---

🧠 I. Large AI Models

Chinese AI Model Popularity Surpasses US Rivals; Open Source and Low Prices Break Global Ground

Industry statistics and community data show that domestic models' popularity among developers and users has exceeded some US competitors, with downloads and discussion heat rising simultaneously. Open-source routes, competitive API pricing, and optimization for Chinese scenarios are the main drivers. Some metrics have shifted from catching up to leading. Changes in community leaderboards often precede commercial data; shifts in model popularity usually reflect in revenue two to three quarters later. For domestic large model companies, the overseas developer market means new commercialization space. Rising model demand also drives inference compute consumption, directly benefiting domestic inference chips and cloud providers. Whether popularity converts to paid revenue is the next watershed. Currently, most vendors still use free quotas plus low-price packages to gain ecosystem share.

Domestic vendors' competitive focus is shifting from training capabilities to ecosystem and price. Developer community activity becomes the deciding factor for the next round of funding and orders. This trend is already reflected in open-source community toolchains and deployment solutions.

DeepMind Pilots World's First Double-Blind AI Evaluation

Models are tested in encrypted environments where neither evaluators nor evaluated parties know the questions beforehand. Questions are unreadable before testing starts, and the process is auditable, mechanically blocking leaderboard gaming. High scores on public question banks disconnecting from real capabilities is an industry chronic ailment. Vendors tuning parameters specifically for question banks are hard to guard against. Over the past year, multiple benchmarks lost credibility due to this. Double-blind design makes high scores hard to manipulate, bringing evaluation standard disputes to the table. For enterprises, more credible evaluations directly reduce model selection costs. For the open-source community, it means a new round of transparency competition. Top labs' follow-up speed is worth observing.

If this approach is adopted by mainstream labs, existing public leaderboard rankings may undergo a reshuffle, and the business model of the evaluation industry itself will need redesigning.

• Equity Map of China's Top AI Companies: A panoramic overview of money, people, and ownership, serving as a foundation for industrial research. Primary market investors wanting to understand the domestic large model landscape often find shareholder structures and capital backgrounds closer to truth than valuation numbers. This compilation clarifies the capital relationships of over a dozen top companies at once.
• Cocomelon Producer Asks Artists to Start Trying AI: Moonbug, the head studio behind children's entertainment, changes attitude; AIGC penetration in the content industry deepens. The company owns major IPs like Little Baby Bum and Blippi, with huge annual production volume. Once AI tools enter the production pipeline, impact on the animation industry spreads quickly. However, resistance from original artists hasn't disappeared; debates on tool ethics will continue.

💻 II. AI Software

Baidu Dazi Releases Comprehensive Upgrades, Targeting Professional Office Work with "Delivery That Amazes"

Personal, enterprise, professional suite, and workspace capabilities fully upgraded, covering topic planning, research, content creation, financial analysis, etc., directly producing publishable graphics, scripts, and reports. This signifies domestic AI office moving from draft generation to finished product delivery. The entire workflow from topic to reporting is packaged into one Agent, raising the competitive barrier for office software markets. For self-media and finance practitioners, this compresses repetitive labor across multiple stages into one subscription service. Tiered pricing for personal vs. professional suites clarifies the monetization path.

The strategy of targeting self-media first with industry suites corresponds to this client group's rigid demand for finished product delivery. Enterprise versions carry higher average ticket price expectations.

• Google Flow Launches: An AI creative studio for video, image, and custom tools, further lowering creation barriers. Unlike single drawing tools, it puts video, images, and custom tools in one workspace, targeting the complete creative workflow. Competition in generative video tools is fierce. By packaging multimodal creation capabilities now, Google clearly aims to occupy the creative entry point, intensifying direct competition with Adobe.

The key to competition for such tools has shifted from model capability to workflow integration and distribution channels. Ecosystems that start running first gain obvious advantages.

• Meta Hatch Personal Agent Exposed: Internal memos show it can order food, book tables, and access Instagram. Binding between agents and proprietary social products deepens, differentiating from office-scenario Copilot-like products. Direct competition with similar life agents is on the table. Personal data permissions will be the regulatory focus.
• Meta Abandons AI-Centric Layoff Restructuring Plan: Internal assessments found effects below expectations; adjustments involving thousands are shelved. Management expectations for using AI to reshape workforce structure are becoming cautious. This serves as a reminder for enterprise software firms valued on efficiency narratives: actual progress of automation replacing labor is slower than advertised.
• Workday Earnings Lackluster: Recent rally in software stocks stalls; market expectations for AI monetization begin to retreat. Salesforce surged ~23% the previous day on better-than-expected earnings. Workday's lackluster performance cools the broad rally in software sectors. The contrast between the two companies perfectly illustrates the two sides of enterprise software prosperity.

The market will reassess AI pricing power in the software sector accordingly. AI product penetration data in upcoming quarters will become new bargaining points.

• Bilibili Q2 Revenue 7.94 Billion Yuan: Gross margin improved quarter-over-quarter for 16 consecutive quarters, reaching 37.2%. Adjusted net profit 700 million yuan; daily active users 117 million; user scale continues to grow. Sustained gross margin improvement indicates better cost structure and income quality. Content ecosystem retention remains stable.
• Post-2000 Cross-Border Seller Self-Develops AI Magnifier: Photo recognition with voice explanation, targeting children's cognitive scenarios. Small team uses existing visual models to meet overseas niche demands. Hardware cost is low, margin is high. For young children, voice interaction is much friendlier than text input. Putting visual models into low-unit-price hardware is becoming a new experimental field for small teams going global.

🤖 III. Humanoid Robots

Anthropic Tests New Mechanism Internally; Claude Begins Controlling Robots and Lab Equipment

Robotic arms and scientific equipment start accepting Claude's instructions, completing operations like pipetting and sample preparation. Large model capability boundaries expand from dialogue and code to the physical world. Agents can observe, decide, and execute autonomously for the first time—a key step for autonomous intelligence moving from screen to reality. For lab automation and robotics companies, integrating large model brains could significantly lower automation barriers. The lab automation market, long dominated by traditional equipment manufacturers, faces uncertainty. If this mechanism works, biopharma, materials science, and other experiment-intensive industries will be the first beneficiaries.

Amidst large model vendors collectively betting on embodied intelligence, Anthropic chooses to enter via research labs, forming differentiation from consumer robot routes.

• Dreame Robot Vacuums Grow Rapidly in Emerging Markets: Growth rate exceeds 100% for two consecutive years; market share ranks first in Brazil, UAE, Mexico, and other countries. Business covers key nations like Turkey, Saudi Arabia, India, Brazil, Chile. Unlike peers betting on Western main battlefields, Dreame chose differentiated competition in emerging markets. High-growth regional revenue is becoming its second growth curve.

Behind high regional growth is sustained investment in localized operations teams, building channel deployment and after-sales networks according to local habits.

🚗 IV. Autonomous Driving

• MIIT and Three Other Departments Deploy Automotive Quality Improvement Actions: Establish Chief Quality Officer systems, strictly controlling technical review gates for new product admission. Notice also requires advancing revision of standards related to innovative design, reliability, and durability. Special inspections on quality and consistency will increase testing and compliance costs for automakers before new product launches. The automotive industry's competitive focus is shifting from piling up configurations to quality consistency. The more intelligent the vehicle, the greater the test for software reliability. This is a clear signal of stricter admission regulation for smart cars.

🌐 V. Physical AI

Google Releases Planetary Prediction Engine, Automating Global Earth System Models with AI

The engine automatically builds and runs models covering weather, climate, and resource simulation based on Earth AI. Prediction time compresses from days to minutes, making it easier to integrate real-time satellite observation data. Unlike previous earth system models relying on manual parameter tuning, the engine can automatically generate and calibrate models, greatly reducing maintenance costs. This is another landing point for AI for Science, signaling big tech entering infrastructure in meteorology and climate. Climate research institutions have long relied on supercomputers to run physical models. If AI engines can replace part of the compute demand, research costs drop significantly. Commercialization space concentrates in weather services and disaster warnings.

The model can continuously learn new observation data, reducing manual intervention compared to traditional numerical models. This is its biggest difference at the operational level.

• In-Depth Review of AI Chip Status: Architecture debate shifts to token efficiency per watt; compute competition enters refinement stage. Industry argument focus shifts from peak compute to actual runtime efficiency. Tokens per watt become core metrics for customer orders. Rising proportion of inference demand also changes chip design direction. Procurement decisions shifting from training to inference are restructuring capex.
• Nvidia Releases NVLink Fusion: Integrates ultra-high bandwidth memory for next-gen AI infrastructure. Binding high-bandwidth memory with interconnect protocols targets the bottleneck of insufficient VRAM bandwidth during large model inference. Making memory and interconnect an integrated solution also helps Nvidia consolidate its voice in AI server whole-machine solutions, laying technical groundwork for AI factories integrating training and inference.
• Half-Year Report of "First Stock in AI Drug Delivery": MedPac Tech revenue increased 134 times YoY; commercialization inflection point emerges. First interim report after listing tests commercialization substance. Beyond revenue growth, the market cares more about order sustainability. At the industry inflection point where AI drug discovery moves from molecule finding to clinical validation, this sample's reference value is significant.

📈 VI. Macro and Market Data

Nvidia Earnings Ignite AI Rally; Nasdaq Posts Best Single-Day Performance Since August

Nvidia's Q2 revenue grew more than double; data center business remains the largest engine. FY2028 revenue guidance is 70%, implying ~$67.3 billion, far exceeding analysts' expected 44%. After earnings, Nvidia's market cap rose $442 billion in a single day, closing up 8.74%, the largest single-day gain in 16 months. Index-wise, Nasdaq closed up 1.57%, S&P 500 up 0.72%, Dow up 0.20%. This was the best day for tech stocks since August 4. Semiconductor sector strengthened overall; Broadcom up 4.5%, Intel up 4.36%; storage and equipment stocks followed suit. Trading data shows sentiment clearly warming. A-share compute chain followed suit; Hygon Information up 6.50%, Foxconn Industrial Internet up 5.4