Physix FrontierEast Voice | 2026-07-11: Unitree IPO · 100K Humanoid Robots · LongCat-2.0 on Domestic Chips · WAIC 2026
July 11, 2026 · Physix Frontier
Editor's Note
Today's keywords: Unitree IPO / 100K Humanoid Robots / MIIT Confirmation / LongCat-2.0 on Domestic Chips / WAIC 2026 Preview
China's humanoid robot sector hits a historic inflection point: Unitree's IPO registration takes effect in a record 104 days with a ¥4.2B fundraise, UBTECH's U1 series surpasses 13,000 pre-orders on debut, and MIIT confirms annual humanoid robot output will exceed 100,000 units in 2026. On the large model front, Meituan's LongCat-2.0 becomes the first trillion-parameter model trained entirely on domestic chips, while the industry braces for Anthropic's China service cutoff fallout — Alibaba mandates a full migration from Claude to its in-house coding tool Qoder. WAIC 2026 (July 17–20) looms as the next major showcase, with Huawei's Atlas 950 SuperPoD making its physical debut.
I. AI Large Models
1. Meituan Unveils LongCat-2.0: First Trillion-Parameter Model Trained Entirely on Domestic Chips
Summary: Comparable to Google Gemini 3.1 Pro, LongCat-2.0 is reportedly the first trillion-parameter model trained wholly on domestically developed hardware, a milestone for China's AI self-reliance amid U.S. chip export restrictions.
Source: Economic Times
Editor's Take: This is a proof-of-concept that China can train frontier-scale models without NVIDIA hardware. The real question is whether domestic chips can sustain the iteration speed and cost-efficiency needed for competitive model development at scale.
2. Alibaba's T-Head Launches "Zhenwu 810E" GPU Chip Targeting NVIDIA H20
Summary: T-Head Semiconductor released the Zhenwu 810E GPU, benchmarked against NVIDIA's H20, already deployed in multiple 10,000-card clusters on Alibaba Cloud serving 400+ enterprise clients including State Grid and XPeng.
Source: Toutiao / AI Deep Dive
Editor's Take: Alibaba's chip strategy has shifted from backup plan to primary infrastructure. With 400+ paying customers on domestic silicon, the "can we use it?" question is answered — now it's about "how fast can we scale?"
3. ByteDance Orders 350,000 Huawei Ascend Chips Worth ~¥40B
Summary: ByteDance has signed orders for approximately 350,000 Huawei Ascend chips (~$5.6B) for training, while simultaneously engaging Tianshu Zhixin and Enflame for inference workloads.
Source: AI Deep Dive / Toutiao
Editor's Take: ByteDance's bet is the largest single domestic chip order to date. At this scale, it's not a pilot — it's a strategic commitment. If Ascend delivers at scale, it fundamentally reshapes China's AI compute supply chain.
4. DeepSeek Secretly Developing Inference Acceleration Chip
Summary: DeepSeek is reportedly building a custom inference acceleration chip, with the project started approximately one year ago and still in early stages.
Source: AI Deep Dive / Toutiao
Editor's Take: DeepSeek's approach mirrors the global trend — frontier labs are vertically integrating their hardware stack. An inference-specific chip suggests DeepSeek is optimizing for deployment economics, not just training benchmarks.
5. Baidu's Kunlunxin Targets $50B Hong Kong IPO
Summary: Baidu's chip subsidiary Kunlunxin has initiated its Hong Kong IPO process, targeting a ~$50B valuation. Kunlunxin has built full-stack capabilities spanning chips, clusters, and software stacks.
Source: AI Deep Dive / Toutiao
Editor's Take: A $50B valuation for a domestic AI chip company signals strong market confidence in China's silicon independence narrative. Kunlunxin's full-stack positioning differentiates it from pure-play chip designers.
6. Alibaba Bans Claude Internally, Switches to In-House Coding Tool Qoder
Summary: Alibaba issued an internal mandate on July 3 requiring all employees to uninstall Claude by July 10 and migrate to its proprietary coding tool Qoder, powered by the Qwen model series.
Source: AI Deep Dive / Toutiao
Editor's Take: This is the most aggressive corporate response to Anthropic's China service cutoff. By forcing 100% migration in one week, Alibaba is betting that forced adoption of domestic tools will accelerate their maturation — a high-pressure strategy with high potential payoff.
7. Zhipu AI Releases GLM-5.2 Open Model
Summary: GLM-5.2 performs nearly as well as top offerings from Anthropic and OpenAI across several benchmarks, released as open-source amid growing demand for non-U.S.-controlled AI systems.
Source: Economic Times
Editor's Take: GLM-5.2's timing is impeccable — released just as Anthropic's China cutoff creates a vacuum. For enterprises that lost Claude access overnight, GLM-5.2 is the most credible immediate replacement.
II. AI Software & Applications
1. Open Compute Conference 2026: Tackling AI Infrastructure Bottlenecks
Summary: The OCP conference in Beijing focused on trillion-parameter models and multi-agent token explosion, addressing challenges in high-speed interconnects, power delivery, and liquid cooling with 50+ global enterprises including Samsung, ByteDance, Alibaba, and Baidu.
Source: China News Service
Editor's Take: The conference's core message — infrastructure must evolve from "supporting single-model inference" to "supporting swarm intelligence and multi-modal fusion" — reflects the industry's shift from model-centric to system-centric thinking.
2. StepFun Announces First AI Agent Smartphone
Summary: StepFun (founded by Kuang Yinchi) will launch a dedicated AI agent phone brand, manufactured by Huaqin Technology. This positions StepFun ahead of OpenAI's planned 2027 hardware release.
Source: CSDN / STAR Market Daily
Editor's Take: An AI-native phone represents the boldest bet on edge AI agents yet. StepFun's thesis — that AGI requires physical-world interaction, not just language — could reshape the smartphone category if execution matches ambition.
3. China's AI Personification Regulation Takes Effect July 15
Summary: The "Interim Measures for AI Personification Interaction Services" — China's first regulation specifically targeting AI companionship services — takes effect July 15, covering virtual partners, emotional companions, and related agents.
Source: CSDN / Government Announcement
Editor's Take: This regulation signals that China's AI governance is moving from broad strokes to surgical precision. The focus on "emotional AI" services reflects growing societal concerns about human-AI attachment.
4. Greater Bay Area Ascend 10K-GPU Cluster Goes Live in Shaoguan
Summary: A domestically built compute cluster with over 10,000 Huawei Ascend AI accelerators has been deployed in Shaoguan, Guangdong, marking a new phase in China's "compute sovereignty" strategy.
Source: CSDN / Toutiao
Editor's Take: A 10,000+ card domestic cluster is a significant scale milestone. It demonstrates that Huawei's Ascend ecosystem can now support serious training workloads at meaningful scale.
5. ACL 2026: Chinese Researchers Dominate Best Paper Awards
Summary: All three best paper first-authors at ACL 2026 are Chinese researchers, with Chinese authors also dominating the outstanding paper list. Conference submissions surged 45% to a record high.
Source: CSDN / Toutiao
Editor's Take: Chinese NLP/AI research continues to strengthen its global position. The 45% submission growth also signals the field's rapid expansion — though quality control at this scale remains a concern.
6. Beijing's AI for Science Achievements: DPA-4 Tops International Materials Benchmark
Summary: Beijing's DPA-4 large atomic model leads the authoritative international materials discovery benchmark, while the "Panshi 100" model system now serves 2,300+ research institutions and 30,000+ users.
Source: CSDN / Toutiao
Editor's Take: AI for science is one of the most impactful yet under-covered AI applications. DPA-4's success demonstrates that China's AI capabilities extend well beyond consumer-facing language models.
III. Humanoid Robots
1. Unitree Technology IPO Registration Takes Effect — A-Share's First Profitable Humanoid Robot Company
Summary: Unitree's IPO on the STAR Market was approved in a record 104 days, raising ¥4.2B at an estimated ~¥42B valuation. In 2025, the company shipped 5,500+ humanoid robots (global #1) and 30,000+ quadruped robots cumulatively, with revenue surging from ¥159M (2023) to ¥1.7B (2025).
Source: 36Kr / Wanlian Securities
Editor's Take: Unitree is the defining IPO of China's humanoid robot wave. The 10x revenue growth in two years and actual profitability prove this isn't hype — it's a real business. The ¥20.2B allocated to robot model R&D signals that the next competitive frontier is intelligence, not just locomotion.
2. UBTECH Launches Ultra-Bionic U1 Series, First-Day Orders Exceed 13,000 Units
Summary: UBTECH unveiled three U1 variants (Lite at ¥119,800, Pro at ¥169,800, Ultra at ¥880-990K) with 88 degrees of freedom, 20+ emotion recognition at 90%+ accuracy, and a "human-machine symbiosis" strategy for the next decade.
Source: Wanlian Securities / Industry Reports
Editor's Take: 13,000+ orders on launch day is remarkable for a humanoid robot product line. The tiered pricing strategy (¥120K–¥990K) cleverly addresses both consumer and premium commercial segments. The emotion recognition capability suggests UBTECH is betting on companionship as the killer use case.
3. MIIT Confirms 2026 Humanoid Robot Output to Exceed 100,000 Units
Summary: China's Ministry of Industry and Information Technology confirmed that national humanoid robot production will surpass 100,000 units in 2026, a 10x increase from 2025's 14,400 units.
Source: CSDN / MIIT
Editor's Take: 10x year-over-year growth in unit production is the clearest signal that China's humanoid robot industry has crossed from "concept" to "manufacturing." With 140+ domestic manufacturers and over half of provincial governments including robotics in their 2026 work plans, the policy-industry alignment is unprecedented.
4. Boundless Power K15 Receives World's First Industrial-Grade Full-Domain CE Certification
Summary: Boundless Power's K15 robot became the first humanoid robot globally to achieve full industrial-grade CE certification for EU market access, with nearly $100M in global orders now entering delivery.
Source: CSDN
Editor's Take: CE certification is the gateway to European industrial markets. Being first to achieve full-domain certification gives Boundless Power a significant first-mover advantage in what could become a major export market for Chinese humanoid robots.
5. Qiyuan Robot Opens First South China Retail Store in Guangzhou
Summary: Qiyuan (subsidiary of Shangwei New Materials, 688585) opened a humanoid robot experience store in a home furnishing complex, letting consumers interact with the Q1 robot in real living-room scenarios.
Source: CSDN
Editor's Take: A humanoid robot retail store in a furniture mall — this is the "iPhone Store moment" for consumer robotics. Physical retail presence signals confidence in the product and creates the direct consumer feedback loop that's critical for mass-market products.
6. Tesla Optimus V3 Production Line Ready, Targeting 1,000/Week by September
Summary: Tesla's Fremont factory has completed conversion of its Model S/X line to Optimus production. The V3 features 37 joints, 22-DOF dexterous hands (0.08mm precision), and targets $20-30K pricing with public sales expected in 2027.
Source: Eastmoney / PConline
Editor's Take: Tesla is applying its automotive manufacturing DNA to robotics at unprecedented scale. The 1M annual capacity plan and sub-$30K price point would, if achieved, make Optimus the most accessible humanoid robot on the market. The supply chain order data (hundreds of units for August, scaling to 10K/year capacity) adds credibility.
7. Embodied Intelligence Funding Surges Past ¥90B in H1 2026
Summary: Chinese embodied intelligence startups raised over ¥90B in the first half of 2026, a 5x increase year-over-year, with 300+ funding events. Individual rounds of ¥1B+ have become routine.
Source: CSDN / Titanium Media
Editor's Take: ¥90B in six months is staggering. The shift in investor evaluation criteria — from "impressive demo videos" to "actual delivery orders and repeat purchases" — signals the market's maturation. The question is no longer "can they build it?" but "can they sell it at scale?"
IV. Autonomous Driving
1. XPeng Robotaxi Begins Internal Testing — Full Pipeline in 8 Months
Summary: XPeng's Robotaxi program went from announcement (Nov 2025) to internal testing in just 8 months, with the first production vehicle rolling off the line in May. CEO He Xiaopeng called the pace "beyond our initial expectations."
Source: CSDN / Industry Reports
Editor's Take: Eight months from announcement to road testing is remarkably fast. XPeng's advantage is leveraging its existing ADAS platform and manufacturing infrastructure — avoiding the cold-start problem that plagues pure-play Robotaxi startups.
2. XPeng's Second-Generation VLA Achieves Global Generalization Capability
Summary: XPeng's second-generation Vision-Language-Action model is designed for global generalization, significantly reducing deployment costs across different cities and road conditions.
Source: Industry Reports
Editor's Take: Global generalization is the holy grail of autonomous driving. If XPeng's VLA can truly generalize across diverse road environments without per-city retraining, it could leapfrog competitors still relying on HD maps and geofenced operations.
3. Huawei Atlas 950 SuperPoD to Debut at WAIC 2026
Summary: Huawei's Atlas 950 — the industry's largest-scale supernode linking 8,192 accelerator cards — will make its first physical exhibition at WAIC 2026 (July 17–20, Shanghai).
Source: CSDN / WAIC Official
Editor's Take: An 8,192-card supernode is the compute backbone needed for training autonomous driving foundation models at scale. Huawei's physical debut at WAIC signals readiness to compete with NVIDIA's DGX ecosystem in the autonomous driving compute space.
4. China's Robotaxi Market Heats Up with Multiple Players Entering Testing Phase
Summary: Multiple Chinese companies including Baidu Apollo, Pony.ai, and AutoX continue to expand their Robotaxi testing operations across major Chinese cities.
Source: Industry Reports
Editor's Take: China's Robotaxi ecosystem is the world's most competitive, with at least 5 major players in active commercial testing. The regulatory environment — more permissive than the U.S. for driverless operations — gives Chinese players a structural advantage in accumulating real-world miles.
5. Dongfang Suanxin to Debut World's First Software-Defined Near-Memory Computing 3D Chip at WAIC
Summary: Dongfang Suanxin will showcase the DF1000, described as the world's first software-defined near-memory computing 3D chip, at WAIC 2026.
Source: WAIC Preview / Industry Reports
Editor's Take: Near-memory computing architectures could fundamentally change the economics of autonomous driving inference by dramatically reducing data movement costs. The DF1000's debut at WAIC will be closely watched by the AD compute community.
6. First Multimodal Spatiotemporal Foundation Model for Urban Spaces to Be Showcased
Summary: WAIC 2026 will feature the first multimodal spatiotemporal foundation model designed for urban-scale applications, targeting autonomous driving and smart city use cases.
Source: WAIC Preview
Editor's Take: Urban-scale spatiotemporal models represent the convergence of autonomous driving, smart city infrastructure, and digital twin technologies. China's unique advantage — massive, diverse urban datasets — positions it well in this emerging category.
V. World Models / Physical AI
1. WAIC 2026 Preview: 100,000 sqm Exhibition with World Model Focus
Summary: The 2026 World AI Conference (July 17–20, Shanghai) will feature world models as a core theme, with exhibitions exceeding 100,000 square meters for the first time. Key showcases include Huawei Atlas 950, MiniMax M3 multimodal model, and StepFun Agent OS.
Source: WAIC Official / Shanghai Government
Editor's Take: WAIC's elevation to a 100,000 sqm event with "world models" as a headline theme reflects China's strategic prioritization of physical AI. The conference will serve as a bellwether for how far Chinese world model research has progressed.
2. GigaWorld-1: 324,000 Simulated Rollouts for Robot Policy Evaluation
Summary: Published on arXiv, GigaWorld-1 presents WMBench — a systematic benchmark for evaluating world models as robot policy evaluators, based on 324,000 simulated rollouts paired with real robot executions and 12,000+ hours of training video.
Source: arXiv (2607.02642)
Editor's Take: The core finding — that long-horizon action fidelity matters more than short-term visual realism for policy evaluation — should redirect significant R&D resources. World models optimized for "pretty frames" may be fundamentally misaligned with what robotics actually needs.
3. JiYing Technology Releases "JiYing 2.0-s" Solid Mechanics Physics Foundation Model
Summary: The model can solve previously unseen physics problems — with new geometries, loading conditions, and materials — without retraining, achieving accuracy close to numerical simulation methods.
Source: Industry Reports
Editor's Take: A physics foundation model that generalizes to unseen geometries and materials is a major step toward "physical intuition" for AI systems. This is precisely the kind of capability needed for robots to reason about physical interactions in novel environments.
4. Chinese Embodied AI Funding Hits ¥90B in H1 2026
Summary: The massive capital influx into embodied AI — a 5x YoY increase — reflects investor conviction that physical AI is the next frontier after language models. Key themes include manipulation, locomotion, and multi-modal perception.
Source: Titanium Media / CSDN
Editor's Take: The ¥90B figure confirms that China's VC community has made a collective bet on physical AI as the successor to the LLM wave. Whether this capital translates into commercially viable products remains the trillion-yuan question.
5. Open Compute Conference: From Single-Model Inference to Swarm Intelligence Infrastructure
Summary: Industry consensus at the OCP Beijing conference: next-generation AI infrastructure must support thousands of concurrent agents, not just single-model serving. Inspur proposes "native liquid cooling" CPU architectures to handle agent swarm workloads.
Source: China News Service
Editor's Take: The shift from "running one big model" to "orchestrating thousands of agents" fundamentally changes infrastructure requirements. This is the physical AI infrastructure story — the compute layer must evolve as fast as the intelligence layer.
6. Mistral AI Enters Robotics with Robostral Navigate
Summary: French AI lab Mistral launched Robostral Navigate, a model purpose-built for robot navigation, signaling the convergence of frontier LLM capabilities with embodied AI applications.
Source: Industry Reports
Editor's Take: When a leading European LLM lab pivots into robotics, it confirms that "language model → world model → robot policy" is becoming the canonical technology transfer path in AI.
Track Statistics
| Track | Count |
|---|---|
| AI Large Models | 7 |
| AI Software & Applications | 6 |
| Humanoid Robots | 7 |
| Autonomous Driving | 6 |
| World Models / Physical AI | 6 |
| Total | 32 |
Physix Frontier · East Voice | Daily China AI Intelligence Briefing
🔔 East Voice — Physix Frontier's China AI column for global readers
East voice reaches west, Chinese perspective.
© 2026 Physix Frontier · All rights reserved