Krux Star Evening Report: Intel to Lay Off Staff in Data Center Division; Nissan Recalls Over 160,000 Armada and Infiniti SUVs in the US; China to Build Over 3,000 Charging and Swapping Stations for Electric Heavy Trucks 氪星晚报｜英特尔将在数据中心部门裁员；日产在美召回超16万辆阿曼达及英菲尼迪SUV；我国将建设3000个以上电动重卡充换电站

TL;DR

• Moonshot AI reports that API calls constitute 70% of its B2B revenue, indicating a mature, sustainable commercialization loop for enterprise AI services.
• Alibaba releases Qwen-Image-3.0, a third-generation image generation model supporting 4.5k token input, precise 10px text rendering, and native rendering in 12 languages.
• Tencent introduces Hyra-1.0, an autonomous research agent capable of recursive self-improvement for performance-oriented engineering and scientific tasks.
• Microsoft expands partnership with AMD to deploy the Helios rack-level AI platform on Azure for next-generation model inference workloads starting in late 2026.
• Meta and NVIDIA collaborate on material discovery via CuspAI, leveraging AI to find new manufacturing materials for the semiconductor industry.

Why It Matters

This update highlights the critical shift from experimental AI models to robust, revenue-generating enterprise infrastructure, as evidenced by Moonshot’s API-centric business model and Alibaba’s focus on high-fidelity, multilingual image generation. For practitioners, the introduction of self-improving agents like Tencent's Hyra and the integration of specialized hardware like AMD's Helios into major cloud providers signals that autonomous workflows and optimized inference pipelines are becoming standard operational requirements rather than niche innovations.

Technical Details

• Qwen-Image-3.0: Supports up to 4.5k token inputs, enabling complex context understanding. Key technical features include precise rendering of small fonts (10px) and native support for 12 different languages, addressing common hallucination and text-integration issues in generative image models.
• Hyra-1.0: An autonomous agent framework designed for recursive self-improvement. It utilizes self-play, self-evaluation, and user feedback loops to iteratively refine strategies in open-ended scenarios such as game design, content creation, and scientific research.
• AMD Helios Platform: A rack-level AI computing solution scheduled for deployment on Microsoft Azure in late 2026. It is specifically engineered to handle heavy inference workloads for frontier models, offering a scalable alternative to traditional GPU clusters.
• CuspAI Collaboration: A joint initiative between Meta and NVIDIA utilizing AI-driven material science to discover new compounds for semiconductor manufacturing, aiming to accelerate the development of next-gen chip materials through automated discovery pipelines.

Industry Insight

• API-First Monetization: The dominance of API usage in B2B revenue (70%) suggests that successful AI companies must prioritize stable, low-latency, and cost-effective API infrastructure over standalone consumer applications. Enterprises should evaluate vendors based on their ability to integrate seamlessly into existing workflows via standardized interfaces.
• Autonomous Agents in R&D: The emergence of self-improving agents indicates a move toward "AI-for-AI" workflows. Companies in R&D-heavy sectors should begin piloting autonomous agents for iterative testing and optimization, as these tools can significantly reduce human-in-the-loop overhead for repetitive engineering tasks.
• Hardware Diversification: Microsoft's adoption of AMD's Helios alongside NVIDIA underscores the strategic importance of hardware diversification in cloud computing. Practitioners should monitor multi-vendor AI infrastructure trends to avoid vendor lock-in and leverage competitive pricing and performance improvements across different chip architectures.