Physix Frontier🔔 East Voice · July 9, 2026 — China's AI Sovereignty Moment
July 9, 2026 (Thursday) ｜ Physix Frontier · East Voice Column
East speaks, West hears; the Chinese perspective.
📌 Editor's Note
Today's China AI landscape is defined by one word: sovereignty. Across every track — from foundation models to chips to autonomous vehicles — Chinese companies are building end-to-end independence from Western technology stacks. DeepSeek and Zhipu AI are secretly developing custom inference chips to escape reliance on both NVIDIA and even Huawei's Ascend. China's government formally legalized steering-wheel-free Robotaxis, creating the world's most permissive L4 regulatory framework. The humanoid robot sector crosses a critical inflection point with MIIT confirming a 100,000-unit production target — a 500%+ jump from last year. Meanwhile, Chinese AI models are quietly conquering the American market: on OpenRouter, Chinese models now account for 30-46% of all tokens consumed by US companies, up from 4.5% just a year ago.
The pattern is unmistakable: China is no longer playing catch-up. In multiple domains — video generation, humanoid manufacturing, autonomous driving deployment — it's setting the pace. The question for global observers isn't whether China will compete, but how quickly the rest of the world will adapt to China's tempo.
🤖 AI Large Models (8 stories)
🔥 DeepSeek and Zhipu AI Secretly Developing Custom Inference Chips — NVIDIA Shares Dip
Impact: ⭐⭐⭐⭐⭐
On July 7, Reuters and The Information broke simultaneous bombshells: China's two leading AI labs — DeepSeek and Zhipu AI — are both secretly developing custom AI inference chips, aiming to eliminate dependency on NVIDIA and even Huawei's Ascend processors.
The DeepSeek story is particularly striking. The company has been working on its chip project for approximately one year, engaging with external chip design firms, foundries, and memory companies — all while maintaining its trademark silence. The timing isn't coincidental: in June 2026, DeepSeek completed a record-breaking ¥51 billion (~$7B) Series A funding round, with founder Liang Wenfeng personally contributing ~¥20B as the largest single investor. Tencent, CATL, NetEase, JD.com, and IDG Capital all participated. The chip ambition explains why a company famous for refusing outside capital suddenly wanted $7 billion.
Zhipu AI's motivation is equally compelling but more urgent. Its GLM-5.2 model, launched in June, became the fastest-adopted model in Vercel's history — daily token consumption surged 27x and customer count grew 80x in a single week. This explosive demand crashed headfirst into China's chronic shortage of advanced chips. Unable to purchase NVIDIA's cutting-edge GPUs due to US export controls, and finding Huawei's Ascend capacity increasingly contested, Zhipu began exploring domestic ASIC partnerships — with an estimated 2+ year timeline to tape-out.
Why it matters for global readers: This isn't just a China story. It mirrors the exact strategic playbook executed by OpenAI (Jalapeño chip with Broadcom, launched June 2026), Google (TPU, now in its 6th generation), Amazon (Trainium), and Meta (MTIA). The difference is the constraint: Chinese companies must design for domestic foundries with mature process nodes, requiring exceptional design ingenuity to compensate for the manufacturing gap with TSMC's 3nm/5nm lines. If they succeed, it would represent one of the most significant vertical integration achievements in tech history — designing chips on domestic fabs, running domestic models, serving global users.
NVIDIA's pre-market stock dipped ~1.6%, though analysts note NVIDIA's China market share is already "effectively zero" due to export controls, making the actual financial impact negligible. The real pressure falls on Huawei, whose Ascend chips face a new domestic competitor.
(Sources: Reuters 7/7, The Information 7/7, 36Kr, Guancha.cn)
📌 Chinese AI Models Quietly Capture 30-46% of US Enterprise Token Usage
OpenRouter data reveals a seismic shift: since February 8, 2026, Chinese AI models have consistently accounted for over 30% of tokens consumed by American companies on the platform, peaking at 46%. For context, the 12-month average was just 11%, and H1 2025 was a mere 4.5%. AI startup Lindy switched its entire infrastructure from Anthropic to DeepSeek, finding comparable performance at 1/10th the cost — projecting millions in annual savings. Coinbase made DeepSeek and Zhipu's GLM its default models for engineers. As Brookings researcher Kyle Chan noted: "Companies used to prioritize capability without considering cost — now they're choosing the cheapest model that's good enough, and Chinese models are winning that race." Chinese open-source models are estimated to be 60-90% cheaper than frontier offerings from OpenAI and Anthropic.
(Sources: CNBC, Bloomberg, OpenRouter data)
📌 Meituan Open-Sources LongCat-2.0: 1.6 Trillion Parameters, Trained Entirely on Domestic Chips
Meituan released LongCat-2.0, a 1.6-trillion-parameter model that natively supports 1M token context windows — and was trained entirely on China's domestic computing infrastructure. The model serves as Meituan's unified AI backbone, powering internal tools like "Kangaroo Advisor" (merchant analytics) and "Smart Manager" (operations automation). While partially open-sourced, its strategic significance lies in proving that frontier-scale models can be trained without any Western chip dependency. The model shows strong performance on mathematical reasoning (AIME 2026 benchmarks) and code generation tasks.
(Source: GitHub, Meituan Tech Blog)
📌 GPT-5.6 Sol Launches as Chinese Model Wave Intensifies Competition
OpenAI released GPT-5.6 Sol on July 7, its flagship model priced at $5/M input tokens and $30/M output tokens. But the launch's timing is telling — it was strategically placed on the same day Anthropic's Claude Fable 5 limited-tier expired. Meanwhile, DeepSeek V4's official version is expected mid-July, and Zhipu's GLM-5.2 is already disrupting Western developer platforms. The competitive pressure is forcing pricing innovation across the industry.
(Source: Wall Street Insight)
📋 Quick Scan
- 7 National AI Agent Standards released (July 6): China's market regulator published interconnection standards covering identity, capability discovery, tool calling, security, and evaluation — potentially the "HTTP of the agent economy" (Source: State Administration for Market Regulation)
- Doubao and Qwen shut down C-end agent features (July 15): Commercial reality bites — each ¥1 of agent compute generates less than ¥0.10 in direct revenue; both pivot to B2B enterprise agents (Source: Multiple tech media outlets)
- WAIC 2026 set for July 17-20: 100,000 sqm exhibition space, 3,000+ global product debuts, 140+ forums, first self-hosted international academic conference with Turing Award laureates (Source: WAIC Official)
- World's first memristor neuromorphic chip (July 4): Peking University + Chinese Academy of Sciences achieve 50-478x speedup over advanced GPUs on brain cortex reconstruction tasks (Source: Peking University)
- Huawei publishes "Tao's Law" V2 (July 3): A time-scaling theory for multi-level electronic systems, marking China's shift from "following" to "originating" in semiconductor theory (Source: Huawei)
💻 AI Software (7 stories)
🔥 WPS Comate Launches: China's Enterprise AI Office Tools Enter Four-Way Battle
Impact: ⭐⭐⭐⭐
On July 3, Kingsoft Office (WPS) officially unveiled WPS Comate, its organizational-level AI agent platform — positioning it as an "enterprise brain" rather than just another copilot. The product launches mid-July, and its arrival completes what industry analysts are calling the "four-nation war" of Chinese enterprise AI office tools: DingTalk AI (Alibaba), Feishu Aily (ByteDance), Tencent WorkBuddy, and now WPS Comate.
What makes Comate different: Kingsoft's secret weapon is 38 years of document accumulation. The platform already stores nearly 3,000 billion documents across its cloud infrastructure — contracts, reports, spreadsheets, presentations — all digitized and vectorized since 2023 into a searchable enterprise knowledge base. No competitor has this depth of structured document understanding.
WPS Comate's "Three Connections + One Platform" methodology reflects hard-won enterprise learnings:
1. Connect Tokens: Provide company-wide AI compute budgets (Kingsoft internally consumes 300B+ tokens, growing from 200B to 300B in just three weeks, at ¥6.5-7/person/day)
2. Connect Data: Unified access to chat logs, meeting transcripts, documents, and business systems with granular permissions
3. Connect APIs: AI outputs directly become system actions — to-dos, meeting invites, purchase orders
4. One Platform: Centralized security, cost control, and shareable skill packages
Why it matters: In Q1 2026, Kingsoft Office revenue hit ¥1.613B (+24% YoY), with WPS 365 B2B revenue reaching ¥244M (+61% YoY). Internally, employees created 5,000+ custom AI "Skill" packages within the first month. One business unit reports 2026 H1 performance already matching all of 2025, projecting 300-400% YoY growth. The competitive landscape is shifting from "single-tool AI features" to full-platform AI infrastructure wars — and enterprise data moats are the new battleground.
The timing also coincides with Lei Jun's (Xiaomi founder) aggressive ¥5.47 billion HKD stake increase in Kingsoft Software over 10 trading days (June 16 – July 3), signaling major investor confidence in the AI office pivot.
(Sources: Sing Tao Daily, Kingsoft Office)
📌 Feishu PC Launches on HarmonyOS with Native Aily AI Integration
ByteDance's Feishu (Lark's China version) released a native PC client for Huawei's HarmonyOS, marking a significant expansion of the HarmonyOS productivity ecosystem. The client is purpose-built for desktop (not a mobile port), featuring native integration of Feishu Aily — the platform's AI assistant that understands user work context across docs, chats, calendars, and spreadsheets. Key features include built-in screenshot annotation with markup tools, customizable keyboard shortcuts, flexible cloud document management, and cross-ecosystem copy-paste compatibility with WPS Office and WeChat. The launch signals that China's "domestic tech stack" — from OS (HarmonyOS) to productivity software (Feishu) to AI (Aily) — is reaching maturity.
(Source: 21st Century Business Herald)
📌 DingTalk Multidimensional Table Gets Major AI Upgrade — Free for E-commerce, Manufacturing, F&B
DingTalk (Alibaba's enterprise collaboration platform) rolled out a significant AI upgrade to its multidimensional table feature, targeting e-commerce operations (one-click generation of product copy and posters for hundreds of SKUs), chain restaurant management (AI-driven menu item performance analysis), and manufacturing workflows. The update is part of DingTalk v7.7.0, which included nearly 100 UX improvements. The platform is positioning AI-powered low-code tables as the entry point for SME digital transformation — making enterprise-grade AI analytics accessible without any technical expertise.
(Source: DingTalk Official)
📋 Quick Scan
- China's GEO market approaching ¥100B with 169.7% YoY growth: Generative Engine Optimization (optimizing brand visibility in AI-generated answers) becomes a critical new marketing discipline, with 5.15B+ Chinese users querying AI platforms daily (Source: IDC, CNNIC)
- Enterprise AI Agent memory market exploding: From ¥1.44B (2025) to a projected ¥64.25B by 2030 — a 110% CAGR — as agents need persistent memory to function effectively in enterprise contexts (Source: industry research)
- Microsoft launches Frontier Company with $2.5B investment and a 6,000-person on-site AI deployment team, mirroring Amazon's $1B similar initiative announced two days earlier — signaling the global shift from "selling AI tools" to "delivering AI outcomes" (Source: Yicai Global)
🦾 Humanoid Robots (8 stories)
🔥 MIIT Confirms: 2026 Is China's Humanoid Robot Mass Production Year — 100,000 Units Target
Impact: ⭐⭐⭐⭐⭐
At the 2026 World Artificial Intelligence Conference on July 7, China's Ministry of Industry and Information Technology (MIIT) made the industry's most definitive statement yet: domestic humanoid robot production will exceed 100,000 units in 2026. For context, global humanoid robot shipments in all of 2025 were only 18,000-20,000 units, with China accounting for 14,400. This year's target represents a 500%+ increase — one year's output matching all previous years combined.
The policy framework is equally significant. MIIT and SASAC (State-owned Assets Supervision and Administration Commission) jointly launched a "Real-World Training" special initiative requiring state-owned enterprises to open real factory scenarios — loading/unloading, warehouse handling, equipment inspection — for humanoid robot deployment. The mandate: establish 100 standardized commercial deployment scenarios by year-end, transitioning the industry from "exhibition demos" to sustained procurement demand.
The reducer supply chain is the critical bottleneck and investment opportunity. Each humanoid robot requires 20-26 precision reducers (harmonic + RV types), representing 35-40% of total BOM cost. In 2026, domestic demand for humanoid robot reducers is projected to exceed 2 million units — a 6x increase from 2025. Key players:
- Leaderdrive: National harmonic reducer leader with 60%+ domestic share; sole Chinese supplier verified for Tesla Optimus; humanoid-specific 5M-unit production line launching Q3 2026; orders booked through 2027; Q1 revenue +180% YoY
- Shuanghuan Drive: #1 in domestic RV reducers; Q2 humanoid revenue ¥280M (doubled YoY); supplies Unitree H2, Agibot, and XPeng humanoid
- Zhongda Leader: Only A-share company producing all three reducer types (harmonic, RV, planetary); primary planetary reducer supplier to Unitree
Why it matters for global readers: China's humanoid robot industry is crossing from "technology validation" to "commercial scaling" at a speed unmatched globally. The combination of government mandate (forced deployment in state-owned factories), massive domestic demand (the world's largest manufacturing base), and a mature EV-derived supply chain (servos, sensors, computing modules) creates a uniquely powerful scaling engine. The 2035 global market is projected at ¥400B. Whoever wins the manufacturing race in the next 2-3 years will define the industry for a generation.
(Sources: MIIT, People's Daily Online, Toutiao)
📌 Unitree Technology STAR Market IPO Approved in Record 104 Days — "A-Share Humanoid Robot First Stock"
China Securities Regulatory Commission (CSRC) officially approved Unitree Technology's STAR Market IPO on July 2, just 104 days from application acceptance — a record for the STAR Market's pre-review mechanism. The company plans to raise ¥4.202 billion at an estimated valuation of ~¥42 billion. 85% of proceeds go directly to R&D: embodied AI models (¥2.02B), robot body development (¥1.11B), new product lines, and a manufacturing base.
Financials are striking: revenue grew from ¥159M (2023) to ¥1.699B (2025) — a 226.78% CAGR — while gross margins expanded from 44.2% to 60.1%. Unitree shipped 5,500 humanoid robots in 2025, capturing 32.4% global market share. Meituan, Tencent, Alibaba, Ant Group, Sequoia China, and Shunwei Capital (Lei Jun's fund) are all investors. The IPO triggered a sector-wide rally: 40+ stocks hit daily limit-up, with the humanoid robot sector market cap exceeding ¥11.8 trillion.
(Sources: Shanghai Stock Exchange, Securities Daily, IT Home)
📌 UBTECH Launches U1 Ultra-Bionic Humanoid Series — Orders Exceed 13,000 Units
UBTECH's June 30 global launch event unveiled the U1 series, positioned not as industrial tools but as emotional consumer products — daily companionship, emotional support, and social interaction. Pricing ranges from ¥119,800 (U1 Lite) to ¥990,000 (U1 Ultra male version). Despite 2-4 hour battery life (industry standard for full-size humanoids), orders surpassed 13,000 units. The product features 88 high-DOF joints, a "nurture-style" emotional AI model with encrypted local memory storage, and extensive appearance customization. Industry observers note U1's innovation isn't technical specs but category creation: transforming humanoid robots from "tech exhibition pieces" into "lifestyle proposals."
(Source: UBTECH Official, Multiple Media Outlets)
📌 20+ Chinese Humanoid Companies in IPO Pipeline — Capital Market Year Begins
The sector is experiencing an unprecedented capitalization wave. Agibot — with 15,000 units produced and ¥22B+ valuation — plans Hong Kong listing in H2 2026. Robot Era completed a ¥2.5B pre-IPO round. Fourier Intelligence has begun STAR Market preparation. In total, 20-30 companies across the humanoid value chain are in various stages of IPO filing, creating what analysts call the "2026-2027 golden capitalization window."
(Sources: Sina Finance, Multiple Brokerage Research Reports)
📋 Quick Scan
- Tesla Optimus: Fremont factory production line installation underway, first line operational within one week; year-end mass production target confirmed by Elon Musk; long-term design capacity of 1M units/year (Source: Tesla)
- BYD to deploy up to 20,000 humanoid robots internally by year-end 2026, signaling major industrial adoption (Source: industry reports)
- Domestic harmonic reducer market share surpasses 60%, with RV reducer shipments doubling YoY — rapid import substitution accelerating (Source: Industry Data)
- Humanoid robot prices drop below ¥10,000 for entry-level models, entering "major appliance" pricing territory (Source: Industry Report)
🚗 Autonomous Driving (7 stories)
🔥 Historic Breakthrough: China Legalizes Steering-Wheel-Free Robotaxis for Paid Operation
Impact: ⭐⭐⭐⭐⭐
On July 1, 2026, China achieved what may be the world's most significant autonomous driving regulatory milestone: the multi-ministry "Intelligent Connected Vehicle Access and Road Traffic Pilot Management Measures" entered full implementation. For the first time, L4 autonomous vehicles with no steering wheel, no accelerator pedal, and no driver's seat received official license plates for legal paid operation on public city roads.
This isn't a limited pilot in a controlled zone. The regulatory framework — jointly established by MIIT (vehicle safety certification), Ministry of Transport (operating permits), and Ministry of Public Security (accident liability) — creates a complete legal closed loop. The liability model is groundbreaking: when a driverless vehicle causes an accident, the vehicle owner and operating platform bear full legal responsibility, replacing the traditional driver-centric liability framework.
The economic case is compelling. With traditional taxis, labor accounts for 60%+ of operating costs. Beijing Yizhuang data shows Robotaxi performance metrics: 20+ trips per vehicle per day, average wait times under 3 minutes during peak hours, and per-kilometer costs below ¥2 — competitive with or cheaper than conventional taxis.
Baidu's Apollo RT6, the 6th-generation production Robotaxi, exemplifies the technology readiness: 1,200+ TOPS computing power, fully redundant wire-controlled chassis, vehicle cost under ¥204,600 — roughly 1/7th the cost of Waymo's vehicles and 30% cheaper than Tesla's CyberCab.
Why it matters for global readers: While Waymo leads in US commercial Robotaxi operations, China's regulatory approach is fundamentally different — and potentially more scalable. Where the US relies on state-by-state patchwork regulation, China created a unified national framework that simultaneously addresses vehicle certification, operating licenses, AND liability assignment. The BBC reported from Beijing Yizhuang that Baidu, WeRide, and Pony.ai already operate driverless Robotaxis alongside regular traffic. With 50+ cities now open for Robotaxi passenger service and some zones offering 24/7 fully driverless operation, China is building the world's largest real-world autonomous driving dataset.
The social impact is already visible: China has ~1.4M registered taxis and 7M+ ride-hailing driver certificates. Every city where Robotaxis expand represents real income displacement — making this as much a governance challenge as a technology one.
(Sources: MIIT, Ministry of Public Security, Ministry of Transport, BBC)
📌 UN Adopts China-Led Autonomous Driving Global Technical Regulation
The United Nations formally approved the ADS Global Technical Regulation (ADS GTR), a landmark framework covering L3-L4 autonomous driving that was led by China. The regulation achieves mutual recognition across China, EU, US, and Japan — dramatically reducing the cost of multi-market certification for Chinese AD companies. XPeng plans to bring its VLA (Vision-Language-Action) model overseas by 2027 with multi-language support. Pony.ai and WeRide are already operating commercial Robotaxi services in Europe and Singapore. The EU, with 12M annual new car sales and less than 15% AD penetration, represents the primary target market.
(Source: AutoNews China, UN regulation documents)
📌 Pony.ai Expands to Chongqing — First SW China Robotaxi Market
Pony.ai, one of the world's leading L4 autonomous driving companies, established a subsidiary in Chongqing's High-Tech Zone — marking its first expansion beyond tier-1 coastal cities. The company's fleet already exceeds 1,159 vehicles across Beijing, Shanghai, Guangzhou, and Shenzhen, with year-end targets of 3,000+. The Chongqing deployment is strategically significant: the city's mountainous terrain, complex overpasses, and dense urban environment represent some of the most challenging driving conditions in China — a powerful stress test for the technology.
(Source: Pony.ai, Chongqing High-Tech Zone)
📋 Quick Scan
- L3 officially legalized: First permits issued to Changan Deepal SL03 and BAIC Arcfox Alpha S (L3 edition) for highway/fast-road conditional autonomous driving (Source: MIIT)
- National unified AD traffic regulation (GAT 2388-2026) effective July 1: Introduces special "Z" license plates for L3 vehicles; establishes clear responsibility frameworks (Source: Ministry of Public Security)
- 50+ Chinese cities now open for Robotaxi passenger service, with some zones offering fully driverless 24/7 operation (Source: Local Policy Summary)
- Solid-state lidar drops to ~¥1,000 level, becoming standard equipment in vehicles priced from ¥100,000 — combined with 4D mmWave radar fusion for all-weather reliability (Source: Industry Data)
🌍 World Models / Physical AI (7 stories)
🔥 Seedance 2.5 Launches: ByteDance's Video Generation Enters "Director-Level" Era
Impact: ⭐⭐⭐⭐
On July 6, ByteDance officially launched Seedance 2.5 on its JiMeng (Dreamina) AI creation platform, with enterprise API access opening within one week. The model represents a generational leap from the already-impressive Seedance 2.0, achieving what industry observers call the "end of AIGC's childhood."
Three breakthrough capabilities define Seedance 2.5:
1. Director-level shot composition: The model autonomously handles cinematic shot selection — knowing when to use close-ups for emotional emphasis, wide shots for environmental context, and rapid cuts for tension escalation. In narrative aesthetics, reviewers note it "surpasses Sora 2.0."
2. Quad-modal input: Text, image, audio, and video can all serve as simultaneous inputs, with a novel @ syntax for precise asset referencing (e.g., "@image1 as protagonist, @video1 for camera movement, @audio1 as background score"). Character consistency across multi-shot sequences achieves a narrative break rate of just 2.6% versus the industry average of 9%+.
3. Subject transfer: Users can transfer their appearance onto any video subject with exact motion and lip-sync replication — essentially motion capture from a single photo and video. The feature generated both excitement and controversy, prompting ByteDance to implement strict identity verification for real-person transfers and invisible digital watermarking on all outputs.
Why it matters for global readers: China's video generation ecosystem has evolved into a genuine three-way race with US competitors. 36Kr's comprehensive review ranks the domestic leaders: Seedance 2.0/2.5 (ByteDance, "best overall"), Kling 3.0 (Kuaishou, "best for physical interaction and Chinese cultural context"), and Tongyi Wanxiang/Wan 2.2 (Alibaba, open-source enterprise deployment). Each has carved distinct niches — ByteDance dominates consumer creativity, Kuaishou excels in physically accurate human-object interactions, and Alibaba targets enterprise/commercial video with open-source flexibility.
The competitive dynamic is accelerating the entire field. Seedance 2.5's enterprise API opening signals ByteDance's ambition to become infrastructure for the global creative industry, not just a consumer app.
(Sources: Volcano Engine, 36Kr, AITOP100)
📌 WAIC 2026 to Host First-Ever AI Film Forum — World Model Meets Hollywood
Wondershare Technology and Shengshu Technology will co-host WAIC's first dedicated AI Film forum on July 20, themed "New Paradigm of AI Film Production Driven by World Models." The event assembles the full AI filmmaking value chain: Wondershare founder, Shengshu founder Prof. Zhu Jun (Tsinghua AI Institute), Alibaba Cloud's Liu Weiguang, plus TV/film industry leaders including producers from Bona Film Group and Linghe Culture. Wondershare's "JuChang" platform — China's first full-pipeline AI premium video creation tool — will demo its new "Infinite Canvas + Multi-Module Agent Workflow" dual-mode interface. Shengshu, founded in 2023, has built a general world model covering multimodal generation, real-time interaction, and physical-world action.
(Source: Shanghai Securities News)
📌 AI Video Generation Reaches Full Maturity — 36Kr's Definitive Tool Comparison
36Kr published a comprehensive two-week evaluation of dozens of AI video tools, narrowing to six finalists across domestic and international offerings:
- Domestic winners: Seedance 2.0 (ByteDance, "leading overall"), Kling 3.0 (Kuaishou, "best Chinese cultural context"), Tongyi Wanxiang Wan 2.2 (Alibaba, "enterprise-grade open-source")
- International competitors: Sora 2.0 (OpenAI, "Hollywood-level physics"), Veo 3.1 (Google, "best audio-visual sync"), Runway Gen-4.5 ("professional post-production toolkit")
The assessment: AI video tools have completed the "usable → good → professional" triple jump in just two years, with creative barriers reaching historic lows.
(Source: 36Kr)
📋 Quick Scan
- Kling 3.0 (Kuaishou): Achieves near-perfect scores on the famous "Smith eating noodles" physical interaction benchmark; excels in Chinese narrative comprehension and cost-effective long video generation (Source: 36Kr Review)
- MoWorld — Huawei Ascend native: 14B-parameter MoE world model achieves 50+ FPS real-time inference on Huawei Ascend 910C CloudMatrix384 NPU, demonstrating domestic chip viability for world model workloads (Source: Academic Paper)
- Li Fei-Fei's WorldScore benchmark: First unified evaluation framework comparing 3D scene generation, 4D scene generation, and video generation models — tested on 19 models across 3,000 test cases (Source: Stanford/CSDN)
- WAIC "World Model Six Dragons" roundtable: Six leading players (Daxiao Robot, Ant Lingbo, Jijia Shijie, Wujie Dongli, Agibot, Zibianliang Robot) debate technical routes, data sources, and deployment paths for world models transitioning to physical execution (Source: WAIC Official)
📊 China AI Market Snapshot
| Indicator | Value | Source |
|---|---|---|
| Chinese model share of US enterprise tokens (OpenRouter) | 30-46% (up from 4.5% in H1 2025) | OpenRouter |
| DeepSeek A-round valuation | ~¥400B ($55B) | Public Reports |
| GLM-5.2 Vercel token growth (1 week) | 27x daily, 80x customers | Vercel |
| Meituan LongCat-2.0 parameters | 1.6 trillion | GitHub |
| 2026 humanoid robot production target (China) | 100,000 units | MIIT |
| Unitree IPO valuation | ~¥42B ($5.8B) | Prospectus |
| Unitree 2025 humanoid shipments | 5,500 units (32.4% global share) | Prospectus |
| Robotaxi fleet: Pony.ai | 1,159 vehicles, targeting 3,000 by year-end | Public Data |
| Baidu Apollo RT6 cost | ¥204,600 (~1/7 of Waymo) | Baidu |
| Cities open for Robotaxi service | 50+ | Local Policies |
| China GEO market growth | 169.7% YoY, approaching ¥100B | IDC |
| WPS internal token consumption | 300B+ (200B→300B in 3 weeks) | Kingsoft Office |
| Kingsoft Office Q1 2026 revenue | ¥1.613B (+24% YoY) | Financial Report |
| AI job openings growth (Jan-Apr 2026) | 8.7x YoY | Maimai |
| AI scientist average monthly salary | ¥130,000+ | Maimai |
| 2026 tech giant AI capex (global) | $800B | Industry Statistics |
🧠 East Voice · Editor's Take
The Chip Is the New Model
For the past two years, the AI industry's defining competition was about who could build the smartest model. This week, China made it clear: the next defining competition is about who can build the chip to run it.
DeepSeek and Zhipu joining the custom silicon club isn't surprising — every major AI lab worldwide is pursuing the same strategy. What's remarkable is the constraint under which they operate. Cut off from TSMC's cutting-edge nodes and increasingly dependent on a Huawei Ascend ecosystem that can't meet their exploding demand, these companies are attempting something unprecedented: designing world-class inference chips for domestic foundries with mature process nodes. It's like trying to build a Formula 1 engine using only parts available at a local auto shop — the engineering creativity required is arguably greater than what OpenAI or Google face.
The deeper signal is strategic: China's AI industry is completing a full-stack vertical integration that no other nation has attempted. Domestic models (DeepSeek, GLM, Qwen) + domestic chips (Ascend + upcoming custom ASICs) + domestic foundries + domestic deployment (the world's largest manufacturing base for humanoid robots and the most permissive Robotaxi regulations globally). If even half of this stack works at scale, it creates an AI ecosystem that operates entirely outside the Western technology supply chain — not by choice, but by necessity born of export controls.
For global businesses, the practical implication is immediate: Chinese AI models already offer 60-90% cost savings at competitive performance. As the infrastructure stack matures, that cost advantage will only widen. The question isn't whether to use Chinese AI — it's whether you can afford not to.
📅 Coming Up This Week
- July 15: New "Interim Measures for AI Anthropomorphic Interaction Services" take effect; Doubao/Qwen C-end agent shutdown
- July 17-20: WAIC 2026, Shanghai — 100,000 sqm, 3,000+ product debuts, Turing laureates
- Mid-July: DeepSeek V4 official release expected
- Mid-July: WPS Comate official launch
- Late July: Tesla Optimus Fremont production line goes live
- H2 2026: Agibot, Robot Era HK IPO filings
🔔 East Voice — Physix Frontier's China AI column for global readers
East speaks, West hears; the Chinese perspective.
© 2026 Physix Frontier · bbs.physixfrontier.com
