观察海外 AI 产业的发展路径，我们发现从大厂密集投入 Capex（2023 年）到 Token 的爆发 （2025 年）并非一蹴而就，中间存在两年左右的滞后期。海外的时间线看，2023→2024→2025 分别经历了 Capex 高投入→模型密集迭代+云商收入回归高增→Token 高增的三个阶段。这背后的逻辑是因为 Capex 的高投入会先用于模型的训练中，大模型的训练十分昂贵对资金要求 较高。一直到 2025 年尤其是下半年开始，需求侧的核心指标 Token 数才开始高速增长。我们认为这背后的原因是随着持续两年的投入，许多场景如搜广推、Coding、多模态等逐步跑通。

科技VS历史：动态PE：不及2021年，远不及2000年科网；ERP：甚至不及1998牛市中段 科技VS其他板块：标普/DAX、美股成长/价值已回落至中位水平，没有过高科技龙头VS科技整体：M7/科技整体，估值比已至中低位 结合基本面看：PB-ROE视角，科技泡沫显著低于科网水平AH科技VS美股科技：A股科技已修复，港股科技仍有性价比

10 月 31 日，阿里官宣：淘宝闪购支持医保支付购药的城市已经超 40 个、覆盖近 300 个区县，近 1 万家定点零售药店 5 万多个医保品，累计服务用户已超过 800 万人次。淘宝闪购已提供常态化的 24 小时全天候送药上门服务，在使用医保支付后，可以大大降低市民买药的个人现金支出，很多常用药到手仅需自付包装费和物流费。

The global AI-powered receipt automation market is a part of the global application software market. The global application software market was valued at $349.96 billion in 2021 and reached $439.86 billion in 2023. The global application software market witnessed an incremental growth of $89.9 billion and registered a CAGR of 12.1% from 2021 to 2023.

The global AI-powered radio access network market encompasses the integration of artificial intelligence and machine learning technologies within the cellular network infrastructure that connects end user devices to the core network. This technology functions by employing sophisticated algorithms to continuously analyze real time network data, enabling the autonomous optimization of RAN operations. Key applications include predictive analytics for network maintenance, intelligent traffic steering, dynamic energy consumption management, and automated resource allocation for 5G network slicing.

AI-driven customer support agents encompass advanced software solutions leveraging artificial intelligence and machine learning to automate and enhance customer interactions. These intelligent agents, often deployed as chatbots or virtual assistants, are designed to understand, process, and respond to customer inquiries across various communication channels, including websites, mobile applications, and social media. Their operational mechanics involve natural language processing for understanding user input, natural language generation for crafting human-like responses, and machine learning algorithms for continuous improvement through interaction analysis

AI voice recognition encompasses the technologies, software, and hardware systems engineered to identify, interpret, and process human speech. This technology fundamentally relies on artificial intelligence, particularly machine learning models and deep neural networks, to perform automatic speech recognition (ASR). The core function of ASR systems is to convert spoken language into a machine-readable format, such as text or executable commands. More advanced solutions integrate natural language understanding (NLU) to discern user intent and context, enabling sophisticated human-computer interactions.

The global agentic AI in law enforcement and surveillance market encompasses autonomous systems capable of independent decision making and action without direct human intervention. These technologies are engineered to analyze complex data sets, learn from behavioral patterns, and execute tasks in real time to enhance security and operational efficiency.

Agentic AI for data engineering represents a transformative paradigm wherein autonomous AI agents are designed to perform complex data-related tasks with minimal human intervention. These agents operate by perceiving their environment, formulating goals, planning actions, and executing them to achieve desired outcomes in data pipelines. This includes automated data collection, cleansing, transformation, integration, and validation. Technology leverages advanced machine learning models, natural language processing, and reinforced learning to understand data structures, identify anomalies, and optimize data flows.

The global AI servers in financial services market encompasses the specialized high performance computing infrastructure, including hardware, software, and integrated solutions, designed to execute artificial intelligence and machine learning workloads within the financial industry. These servers are characterized by advanced processing units, such as GPUs and other AI accelerators, high bandwidth memory, and low latency networking capabilities. They are engineered to manage and process the immense volumes of data required for training sophisticated algorithms and deploying AI models at scale.