---

🧠 I. Large AI Models

OpenAI Power Structure Reshaped: Greg Brockman's Role Expands, Governance Landscape Shifts Pre-IPO

According to The Verge, following lawsuits with Musk and multiple executive changes, OpenAI co-founder Greg Brockman's role has significantly expanded, becoming the core figure in the company's actual daily operations. Reports citing observations inside and outside OpenAI indicate that decision-making is clearly concentrating around Brockman, from model roadmaps to organizational personnel. Both internal and external parties feel that this is now Brockman-led OpenAI. Over the past year, OpenAI fought a highly publicized lawsuit with Musk and underwent multiple adjustments in safety and commercialization teams. Brockman gradually took over daily operations during this turmoil. For private companies, concentration of power in key figures directly impacts IPO narratives and governance discounts. Shorter decision chains and stronger founder control are aspects the market welcomes; however, insufficient checks and balances and increased dependency on key individuals may become deduction points during investor due diligence. OpenAI remains the world's most valuable private AI company. Its primary and secondary market pricing, potential IPO timing, and governance voice of shareholders like Microsoft will be reassessed as the power landscape settles. Meanwhile, OpenAI's commercialization continues to accelerate, with product rhythms and organizational adjustments entering IPO preparation simultaneously. The final form of the governance structure will directly impact exit returns for this group of shareholders.

Moonshot AI and Zhipu Model Capabilities Approaching OpenAI and Anthropic Flagships

Bloomberg reports indicate that partial models from Chinese leading firms Moonshot AI and Zhipu have approached the flagship levels of OpenAI and Anthropic. Both companies have recently made dense releases regarding long context, reasoning, and Agent capabilities: Moonshot AI's Kimi series continues to emphasize ultra-long context, while Zhipu maintains a dual-track approach of open-source and commercial models. On certain key benchmarks, both have entered the same generational interval as overseas flagships. Over the past two years, domestic models were generally considered one to two generations behind overseas frontiers. Now, as the gap narrows, the competitive narrative at the model layer changes directly. For domestic model companies, valuation systems and overseas pricing gain stronger support. For OpenAI and Anthropic, product pricing and API strategies in the Chinese market face greater reference pressure. The discussion of domestic substitution shifts from the application layer up to foundation models themselves, raising the pricing anchor for domestic models in the primary market. Overseas developer community usage of domestic models is also rising. The China-US price gap at the model layer is narrowing, and global competition in API pricing enters a substantive phase.

💻 II. AI Software

Ramp Launches Proprietary AI Model Router, Enterprise Software Vendors Compete for Inference Distribution Entry Points

Corporate expense management platform Ramp launched its proprietary AI model routing tool, Router, on August 20 (US Eastern Time). It provides unified routing, monitoring, and billing entry points for inference calls, allowing enterprises to automatically schedule tasks between different models based on cost, latency, and quality. Previously, Stripe announced the acquisition of model aggregation platform OpenRouter. The dense layout by payment and financial software vendors in the model distribution layer indicates that AI inference pricing and settlement are becoming a second growth curve for software companies. Ramp cuts in from internal corporate cost control, complementing OpenRouter's developer-side positioning. As the de facto cash register of the AI era, whoever secures the routing layer position first controls the accounting rights of inference traffic. The business model of software companies charging by AI usage gains another reference sample. Cost visibility is replacing functional differences as the primary consideration for enterprise AI procurement. Router can also form data synergy with Ramp's existing expense management core business, linking every large model expenditure back to business documents. Inference costs are being incorporated into routine corporate financial subjects.

• Slack Launches Collaborative Vibe Coding Channels — Teams can co-program with AI agents within conversation flows, replacing the pattern of switching back and forth between multiple tools. AI programming begins entering the team collaboration infrastructure layer. The focus of enterprise-grade code generation shifts from personal assistants to organizational collaboration. Collaboration data will become new fuel for the next round of programming model training.
• Meta Opens Vibe-Coding Game App 'Pocket' Nationwide in the US — Users can generate and share games using natural language, moving from testing to mass market. Meta lays the entry point for AI creation directly into its billion-user product matrix. A new consumer entry point for natural language-generated content emerges. Games, as a testing ground for AI-generated content, verify the speed of democratizing creative tools.

🤖 III. Humanoid Robots

Chenhunxian Tech Demonstrates 1×N+1 Architecture at WRC 2026: One Brain Connecting Multiple Body Forms

At the 2026 World Robot Conference, Chenhunxian Tech demonstrated its 1×N+1 embodied intelligence architecture, where one brain can connect to multiple body forms. Live demos showed robots autonomously orchestrating operational processes based on tasks and generalizing across different scenarios. The company stated that the core of the architecture is consolidating generative task orchestration into a unified brain layer, rather than training skills separately for each machine. Compared to the previous generation's route of debugging unit-by-unit and customizing scene-by-scene, one-brain-multiple-machines means skill training costs can be amortized across multiple bodies. Swapping a robotic arm or workstation on a production line no longer requires retraining the entire model. Embodied intelligence is moving from one-machine-one-strategy to one-brain-multiple-machines. Industry valuation logic may also shift from selling hardware to selling brain licenses. This is one of the most notable route divergences at this year's WRC exhibition, forming a sharp contrast with last year's scenes relying on remote controllers. The company also demonstrated real-time task switching for robots in manufacturing scenarios, stating such capabilities can directly integrate with existing production lines. Manufacturing clients have already entered small-batch verification stages within the year.

• Bloomberg: Chinese Humanoid Robots Shift from Exhibition Demos to Industrial Frontline Applications — Reports indicate Chinese manufacturers are pushing humanoid robots toward scalable industrial deployment scenarios. Return on investment replaces display effects as the screening criterion, forming a sharp contrast with remote-controlled performances at WRC a year ago. The industry discussion focus shifts from "can it move" to "how long until payback," marking the start of the commercial verification phase. Order and yield data will become new metrics for screening companies in the next stage.
• CAICT: Over 70 Embodied Intelligence Training Grounds Built and Operational Nationwide — Reports show training grounds cover more than half of provincial-level administrative regions, with numbers under construction and planned still increasing. Infrastructure-first has become an industry consensus. Training ground operations and data services are becoming a new middle layer in the robotics industry. Gaps in data collection capabilities will first manifest in model generalization levels.

🚗 IV. Autonomous Driving

Luxeed RX Debuts L3-Level Autonomous Driving Architecture, Pre-Orders Open

At the Stelato G9 and HarmonyOS Intelligent Mobility Alliance new product launch, the Luxeed RX coupe SUV opened pre-orders. Richard Yu revealed it features a future-oriented L3-level autonomous driving architecture design. Product head Guo Rui stated more details on the architecture would be released later. Launched the same day, the Stelato G9 broke 3,100 firm orders within one hour, starting at RMB 429,800, validating the order absorption capacity of HarmonyOS Intelligent Mobility Alliance's high-end models. Huawei extending L3 architecture to the coupe category covers the full spectrum from hardcore off-road to coupes. If regulations and data collection progress smoothly, Huawei-affiliated smart driving vehicle coverage and data scale will reach new heights, supporting order expectations for supply chain segments like LiDAR, smart driving chips, and domain controllers. L3 commercialization pace becomes the core variable for the smart driving sector in the second half. The Huawei Smart Selection camp has completed multi-model layout in the RMB 200,000 to 500,000 price range. L3 architecture is expected to become the common foundation for the next round of product differentiation.

📈 VI. Macro & Market Data

US Stocks Fall Collectively: Long-Term Treasury Yield Rebound Erases Buyback Benefits, Magnificent Seven Retreat Across the Board

At the close on August 20 (US Eastern Time), the Dow Jones fell 1.32% to 52,759.21 points, the Nasdaq fell 1% to 26,067.17 points, and the S&P 500 fell 0.87% to 7,641.16 points. The US Treasury doubled the cap on government bond buybacks to at least $4 billion the previous day, but long-term yields gained only one day of relief before rising again. The 10-year US Treasury yield rebounded to 4.704%, with financing cost concerns suppressing stocks anew. Walmart's comparable store sales missed expectations, plunging 9.2%, dragging the Dow down over 700 points. The Magnificent Seven closed lower across the board, with Amazon leading the decline at 2.16%. Nvidia remained relatively resilient, falling only 0.33%. Structural highlights appeared in the hardware sector: Storage and optical communications bucked the trend, with SK Hynix up over 4%, Micron up over 3%, and Lumentum up over 6%. AI hardware prosperity hedged against liquidity tightening, with funds picking and choosing within the compute chain. The market also focused on Fed FOMC meeting minutes. If long-term rates continue to rise, volatility in high-valuation tech stocks will be further amplified. Meanwhile, tensions between Iran and the US pushed oil prices higher, adding disturbance to inflation expectations. Stock-bond linkage makes this round of valuation repair fragile.

Hudson River and CoreWeave Reach Multi-Billion Dollar AI Cloud Agreement

Bloomberg reports that quantitative hedge fund Hudson River Trading signed a multi-billion dollar, multi-year AI computing agreement with AI cloud provider CoreWeave. CoreWeave had previously secured contracts with major clients like Microsoft and OpenAI. The addition of hedge funds expands its client structure from cloud providers and model companies to financial quant institutions. Quant strategies have rigid demand for low-latency compute and are willing to lock in fixed costs with 3-5 year contracts. These buyers ignore concepts and focus solely on transaction costs per unit of compute, representing the highest value tier in AI cloud orders. AI compute demand thus sees a third growth pole beyond model companies. Cloud provider order visibility improves. Demand transmission in the optical module and server chains is a key observation point. The binding model between AI cloud providers and hedge funds may be replicated by peers. For CoreWeave, an increased proportion of financially stable clients helps improve reliance on debt instruments in its financing structure. Such long-term agreement orders are becoming the hardest assets in AI cloud providers' balance sheets.

• Alibaba's T-Head Second-Gen Chip Expected to Tape Out in H2 This Year — Wu Yongming revealed in the earnings call that this generation chip possesses strong computing power and interconnect bandwidth, internally believed to be fully capable of replacing large-scale model training. The Zhenwu M890 supernode is advancing simultaneously. The same earnings report showed Alibaba Cloud revenue growth hitting a 22-quarter high, with AI-related product revenue growing triple digits for 12 consecutive quarters. Domestic training-grade compute supply accelerates. Cambricon and Hygon fell 3.77% and 5.16% respectively on the day, as capital began to re-evaluate...