国君研究|全球AI应用趋势 · 合集

AI ASIC Chips - ASIC chips are designed for specific scenarios and have a complete software and hardware ecosystem, offering significant price and power consumption advantages compared to GPUs [1][2] - Although single ASIC chip computing power is still lower than the most advanced GPUs, the computing efficiency of ASIC clusters may be higher, making them suitable for AI inference and training [1][2] - The ASIC market is in its early stages, with a low market share in AI acceleration chips (16% in 2023), but it is expected to grow rapidly, reaching over $40 billion by 2028, with a CAGR of 45% [1][3] - ASIC chips like Google TPU v6 and Microsoft Maia 100 achieve 90% and 80% of H100's non-sparse computing power, respectively, while being more cost-effective in inference scenarios [2] - ASIC clusters, such as Google TPU v5p, can scale up to 8,960 chips per pod and maintain linear acceleration, potentially surpassing GPU clusters in computing efficiency [2] - The software ecosystem for ASIC chips is gradually maturing, with cloud providers developing full-stack software ecosystems and open-source platforms like ROCm and oneAPI [2][3] AI Human-Computer Interaction - LLM-driven multimodal interaction methods are leading a new productivity revolution, with AI agents potentially reshaping the ecosystem of end-side operating systems and business models [4][5] - Apple, Google, and Microsoft are expected to benefit from AI-driven human-computer interaction changes, with Apple Intelligence, Google's Project Astra, and Microsoft Copilot leading the way [4][5] - AI agents will integrate and unify various app entrances, becoming "super apps" at the operating system level, with Siri, Project Astra, and Copilot+PC leading the transformation [4][5] - The evolution of human-computer interaction will bring new business models, such as Apple's potential to charge for Siri calls and Google's ability to enhance its app ecosystem through AI [5] AI Social Networks - AI is transforming traditional software business models, with social networks expected to strengthen their value through hybrid AI structures, enhancing network effects and social utility [7][8] - Social networks are a highly certain sector in the AI era, with demand and business model certainty driven by multi-level social applications and improved matching efficiency [7] - The future of AI social networks will involve interactions between humans, robots, and AI agents, creating a hybrid network that enhances human prosperity [7] - Global social network leaders like Tencent, Meta, and Google are recommended due to their user base, ecosystem, and model capabilities, with AI agents potentially doubling network nodes and tripling network effects [8] AI Smartphones - AI smartphones are expected to become the core entry point for AI applications, surpassing AI PCs due to their ubiquity and portability, with future AI phones evolving into autonomous agents with digital personalities [10][11] - The AI smartphone market is projected to drive a new wave of device upgrades, with shipments reaching 380 million units by 2025, a 134% year-on-year increase [10] - AI smartphones will benefit hardware and model providers, with companies like Qualcomm and TSMC expected to see improved profit margins and increased demand for advanced SoCs [10][11] - The deployment of large models on devices is technically feasible, with advancements in memory management and NPU performance enabling high-performance computing on smartphones [10]