AI Agent Industry Chain Panorama: Investment Opportunities from Compute Chips to Vertical Applications

First Layer: Compute Infrastructure – The Most Certain “Shovel Sellers”

This tier is the short‑term most certain segment: more agent workloads mean higher demand for chips. Zhongshang Industry Research Institute reports China’s AI accelerator chip market grew from ¥30.1 bn in 2021 to ¥142.5 bn in 2024 and is projected to reach ¥381.39 bn by 2026. Domestic share is expected to rise from ~40% in 2025 to >50% in 2026, with Huawei Ascend leading at 812,000 units shipped, followed by Cambricon, HaiGuang, Loongson, etc. A note adds that agent‑era compute needs extend beyond GPU inference to tool calls, environment setup, task scheduling, long‑term memory, and high‑concurrency CPU coordination, pulling optical modules, storage, power, liquid‑cooling, and PCBs upstream. Investment view: the layer delivers proven performance and strong growth, but is sensitive to domestic substitution progress and external regulations; the safest bets are firms tightly bound to leading cloud suppliers and capable of scaling multi‑GPU clusters.

Second Layer: Cloud Computing Platforms – From Selling Compute to Capturing the Entry Point

Cloud providers have shifted from pure GPU‑hour rentals to becoming the deployment gateway for agents, controlling model distribution and scaling negotiations. The Chinese cloud market is projected to expand from ¥828.8 bn in 2024 to ¥1,398.6 bn by 2026, with Alibaba Cloud, Huawei Cloud, Tencent Cloud, and China Telecom Cloud leading, while players like UCloud, Kingsoft Cloud, and Wangsu focus on one‑click deployment, model hosting, API gateways, and edge security.

The investment logic is lighter than hardware: lower capital intensity, clear business model, and faster cash‑flow improvement. Short‑term capital prefers cloud because of rapid rollout and stable cash, but warns that providers staying only in “shovel selling” may be squeezed by upstream chip and downstream platform dynamics; true moat lies in binding agent runtimes and model distribution into their ecosystems.

Third Layer: Large Models – From Price Wars to Tiered Pricing

Large models act as the “brain” of agents. Early years saw aggressive price competition; by 2026 a signal emerged when Zhipu raised model prices, indicating a shift toward tiered subscription and higher pricing power. The driver is agents’ demand for unit‑cost efficiency, long context, multi‑step reasoning, and coding capability, which favors cost‑effective domestic models.

In February 2026, China’s weekly model call volume surpassed the United States for the first time. Domestic models such as MiniMax (long context and programming) and Zhipu (leading coding tech) have captured significant share in open‑source ecosystems like OpenClaw. Investment opportunities focus on vertical‑specialized models (finance, healthcare, industry) and system‑level models that combine custom chips with ecosystem reach, rather than another generic large model.

Fourth Layer: Agent Platforms and Frameworks – The Largest “Operating‑System‑Level” Opportunity

If chips are the foundation, cloud the utilities, and models the brain, then agent runtimes and development frameworks are the “body” that gives the brain hands. OpenClaw’s rapid rise is likened to an “Android moment” for agents: it provides a runtime with a “skill + plugin” architecture that lets AI operate computers, invoke tools, and execute tasks. GitHub stars have surged to the 300 k range and the plugin marketplace ClawHub is live, echoing the early Android ecosystem explosion.

Players split into open‑source camps (OpenClaw, LangChain, AutoGPT, Dify) and giant‑vendor camps (Microsoft Copilot + Agent, Salesforce Agentforce, ByteDance, Tencent WorkBuddy, Coze). The layer is touted as the next OS‑level opportunity, but its moat is unclear: framework iteration is fast, developer loyalty low, and monetization paths unproven. Success may hinge on underlying protocols such as MCP (tool connection) and A2A (agent‑to‑agent collaboration); whoever defines these standards could control the “OS ticket.” Investment bets on platform‑scale upside but must tolerate higher uncertainty.

Fifth Layer: Application Layer – Vertical SaaS’s Shortcut to Market Dominance

The top tier is closest to revenue and most visible to end users, offering the greatest chance for Chinese firms to “leapfrog.” IDC projects China’s AI agent market to reach ¥182.34 bn in 2025 (78% YoY) and enterprise agents to exceed 350 million by 2031 (CAGR > 135%). Industrial penetration is already 47%. The 2026 government work report mentions “intelligent agents” for the first time, with the State Council’s “AI+” action targeting >70% adoption by 2027 and >90% by 2030.

Success here depends on deep industry know‑how rather than raw compute or model size. Examples: Zhongkong Technology’s TPT time‑series model has deployed in over 110 petrochemical projects, generating >¥1.7 bn economic benefit; TuorSi’s focus on government‑finance compliance has delivered 40+ projects. Established SaaS vendors (Kingsoft Office, Yonyou, Glodon) can embed agents into core workflows more easily than pure startups because they already solve complex deployment challenges. Investment view treats this tier as the “third ladder” – driven by expectations with the largest upside. High‑impact use cases concentrate in domains with standardized data, high labor cost, and clear ROI (legal, finance, customer service, marketing). Business models are evolving from pure subscription to result‑based revenue‑as‑a‑service (RaaS), where providers share a portion of the cost savings they generate.

Investment Ladder Across the Five Layers

The five tiers form a transmission chain: booming applications → token inflation → compute re‑valuation → cloud‑model price restructuring → platform‑application resurgence. Institutions segment investments into three ladders:

First ladder (high certainty, proven performance): Compute infrastructure – optics, PCBs, servers, liquid cooling.

Second ladder (high elasticity): Domestic compute chain – chips, devices, advanced packaging, tied to domestic substitution progress.

Third ladder (largest expectation gap): Edge AI and vertical applications – industry SaaS, embodied intelligence, expected to materialize after 2026.

Certainty declines while imagination space expands down the ladder; there is no single optimal choice, only the fit for each investor’s risk‑return profile.

In analogy, large models are the “electricity” and AI agents the “appliances.” While power (compute, models) is essential, the real transformation for ordinary users comes from the appliances themselves – the platforms and vertical applications that make agents truly useful.