Driscoll's Responds to 'Detected Carcinogens'; Meituan Denies Launching 'Bachelor's Special Delivery' and 'Master's Premium Delivery' Services; US Calls on Samsung Electronics and SK Hynix to Share Excess Profits 8点1氪丨怡颗莓回应“检出致癌物”；美团否认上线“本科专送”“硕士优送”等配送服务；美方喊话三星电子、SK海力士分享超额利润

TL;DR

• Google is developing "Frozen v2," a custom server chip embedding Gemini model architecture to achieve 6-10x better token processing per watt compared to current TPUs.
• Moonshot AI reports that API calls constitute 70% of its B2B revenue, signaling a sustainable commercialization loop for enterprise AI agents.
• Alibaba released Qwen-Image-3.0, a third-generation image generation model supporting 4.5k token inputs for complex knowledge diagrams and UI design.
• Tencent launched Hyra-1.0, a recursive self-improving research agent capable of iterative strategy refinement in engineering and creative tasks.
• Beijing plans to build "Token Factories" and add 50,000 P of intelligent computing power in H2 2026 to support the token economy.

Why It Matters

This update highlights a critical industry shift from pure model capability to infrastructure efficiency and specialized hardware optimization, as demonstrated by Google's Frozen v2. It also underscores the maturation of AI commercialization, where API-driven B2B models and autonomous agent ecosystems are becoming primary revenue drivers for major tech firms. For practitioners, these developments signal the importance of integrating specialized hardware and building robust, self-improving agent workflows for enterprise applications.

Technical Details

• Google Frozen v2: A proprietary server chip designed to permanently embed parts of the Gemini model architecture directly into silicon. This reduces computational load and data transmission needs, aiming for 6-10x higher token throughput per unit of power compared to standard TPUs.
• Alibaba Qwen-Image-3.0: The latest iteration in the Qwen image series, featuring support for up to 4.5k tokens in input. It specializes in generating complex visual outputs like knowledge diagrams with formulas and geometric shapes, supporting 12 languages and 20+ font families natively.
• Tencent Hyra-1.0: An autonomous agent framework focused on recursive self-improvement. It utilizes self-play, self-evaluation, and user feedback loops to iteratively refine strategies in research, engineering, game design, and content creation.
• Moonshot AI KimiHosted Agent: A new platform offering standardized APIs for enterprise use cases such as PPT generation and investment research system construction. It provides sandbox and harness capabilities, allowing direct integration into internal corporate systems.
• Beijing Token Economy Infrastructure: Strategic initiative to construct "Token Factories" and distribution platforms. The goal is to create a full-stack autonomous technology system combining RISC-V chips, AIOS, and agent applications, targeting 130,000 P of total computing power by year-end.

Industry Insight

• Hardware-Software Co-Design: The emergence of chips like Google's Frozen v2 indicates that future competitive advantages will rely heavily on vertical integration between model architecture and physical hardware, optimizing for energy efficiency and latency rather than just parameter count.
• Agent-Centric Business Models: The success of Moonshot's API revenue and Tencent's self-improving agents suggests a pivot toward autonomous, task-oriented AI services. Companies should prioritize building stable, scalable agent infrastructures over simple chat interfaces for B2B value capture.
• Regulatory and Economic Shifts: Beijing's push for a "Token Economy" and specific computing power targets reflects state-level strategic planning to dominate the next phase of AI infrastructure. Global firms must monitor regional policy shifts regarding data sovereignty, compute allocation, and emerging economic frameworks around AI-generated content and tokens.