---

🧠 I. Large AI Models

Anthropic Q2 Revenue Hits $11.5 Billion, Becoming First Frontier AI Lab to Report a Profitable Quarter

According to Forbes, Anthropic disclosed Q2 2026 booked revenue of $11.5 billion, a year-over-year increase of over 14x (compared to $787 million in the same period last year). Adjusted operating profit turned positive, making it the first frontier AI lab to show investors a profitable quarter, with cumulative booked revenue for the first half of the year reaching approximately $16.2 billion. Industry trackers estimate that about 80% of its revenue comes from API and enterprise business, with Claude Code being the fastest-growing second engine; API gross margins have long remained above 80%. Losses primarily stem from step-wise training investments for next-generation models, while the revenue side has already outpaced expenses in a supply-constrained, positive-margin state. For the first time in over three years, the industry assumption that "labs can never burn through compute bills" has been overturned by numbers. Previously, discussions at the shareholder level had emerged regarding an October IPO with a $2 trillion valuation; this disclosure provides hard backing for price inquiries. Additionally, Bloomberg reports that its pre-IPO credit line will also be raised to over $10 billion, further thickening its IPO ammunition. In contrast, OpenAI's publicly promoted revenue of over $40 billion is an annualized run rate; the two metrics are not comparable. For OpenAI, another lab preparing for an IPO, buyers have now seen a profitable financial statement, and this reference point will accompany its entire pricing process.

• OpenAI Offers Zero Data Retention Option for Frontier Models — Enterprise clients can apply for model instances that do not retain data throughout training and inference. This moves a step further toward industries with strict compliance requirements like finance and healthcare, redefining the privacy threshold for enterprise-grade AI and shortening the decision-making chain for large client orders.
• AI Chip Startup Fractile Negotiating Valuation at $6.5 Billion — According to Bloomberg, the company previously reached a partnership with Anthropic. The market pricing of inference chips is catching up to training chips; dedicated inference hardware is becoming the new battlefield for reducing costs for large models, and giant orders are mass-producing new chip unicorns.

💻 II. AI Software

Stripe Plans to Acquire OpenRouter for Approx. $7.5 Billion; Payment Giant Directly Buys AI Model Distribution Entry Point

Stripe announced plans to acquire AI model aggregation platform OpenRouter on August 19 (US Eastern Time). According to the New York Times citing sources, the deal price is approximately $7.5 billion, with about $1.5 billion going to the founders. Less than three months ago, OpenRouter completed a $113 million funding round at a valuation of approx. $1.3 billion; the current price is nearly six times that valuation, making this deal a new pricing anchor for AI infrastructure M&A. OpenRouter is the primary entry point for developers calling open-weight models; developer traffic for domestic open-source models like DeepSeek and Z.ai is largely distributed through it, naturally controlling the routing, accounting, and billing of token flows. Stripe, valued at nearly $16 billion for payment infrastructure, aims to bring all AI inference pricing and settlement into its own ecosystem, as stated by CEO Patrick Collison. OpenRouter expressed hope to maintain its neutral positioning of "multi-model prosperity." For model labs, distribution channels are upgrading from developer tools to strategic assets, requiring reassessment of capabilities to reach developers. For the primary market, the valuation system at the model layer may be repriced accordingly. For regulators, whether a payment giant acquiring a model distributor triggers antitrust review will be a subsequent observation point. This deal simultaneously draws a line for all API relay and model gateway businesses: either merge into giants or be bought by them.

• Meta AI Launches Mac Desktop App — Chatbots enter the system desktop layer, allowing direct sharing of window content to participate in daily workflows. Competition for the desktop AI entry point gains another pole; Mac ecosystem AI assistants are no longer just Apple's own, and the matchup against the ChatGPT desktop version has begun.
• AI Accounting Startup Rillet Secures $100 Million Series C — Led by Iconiq, with a valuation reaching $1 billion, only two years after emerging from stealth mode. Vertical AI applications continue to attract capital in the primary market, with financial automation being bet upon as a new major category.
• Expedia Internal Memo Leaked: AI-Driven Restructuring Leads to Departure of 8 Executives — The online travel giant writes AI efficiency directly into organizational adjustments. Personnel structures of traditional internet companies are beginning to be rewritten by AI; cost reduction at the organizational level has just begun, and the industry is expected to follow suit successively.
• Alipay No Longer Forces App Redirection — When Agents like OPPO Xiaobu call payment services, they can do so without redirecting to the Alipay app. Super entry points begin loosening constraints for the Agent ecosystem; DAU is no longer the sole obsession, and service capability itself becomes the new moat. The entry logic of the payment industry is being rewritten by Agents.

🤖 III. Humanoid Robots

HyperDimension Power KAI Completes World's First Autonomous Table Tennis Match by Humanoid Robot, Debuts at WRC 2026

At the 2026 World Robot Conference opening on August 19, HyperDimension Power showcased the world's first complete table tennis match autonomously played by a humanoid robot and released the SMASH 2.0 high-dynamic humanoid table tennis robot. Themed "Human-Machine Symbiosis, Production-Demand Integration," this conference features 50 competition items, covering 29 competitive events and 21 scenario-based events. Table tennis imposes extremely stringent indicators for real-time response to high-speed balls, landing prediction, and wrist force control—an order of magnitude higher than demo-level interactions like handing flowers or shaking hands. On the same booth, Xiaomi's new generation humanoid robot interacted with audiences in real-time as a "florist," and Honor's self-developed robot announced participation in the second Humanoid Robot Games. Chinese manufacturers collectively showcased their commercialization pace at this WRC. Just the day before, Unitree Robotics listed on the STAR Market, surging 629% at opening. The massive attention and liquidity for the embodied intelligence sector had just been ignited; industrial heat and secondary market sentiment completed a connection this week.

• Octopus Dynamics Debuts at WRC with "Brain-Hand-Data" Technology System — Viewing data closed-loop as a new paradigm for embodied intelligence R&D, emphasizing the scarce value of real-scenario data for models. The route debate shifts from sensors and actuators to data organization methods.
• Dialogue with Luming Robot Founder Yu Chao: Embodied Intelligence's Second Half Shifts from Building Robots to Building Skills — His numbers are blunt: An 89% task success rate in the lab drops to only 12% on real production lines. Industry competition is shifting from hardware arms races to data and skill accumulation; bubble narratives supported by demo videos are beginning to be punctured by real deployment data.

🚗 IV. Autonomous Driving

Aion Ray7 Globally Firsts Huawei DriveONE Chip-Level Drive-Brake Fusion, Pure Electric Chassis Technology Advances Further

Aion's first model after refreshing its logo, the Ray7, debuted in Songshan Lake, Dongguan. Officially claimed to globally feature Huawei's DriveONE chip-level drive-brake fusion technology, it is positioned as a mid-to-large pure electric sedan targeting young urban users. Signing ceremonies were held on launch day, announcing deepened technical cooperation. Huawei Car BU's "chip-level" fusion solution lands on a vehicle product for the first time. Architectural integration of electric drive, chassis, and smart driving at the chip level is becoming a new focus for differentiated competition among Chinese automakers. It also means Huawei's packaged depth extends from software and computing power to the whole-vehicle electronic architecture, redrawing boundaries between OEMs and suppliers, and re-evaluating the balance between self-research and external procurement.

• GAC Trumpchi Xiangwang S9 Launches OTA 3.0 — Upgrades include Huawei Qiankun Smart Driving ADS V4.1 and HarmonySpace 5 cockpit, adding features like cruise parking, lateral protection, and dual-control parking ventilation. Existing owners continuously gain advanced smart driving capabilities; penetration of software payment and subscription models on vehicles continues to deepen.
• Seres Semi-Annual Report: Revenue 57.493 Billion Yuan, R&D Investment 7.007 Billion Yuan (+34.8%) — Wenjie cumulative deliveries in H1 grew 10.2% YoY. Wenjie M9 retained the sales champion title in the 500k+ segment for two consecutive months. The profitability model driven by premiumization and intelligentization continues to be validated, further solidifying the commercial loop of the Huawei smart driving camp.

🌐 V. Physical AI

Amazon Prime Air Drone Delivery Expands to Nearly 500 US Cities Within the Year

Amazon announced a significant expansion of its drone delivery service Prime Air, aiming to cover nearly 500 US cities by year-end, expanding by orders of magnitude from current pilot scales. Logistics unmanned operations are moving from single-point trials to scaled operations. Physical AI networks composed of drones, ground robots, and automated sorting are starting to deliver on the efficiency promise of the "last mile." Cost structures for low-altitude logistics will be repriced, posing new regulatory challenges for urban airspace management. Regulation and technology need to evolve at the same pace.

• Wired Visits Cambridge General Embodied Robot Team — Robots capable of learning on-site quickly master new tasks like clever toddlers. Generality is no longer a roadmap slogan but visible testing results at booths.
• Gravis Robotics Receives Increased SoftBank Investment — Construction robotics company gets top-tier institutional endorsement under the narrative of infrastructure automation. Application scenarios for physical AI extend from factories and roads to construction sites; boring scenarios actually become the first cash cows to land.

📈 VI. Macro & Market Data

Overnight US Stock AI Heavyweights Stabilize and Rebound, Tesla Leads Gains >4%; A-Share Computing Power Sold Off Consecutively

As of the US Eastern close on 8/19, following the previous day's near 5% plunge in the Philadelphia Semiconductor Index, US stock AI heavyweights generally stabilized: Tesla closed up 4.23% leading the pack, Amazon rose 2.46%, Palantir rose 2.13%, Microsoft, Meta, and Alphabet saw slight gains. Only Nvidia dipped slightly by 0.99%, narrowing significantly from the previous night but failing to recover prior losses, indicating divergence is still being digested. On the A-share side (close on 8/19), Unitree's listing drained massive funds, compounded by the overnight transmission of the Philly Semi crash. Cambricon fell 9.64%, Innolight fell 9.36%, Eoptolink fell 9.00%, Hygon Information fell 7.26%, Foxconn Industrial Internet fell 6.60%, Kingsoft Office fell 4.44%. Optical modules, computing power, and AI application sectors sold off collectively, with humanoid robots being the sole highlight across the board. Sentiment on both sides of the strait rarely diverged so starkly: one side saw massive subscriptions and surges for new robot stocks, while the other saw massive sell-offs of old-track computing power stocks. The direction of capital voting with its feet is clear at a glance.

Xiaomi Q2 Revenue Breaks 100 Billion Again: 108.9 Billion Yuan, Operating Profit Surges 104.6% QoQ

Xiaomi's single-quarter revenue in Q2...