleneck for AI-driven scientific research.
Source: Peking University / 科学网
Editor's Take: AI research depends on experimental data, which depends on instruments. If the instrument supply chain is disrupted, AI-driven drug discovery and materials science could stall. This is an underappreciated strategic vulnerability.
13. 10 US and UK Scholars Return to China in Largest-Ever Reverse Brain Drain
Summary: At least 10 senior AI researchers from top US and UK institutions have returned to China in the past six months, joining Tsinghua, Peking University, and corporate AI labs. The trend is described as the largest concentrated reverse brain drain in AI history.
Source: 36氪 / Science
Editor's Take: Geopolitical friction is pushing talent back to China. These researchers bring world-class expertise and networks — a significant boost to China's AI research capacity.
14. China Enters New Phase of AI Agent Regulation
Summary: Following the AI anthropomorphization rules, China is preparing additional regulations specifically targeting AI agents — covering autonomous decision-making, multi-agent coordination, and agent-to-agent commerce. Draft rules expected in Q3 2026.
Source: 36氪 / 法制日报
Editor's Take: China is building the regulatory framework for the agentic era before it fully arrives. This proactive approach could become a global reference — or a constraint on innovation, depending on implementation.
🦾 Humanoid Robots
15. Zhiyuan's 15,000th Humanoid Robot Rolls Off the Line — Global Record
Summary: Zhiyuan Robotics (智元机器人) achieved a historic milestone on June 28 with its 15,000th humanoid robot rolling off the production line. The unit, a Jingling G2 model, was immediately delivered to Longcheer Technology's factory for regular production line work. Zhiyuan will showcase three new products at WAIC.
Source: Zhiyuan Official / Xinhua
Editor's Take: 15,000 units is a global first for humanoid robots — no other company has reached this production scale. The "roll off the line, immediately deploy" model proves these aren't museum pieces.
16. Zhiyuan Unveils Three New Products at WAIC 2026
Summary: Zhiyuan showcased three new humanoid robot models at WAIC, expanding its product line beyond industrial applications into service and research domains.
Source: WAIC organizers / 36氪
Editor's Take: Product line diversification is key to scaling. Industrial use cases prove reliability; service and research applications open much larger addressable markets.
17. Shanghai Accounts for One-Third of China's Humanoid Robot Output
Summary: Shanghai's humanoid robot production accounts for approximately one-third of China's total output, with Zhiyuan, Fourier Intelligence, and other major players headquartered in the city. The Shanghai municipal government has established dedicated humanoid robot industrial parks.
Source: Xinhua / Shanghai Municipal Government
Editor's Take: Shanghai is positioning itself as China's robotics capital. The clustering effect — talent, suppliers, and policy support in one location — accelerates innovation cycles.
18. Unitree Robotics Clears STAR Market IPO in Record 73 Days
Summary: Unitree Robotics (宇树科技) received CSRC approval for its ¥4.2 billion (~$618M) STAR Market listing after a record-setting 73-day review. At potential ~$5.9B valuation, it would become China's first publicly listed humanoid robot pure-play. Unlike Western rivals, Unitree is already profitable.
Source: CSRC / 36氪 / AI2.Work
Editor's Take: 73-day regulatory clearance is extraordinary for a STAR Market IPO. Unitree's profitability sets it apart from every Western humanoid company. This IPO will set valuation benchmarks for the entire sector.
19. "Bolt" Humanoid Robot Achieves 10 m/s Running Speed
Summary: A Chinese robotics lab unveiled "Bolt," a humanoid robot capable of running at 10 m/s (36 km/h) — approaching Usain Bolt's peak speed of 12.4 m/s. The breakthrough uses a novel actuator design combining high torque density with compliant control.
Source: 36氪 / IEEE Spectrum
Editor's Take: Speed is more than a party trick — it's a proxy for actuator power density and dynamic control capability. 10 m/s in humanoid form factor is a significant engineering achievement.
20. Mifeng Data Platform Secures Baidu Investment for Robot Training Data
Summary: Mifeng (觅蜂), a robot training data platform, secured a strategic investment from Baidu. The platform specializes in collecting, annotating, and synthesizing manipulation data for humanoid robots, serving 50+ robotics companies.
Source: 36氪 / Baidu
Editor's Take: Data infrastructure is the picks-and-shovels play for the robotics gold rush. Mifeng's platform approach could become the "Scale AI of robotics" if it achieves network effects.
21. Chinese Humanoid Robot Sector Raises ¥93.5B in H1 2026
Summary: The Chinese humanoid robot industry raised ¥93.5 billion (~$13B) in the first half of 2026, surpassing the total for all of 2025. Morgan Stanley doubled its 2026 China humanoid shipment forecast to 50,000 units. Barclays projects a $200B market by 2035.
Source: 36氪 / Morgan Stanley / Barclays
Editor's Take: ¥93.5B in six months — the capital is flowing faster than the robots can be built. The question is whether deployment can keep pace with investment, or if we're entering a bubble phase.
🚗 Autonomous Driving
22. Huawei's Yinwang Helps Write UN's First Unified Autonomous Driving Regulation
Summary: Huawei's autonomous driving subsidiary Yinwang (引望), as a core member of China's expert group, deeply participated in developing the UN ADS GTR — the world's first unified global technical regulation for autonomous driving systems. The regulation was formally approved at UN/WP.29's 199th plenary session in Geneva, co-led by China, EU, UK, US, Canada, and Japan.
Source: IT之家 / Huawei Official / Phoenix Auto
Editor's Take: China is exporting standards, not just products. Huawei's 190M+ vehicles and 12.8B km of driving data gave it the real-world evidence to shape global rules. This is a paradigm shift from "following international standards" to "writing them."
23. XPeng VLA 2.0: 50.44% of Miles Now Driven Autonomously
Summary: XPeng's second-generation VLA (Vision-Language-Action) system, live for six months, now handles 50.44% of total driving miles. CEO He Xiaopeng completed a full end-to-end robotaxi journey in the GX vehicle with 4 Turing chips (3,000 TOPS). The system has been deployed to L4 Robotaxi fleets for internal testing.
Source: Phoenix Tech / Sina Auto
Editor's Take: 50% autonomous driving penetration in production vehicles is a genuine milestone. Camera+VLA at <80ms response, no LiDAR, no HD maps — XPeng is betting on software-defined autonomy for global scalability.
24. MONA L03 Global Launch in Munich: First China-Europe Simultaneous Release
Summary: XPeng's MONA L03 will hold its global launch event in Munich on July 16 (today), becoming the first Chinese vehicle simultaneously released in China and Europe. XPeng plans to enter 64 countries by end of 2026. The vehicle features the same VLA 2.0 model trained on Chinese roads, successfully tested on European roads without additional training.
Source: XPeng Official / Phoenix Tech
Editor's Take: "One model, two continents" — XPeng's approach proves that VLA trained on Chinese data can generalize to European roads. If MONA L03 succeeds in Europe, it validates China's AI-first automotive export strategy.
25. UN Approves World's First Unified Autonomous Driving Regulation (ADS GTR)
Summary: The UN's ADS GTR establishes a unified global framework covering L3-L5 autonomous driving systems, including safety requirements, manufacturer safety management systems, product safety files, and multi-pillar testing verification. It enables mutual recognition of type approvals across countries.
Source: UN/WP.29 / MIIT
Editor's Take: This is the most significant autonomous driving regulation in history. One test report, multiple country approvals — this eliminates the biggest barrier to global AV deployment: fragmented regulatory standards.
26. AutoX Achieves Full Driverless Operation Across All Shenzhen Districts
Summary: AutoX received approval for fully driverless (no safety driver) Robotaxi operations across all nine districts of Shenzhen, becoming the first company to achieve citywide full-autonomous coverage in China. The fleet covers urban, suburban, and highway scenarios.
Source: Shenzhen Government / 36氪
Editor's Take: Citywide coverage across nine districts — urban, suburban, highway — is the most comprehensive Robotaxi deployment in China. Shenzhen is becoming the world's most permissive large-city AV testing ground.
27. Huawei Qiankun ADAS: 1.9M Vehicles, 12.8B Kilometers, 4.92x Safer
Summary: Huawei's Qiankun (乾崑) intelligent driving solution is deployed in over 1.9 million production vehicles with cumulative assisted driving mileage exceeding 12.8 billion kilometers. Safety data shows vehicles using Qiankun ADAS are 4.92x safer than the Chinese road average when the system is engaged.
Source: Huawei Official / IT之家
Editor's Take: 1.9M vehicles and 12.8B km is a data moat that no competitor can quickly replicate. The 4.92x safety figure is a powerful marketing tool — and the foundation for Huawei's L3 regulatory applications.
28. Tesla's 20 Robotaxis vs Waymo's 3,500: The Scale Gap Widens
Summary: Tesla's Austin Robotaxi service operates approximately 20 vehicles, while Waymo has scaled to ~3,500 across multiple US cities. Tesla's safety record shows only 1 incident since April (rear-ended by human driver). Waymo reports 92% reduction in fatal accident rates.
Source: Phoenix News / IT之家 / Waymo
Editor's Take: The 175:1 fleet ratio tells the story. Tesla's camera-only approach may be more scalable long-term, but Waymo's data advantage from 3,500 vehicles is formidable. Two fundamentally different bets on the future of autonomy.
🌐 World Models
29. Xiaomi Open-Sources Robotics-U0: 38B-Parameter Unified Embodied World Model
Summary: Xiaomi released and open-sourced Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive world foundation model. It unifies four capabilities — embodied scene generation, embodied transfer, robot interaction video generation, and general text-to-image — in a single framework. Ranked #1 on WorldArena benchmark (126 models). FlashAR+ acceleration achieves 82.9x speedup. Real-robot tests show 26.3% improvement in task completion in unseen environments.
Source: Xiaomi Robotics / Zhidongxi
Editor's Take: U0 is the first model to unify embodied data generation across four task types. The 82.9x speedup (from 7.5 minutes to 5.44 seconds per image) makes scaled data production feasible. Outperforming GPT-Image-2.0 on controllability is a strong statement.
30. Ant Group Launches LingBot-World-Infinity: 1-Hour Coherent Video Simulation
Summary: Ant Group's LingBot-Brain 2.0 series includes LingBot-World, which evolves video generation into an interactive world simulator. It supports 720p/60fps real-time interaction and can generate up to 1 hour of coherent simulation video. Users can control via keyboard or natural language to change world state in real time.
Source: Ant Group Official
Editor's Take: From "generating video" to "interactive world simulation" — this is the critical leap from观赏 to utility. 720p/60fps real-time performance makes it viable for robot policy training. The fully open-source release is a bold move.
31. Three Schools of World Models Take Shape
Summary: The world model field has crystallized into three technical schools: Video Generation (Sora/Seedance — visual fidelity), Physics Simulation (Cosmos 3/LingBot-World — physical accuracy), and Unified Models (U0 — multimodal fusion). Each excels in different application scenarios.
Source: Industry Analysis
Editor's Take: These three schools aren't mutually exclusive — convergence is likely. The ideal world model combines video generation's visual quality + physics simulation's accuracy + unified models' flexibility.
32. VLA Emerges as Core Paradigm for Embodied Intelligence
Summary: Vision-Language-Action (VLA) models have become the core technical paradigm for embodied intelligence in 2026. XPeng's VLA 2.0 achieves 50% autonomous driving miles, Figure's Helix 02 validates VLA in factory settings, and NVIDIA's GR00T 1.7 provides the standardized VLA foundation model.
Source: Industry Analysis
Editor's Take: VLA is the bridge connecting "seeing," "understanding," and "acting." From autonomous driving to humanoid robots, VLA is becoming the universal language of embodied AI.
33. FlashAR+: 83x Acceleration for Embodied Data Generation
Summary: Xiaomi's FlashAR+ inference acceleration scheme, deployed with U0, combines diagonal parallel decoding with efficient cache scheduling and vLLM integration to compress image generation time from 450.77 seconds to 5.44 seconds per 1024×1024 image — an 82.9x speedup.
Source: Xiaomi Technical Report
Editor's Take: At 5.44 seconds per training image, mass-producing diverse embodied training data becomes economically viable. This is the engineering breakthrough that turns U0 from a research model into a production tool.
34. Zhiyuan's WITA: China's First Compliant-Filed Embodied AI Platform
Summary: Zhiyuan Robotics announced that its WITA embodied AI platform has become the first to receive regulatory compliance filing in China, establishing a precedent for the governance of embodied AI systems in commercial deployment.
Source: 36氪 / Zhiyuan Official
Editor's Take: Regulatory compliance filing is unglamorous but critical — it sets the standard for how embodied AI systems are governed. First-mover advantage in regulatory alignment can be a significant competitive moat.
35. U0 vs GPT-Image-2.0: Head-to-Head Embodied Generation Benchmark
Summary: In a 300-sample comparison test, Xiaomi's U0 significantly outperformed GPT-Image-2.0 on depth consistency, structural fidelity, and semantic alignment for embodied scene generation. GPT-Image-2.0 tends to produce object misalignment and spatial distortion in cross-view scenarios, which would corrupt robot training data.
Source: Xiaomi Technical Report / WorldArena
Editor's Take: For robot training data, geometric consistency matters more than visual beauty. U0's advantage isn't just "better pictures" — it's that the generated data actually works for training robots. This is the benchmark that matters for embodied AI.
🔔 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.