East Voice — China AI Daily Briefing

Saturday, July 26, 2026

---

Editor’s Note

This week marks a defining moment for China’s AI ambitions. DeepSeek — the startup that stunned Silicon Valley with its efficient models — has suspended its second fundraising round (targeting $71B valuation) after leaked founder remarks about China’s chip dependency sparked controversy. Meanwhile, the humanoid robotics sector is on fire: Agibot (智元) is heading for a Hong Kong IPO at up to HK$50B valuation, Unitree debuted the world’s first mass-produced manned mech in Chengdu, and China’s quadruped robots now command nearly 70% of global sales. On the autonomous driving front, WeRide secured its fourth consecutive win at China’s Intelligent Driving Championship, while Baidu’s Apollo Go expanded into Central Asia. The message is clear: China’s AI industry is transitioning from lab breakthroughs to industrial-scale deployment.

---

Track 1: AI Large Models（8 items）

1. DeepSeek Suspends $1.5B Fundraising Round Amid Leaked Founder Remarks
DeepSeek has verbally informed prospective investors that planned investment agreements will not proceed in the coming days. The suspension stems from founder Liang Wenfeng’s frustration over leaked meeting transcripts in which he acknowledged China’s reliance on Nvidia chips and a 12-18 month gap behind US frontier models. The company had been targeting at least ¥10B ($1.38B) at a pre-money valuation of ¥480B ($71B), up from $50B in its record $7B first round backed by Tencent, CATL, and the National AI Industry Investment Fund. IPO preparations continue.

Source: Bloomberg / Yicai / Crypto Briefing
Editor’s Take: The pause is damage control, not retreat. Liang’s wealth has surged to $36B — making him the world’s richest AI model creator — and DeepSeek’s open-source credentials remain its strongest asset.

2. Kimi K3: 2.8-Trillion-Parameter Open-Source Model Tops Arena.ai
Moonshot AI’s (月之暗面) Kimi K3, with 2.8 trillion parameters, is now the world’s largest open-source model. Within 24 hours of launch, it topped Arena.ai’s front-end code arena, winning first place in six out of seven sub-categories, sending shockwaves through Silicon Valley.

Source: Multiple Chinese tech media
Editor’s Take: Kimi K3 proves that China’s open-source LLM ecosystem can compete head-to-head with the best closed-source models from the US — at a fraction of the cost.

3. Musk Praises China’s AI Prowess, Names Kimi K3
In a wide-ranging interview with The Economist, Tesla CEO Elon Musk stated that “China could very well become the AI leader,” specifically citing his impression with Moonshot AI’s Kimi K3. Musk highlighted China’s decisive advantages in electricity supply (estimating 3x US capacity by 2026), chip self-sufficiency progress, and engineering talent.

Source: The Economist / Toutiao
Editor’s Take: When the world’s most prominent tech CEO publicly acknowledges China’s AI competitiveness — and names a specific Chinese model — it signals a genuine shift in perception, not just diplomacy.

4. China’s Commerce Ministry Deliberates AI Export Controls
China’s Commerce Ministry has convened multiple closed-door sessions with ByteDance, Alibaba, Zhipu, Moonshot AI, and other leading AI companies to develop export control regulations. Key targets include frontier model weights, core training datasets, advanced AI chip designs, and embodied intelligence core technologies — all subject to tiered approval and export licensing.

Source: Reuters / Financial Times / Toutiao
Editor’s Take: This is a mirror response to US chip export controls. China is building its own “small yard, high fence” — but targeting knowledge and model IP rather than hardware.

5. ~200 Silicon Valley Founders Petition Against Banning Chinese AI Models
Nearly 200 Silicon Valley startup founders jointly wrote to President Trump, urging the government not to cut off access to Chinese open-source AI models. The letter argues that even if banned, these models would continue to spread — while US startups would lose competitiveness. Nvidia CEO Jensen Huang echoed this, stating open models enhance safety through external review.

Source: NetEase / Multiple sources
Editor’s Take: The data speaks volumes: Chinese model token usage on aggregation platform OpenRouter accounts for nearly 60% of total volume. Banning what your customers already depend on is a losing strategy.

6. OpenAI President Praises Kimi K3 as “Quite a Good Model”
OpenAI President Greg Brockman publicly acknowledged Kimi K3, calling it “a pretty good model, there’s no doubt about that” — a rare compliment from a direct competitor.

Source: NetEase
Editor’s Take: When your rival compliments you publicly, it means your work has become impossible to ignore.

7. Stanford Report: US-China AI Gap Narrows to 2.7%
Stanford University’s 2026 AI Index Report finds the performance gap between top US and Chinese LLMs has shrunk from 31.6% in 2023 to just 2.7%. The UK AI Safety Institute estimates the iteration cycle gap has compressed to 2-3 months.

Source: Stanford HAI / Toutiao
Editor’s Take: The numbers tell the real story: China’s AI capability is converging with the US faster than anyone predicted, and the gap is now measured in months, not years.

8. Xiaomi Launches ZipVoice Speech Synthesis Models
Xiaomi AI Lab’s Kaldi team released the ZipVoice series based on Flow Matching architecture, including ZipVoice (zero-shot single-speaker TTS) and ZipVoice-Dialog (zero-shot conversational TTS), addressing key limitations in parameter count and synthesis speed.

Source: XM Platform
Editor’s Take: Speech synthesis is a critical building block for embodied AI — Xiaomi is quietly stacking capabilities across the full AI stack.

---

Track 2: AI Software & Applications（6 items）

1. Xiaomi Releases Robotics Foundation Model: 100K Hours of Real-World Data
Xiaomi unveiled Xiaomi-Robotics-1, a foundation model for embodied AI trained on over 100,000 hours of real-world data. The model achieves top rankings on multiple global benchmarks and demonstrates “out-of-the-box” capability — a first for validating scaling laws in robotics with real-world data.

Source: Multiple Chinese tech media
Editor’s Take: This is the first major validation of scaling laws for robotics using real-world (not simulated) data. Xiaomi’s bet: the same approach that worked for LLMs will work for robots.

2. Photonic-Electronic Fusion Chip Could Cut Token Costs by 50%
A breakthrough in photonic-electronic fusion chip technology promises to reduce AI inference token costs by approximately 50%, potentially reshaping the economics of large-scale AI deployment.

Source: Daily Economic News
Editor’s Take: If photonic chips deliver on this promise, the cost structure of AI inference — currently dominated by GPU pricing — could fundamentally shift.

3. Mianbizhineng (面壁智能) Raises Over ¥5B in H1 2026
Chinese AI startup Mianbizhineng (ModelBest) has accumulated over ¥5 billion in funding during the first half of 2026, signaling continued investor confidence in China’s AI foundation model ecosystem.

Source: 36Kr / Zhou Bao
Editor’s Take: Despite global funding headwinds, top-tier Chinese AI labs continue to attract massive capital — the market is betting on China’s AI catch-up story.

4. Kuaishou’s Kling AI Raises ~$3B at $18B Valuation
Kuaishou’s AI video generation subsidiary Kling AI has secured approximately $3 billion in funding, with Abu Dhabi’s BlueFive co-leading the round. Valuation reaches $18 billion.

Source: 36Kr / Zhou Bao
Editor’s Take: Video generation AI is the next frontier — and Middle Eastern sovereign capital is betting heavily on Chinese players.

5. China’s AI Export Controls to Include Embodied Intelligence Core Technologies
The forthcoming export control regulations will specifically restrict foreign capital from acquiring domestic embodied intelligence and AI Agent companies through hostile takeovers, and will place top AI R&D talent under strategic personnel management with post-departure confidentiality requirements.

Source: Toutiao / Reuters
Editor’s Take: China is treating embodied AI talent and IP as national strategic assets — a clear signal that this sector is considered as critical as semiconductors.

6. OpenRouter Data: Chinese Model Token Usage Hits ~60% Share
Data from AI model aggregation platform OpenRouter reveals that token consumption from Chinese AI models now accounts for approximately 60% of total volume on the platform, underscoring global developers’ reliance on cost-effective Chinese alternatives.

Source: NetEase
Editor’s Take: Market share tells the truth that policy debates try to obscure — the world’s developers are already voting with their API calls.

---

Track 3: Humanoid Robots（8 items）

1. Agibot (智元) Launches Hong Kong IPO Process — Target Valuation HK$40-50B
Agibot Innovation has officially initiated its Hong Kong IPO process, with cornerstone investors targeting a valuation of HK$40-50 billion ($5.1-6.4B). The company claims to be the world’s largest general-purpose AI robot company by revenue (2025 full-year + Q1 2026). Revenue grew from ¥300K in Year 1 to ¥60M in Year 2 to over ¥1B in 2025, with Q1 2026 alone exceeding ¥1B. Target: ¥10B revenue by 2027.

Source: Caijing Magazine / Sina Finance
Editor’s Take: Agibot’s “ecosystem model” — where investors and partners also become customers — is controversial but effective. The real test will be whether these robots do useful work beyond the initial purchase.

2. Unitree Debuts World’s First Mass-Produced Manned Mech “GD01” in Chengdu
Unitree Technology unveiled the GD01 — the world’s first mass-produced manned transforming mech — at Chengdu’s Chunxi Road. The 3-meter-tall steel mech can switch between bipedal and quadruped locomotion. Unitree also signed a strategic cooperation agreement with Chengdu municipal government.

Source: The Paper / Red Star News
Editor’s Take: A 3-meter mech on a shopping street is the most effective marketing a robotics company could ask for. Unitree understands: robots need to be seen before they can be used.

3. Chengdu Assembles “Top Five” of China’s Embodied Intelligence
With Unitree’s arrival, all five recognized leaders of China’s embodied intelligence sector — Unitree, UBTECH, Agibot, Galaxy General, and Star Era — have now established operations in Chengdu. The city is positioning itself as China’s robotics capital, with 60+ real-world training scenarios and a target of 10,000 deployed robots by year-end.

Source: The Paper / Red Star News
Editor’s Take: Chengdu is building what no other city has: a complete embodied intelligence ecosystem, from chip to brain to body to deployment. The flywheel is starting to spin.

4. Galaxy General (银河通用) Signs Full-Scene Strategic Cooperation with Hangzhou Qiantang
Galaxy General, valued at over ¥20B, signed a comprehensive partnership with Hangzhou’s Qiantang District covering retail deployment, industrial training, data openness, and a regional HQ. Its Galbot S1 is already operating 24/7 at CATL production lines — the world’s first humanoid robot routinely deployed in new energy manufacturing.

Source: Phoenix News Zhejiang
Editor’s Take: Galaxy General’s approach — “brain + cerebellum + motor control” end-to-end model — is the most technically ambitious in China. The CATL deployment proves it can work in real factories.

5. China’s Humanoid Robots Enter Real Factory Floors
Xinhua reports that humanoid robots are now being tested in Chinese garment factories (Aitu’s robots achieve 97% fabric-separating success rate), electronics assembly (PIABOT’s G2 completed 2,283 tasks in 8 hours without error), and automotive production (12.9-second cycle time vs. human 18 seconds). The Ministry of Industry and Information Technology aims for 10,000-unit deployment scale by year-end.

Source: Xinhua / People’s Daily
Editor’s Take: The 18-month payback period for garment robots is the magic number — once robots cross the economic viability threshold, adoption accelerates exponentially.

6. LimX Dynamics (逐际动力) Completes $200M Pre-IPO Round at ¥15B Valuation
LimX Dynamics has closed a $200M Pre-IPO financing round, reaching a post-money valuation of ¥15 billion ($2.1B), further consolidating the capital race in China’s humanoid robot sector.

Source: 36Kr
Editor’s Take: Another unicorn minted in China’s robotics boom. The Pre-IPO label signals an imminent public listing — the sector’s capitalization wave is far from over.

7. China’s Quadruped Robots Command ~70% of Global Sales
According to China’s Ministry of Industry and Information Technology, domestically developed quadruped robots now account for nearly 70% of global sales. China has over 400 humanoid robot products — more than half the global total.

Source: People’s Daily / Daily Economic News
Editor’s Take: 70% global market share in quadrupeds — China dominates this category the way it dominates solar panels and EV batteries. Scale and cost advantages are nearly insurmountable.

8. Future No Far Robot Raises Nearly ¥1B in Pre-A Round
Future No Far Robot (未来不远机器人) announced completion of a Pre-A financing round totaling nearly ¥1 billion, backed by internet giants, industry leaders, consumer brands, and well-known funds.

Source: GMT Eight
Editor’s Take: Nearly ¥1B in a Pre-A round signals extraordinary investor conviction — this is one of the largest early-stage raises in China’s robotics history.

---

Track 4: Autonomous Driving（6 items）

1. WeRide Wins Fourth Consecutive China Intelligent Driving Championship
WeRide’s (文远知行) end-to-end solution WRD 3.0, co-developed with Bosch and deployed in Chery Exeed Sterra ET, won the second China Intelligent Driving Championship in Wuhu by a margin of over 10 points — marking the first and only “four-peat” in the competition’s history.

Source: Sina Auto
Editor’s Take: Four consecutive wins with an end-to-end architecture — WeRide is proving that its approach works in complex urban scenarios, not just controlled environments.

2. Baidu’s Apollo Go Expands into Central Asia
Baidu’s autonomous ride-hailing service Apollo Go (萝卜快跑) has entered the Central Asian market, marking a significant geographic expansion beyond China and the Middle East.

Source: 36Kr / Zhou Bao
Editor’s Take: Central Asia is an interesting choice — less regulatory friction than Western markets, and a genuine need for mobility solutions in underserved cities.

3. Pony.ai Partners with JD Auto for Robotaxi Maintenance
Pony.ai (小马智行) has partnered with JD.com’s automotive service arm JD Auto (京东养车) to establish dedicated maintenance infrastructure for its autonomous vehicle fleet — essentially creating “4S shops” for robotaxis.

Source: Comprehensive reports
Editor’s Take: The unglamorous side of autonomous driving: someone has to fix the cars. Pony.ai’s partnership with JD Auto is a smart move to build operational infrastructure before competitors.

4. Shenzhen Breaks New Ground with Autonomous Driving Legislation
Shenzhen has enacted pioneering legislation for autonomous driving, establishing a legal framework for liability, insurance, and operational requirements — setting a template for other Chinese cities.

Source: Comprehensive reports
Editor’s Take: Legislation is the unsexy enabler that unlocks everything else. Shenzhen’s move will accelerate deployment across the Pearl River Delta.

5. 9Shi Intelligence (九识智能) Scales L4 Autonomous Driving Without HD Maps
At WAIC 2026, 9Shi Intelligence showcased its L4 autonomous driving solution that operates without high-definition maps, achieving mass-production scale. The company positions itself as a dual-sided smart logistics platform.

Source: WAIC 2026 reports
Editor’s Take: “No HD maps” is the holy grail of autonomous driving scalability — it means you can deploy anywhere, not just in pre-mapped cities.

6. Momenta’s Unmanned Freight Operations Running in Suzhou
Momenta’s unmanned freight business has reportedly been operating in Suzhou for several months, quietly building real-world operational data for autonomous logistics.

Source: Comprehensive reports
Editor’s Take: The quiet, unglamorous work of running unmanned freight in a real city — this is where the actual competitive advantage in autonomous logistics gets built.

---

Track 5: World Models & Physical AI（7 items）

1. ACE ROBOTICS Open-Sources Kairos 3.0-4B: Real-Time World Model
ACE ROBOTICS (商汤科技联合创始人王晓刚创立) released Kairos 3.0-4B, the industry’s first native world model for embodied intelligence with unified “multi-modal understanding-generation-prediction” architecture. Key specs: 72x faster inference than NVIDIA Cosmos 2.5, real-time edge generation on NVIDIA THOR platform, 7-minute coherent interaction predictions, cross-embodiment deployment (single-arm, dual-arm, dexterous hand) without retraining.

Source: The Times Australia / ACE ROBOTICS
Editor’s Take: 72x faster than Cosmos 2.5 with just 4B parameters — this proves that world models don’t need to be massive to be effective. The real-time edge deployment capability is the killer feature.

2. Chinese World Model Tops Li Fei-Fei’s Benchmark — Fully Adapted to Huawei Ascend
A domestic open-source world model with hundred-billion-scale parameters has topped the authoritative benchmark initiated by Stanford’s Li Fei-Fei team. Crucially, the model has completed full-pipeline adaptation to Huawei’s Ascend chips — from training to inference, entirely on domestic hardware.

Source: Toutiao
Editor’s Take: Topping a benchmark is impressive; doing it entirely on domestic chips is a strategic statement. This is what “independent and controllable” looks like in practice.

3. Xiaomi Robotics-1: First to Validate Scaling Laws with Real-World Robot Data
Xiaomi’s Robotics-1 foundation model is trained on 100,000+ hours of real-world data, making it the first to validate scaling laws in robotics using actual (not simulated) data. The model tops multiple global benchmarks and offers “out-of-the-box” deployment capability.

Source: Multiple Chinese tech media
Editor’s Take: Scaling laws worked for language models — Xiaomi is betting they’ll work for robots too. 100K hours of real data is a significant moat.

4. China’s Embodied AI Export Controls Target Core Technologies
The forthcoming export control regulations explicitly include embodied intelligence core technologies, restricting foreign acquisition of domestic companies and placing key R&D personnel under strategic management — signaling that China views physical AI as a national security priority equal to semiconductors.

Source: Reuters / Toutiao
Editor’s Take: When a technology gets export-controlled, you know the government considers it strategically vital. Physical AI has officially joined the “crown jewels” list.

5. Nvidia-SK Group $500B AI Infrastructure Partnership Includes Physical AI
The $500B partnership between Nvidia and SK Group explicitly covers “physical AI” infrastructure, with SK Hynix’s HBM4 memory designed to power next-generation AI systems for robotics and autonomous systems.

Source: Nvidia official / Korean media
Editor’s Take: The AI infrastructure buildout isn’t just about LLMs — physical AI is explicitly named as a target use case, signaling the industry’s confidence in robotics’ compute demands.

6. Agibot’s “Kuotuo” Dual-Engine System Tackles Quadruped Robot Inspection
At WAIC 2026, Agibot’s Kuotuo division unveiled a dual intelligent engine system specifically designed to solve three core pain points in quadruped robot inspection: continuous localization in complex environments, multi-sensor fusion, and long-duration autonomous operation.

Source: WAIC 2026 / Fun Tech
Editor’s Take: Industrial inspection is where quadruped robots generate real revenue today — Agibot’s focus on solving actual customer pain points (not just demos) is the right commercial instinct.

7. Kepler Releases World’s First Hybrid-Architecture Quadruped Robot “Qilin” Series
Kepler Robotics unveiled the Qilin series — the world’s first quadruped robot built on a hybrid architecture, combining electric and hydraulic actuation for improved power and endurance in industrial applications.

Source: Global TMT
Editor’s Take: Hybrid actuation is a pragmatic engineering choice — pure electric lacks power, pure hydraulic lacks precision. Kepler’s approach may find the sweet spot for industrial use cases.

---

Today’s Track Summary

---

East Voice · Physix Frontier · 2026.07.26