How AI Agents Are Redefining the Future of Intelligent Computing

Overview of AI Agents

AI agents trace back to the 1950s Turing test and are defined as intelligent entities that can sense their environment, make decisions, and act. Unlike traditional AI, they can think autonomously, call external tools, and achieve goals by generating their own prompts.

Technology Stack Layers

The AI‑agent market can be divided into three layers:

Infrastructure layer : large models, compute infrastructure, and data infrastructure that support AI agents.

Platform layer : AI‑agent development‑management platforms and LLMOps tools that bridge infrastructure and applications.

Application layer : industry‑specific agents (finance, energy, automotive) and generic scenarios such as process automation, data analysis, and collaborative office tools.

Market Size Estimation

Global data volume reached 103 ZB in 2022 (China 23.9 ZB). It is projected to grow to 284.3 ZB by 2027 (CAGR 22 %). China’s data volume is expected to reach 76.6 ZB in the same period (CAGR 26 %), outpacing global growth. Since July 2023, China has released about 300 domestic large models covering finance, law, education, healthcare, entertainment and other verticals.

Evolution from Training to Inference

With rapid upgrades of large models, AI agents have shifted from pure training to inference‑driven reasoning. OpenAI’s GPT‑o1 exemplifies an inference‑heavy design, where operational Capex can reach millions to billions of dollars, indicating a future surge in inference compute demand.

Interaction Modes and Core Modules

Three interaction modes are identified:

Embedding mode : AI fills missing information and performs small tasks under user prompts.

Copilot mode : AI follows user‑defined workflows, collaborating with humans (e.g., Microsoft Copilot suite).

Agent mode : AI takes primary responsibility; humans set goals and supervise.

The four essential modules of an LLM‑based agent are:

Memory : short‑term context storage and long‑term vector‑store memory for observations and actions.

Planning : decomposes goals into sub‑tasks using chain‑of‑thought or tree‑of‑thought reasoning.

Tool Use : calls external APIs (search, code interpreter, math engine, databases, knowledge bases) to obtain missing information.

Action : executes the planned steps in the environment.

Industry Chain Analysis

The AI‑agent ecosystem consists of upstream infrastructure and technology providers, mid‑stream R&D and integration firms, and downstream application vendors and end users.

Upstream : smart compute centers, large‑model developers, data‑labeling and optimization services.

Midstream : companies that combine LLMs with memory, planning, tool‑use, and action modules to build domain‑specific agents.

Downstream : providers of AI‑agent‑enabled products such as intelligent customer service, personal assistants, autonomous driving, software development tools, and financial‑management solutions.

Future Outlook

GPT‑style agents are early products; as tool‑calling and autonomous reasoning improve, agents will become more self‑directed, driving a shift toward inference‑centric compute and reshaping multiple industries.