China’s Agentic AI Boom: What the OpenClaw Surge Reveals

China Agentic AI OpenClaw adoption marks a structural shift in how artificial intelligence is deployed, commercialized, and scaled across the Chinese market. The rapid rise of OpenClaw reflects falling inference costs, intensifying cloud competition, and growing enterprise demand for autonomous, workflow‑driven AI systems. For foreign businesses, the trend carries direct implications for cloud partnerships, digital transformation strategies, and competitive positioning in China’s fast‑moving AI ecosystem.

China’s rapid adoption of OpenClaw, an open-source AI agent platform that went viral across the country in early 2026, signals a major shift in how artificial intelligence is deployed and commercialized in the Chinese market. While the initial surge in popularity was driven by developers, small businesses, and technology enthusiasts, the phenomenon quickly expanded into a broader ecosystem involving cloud providers, local governments, and major technology companies.

Unlike traditional chatbot-based generative AI tools, agentic AI systems such as OpenClaw can autonomously plan and execute complex multi-step tasks across digital tools, APIs, and enterprise software environments. This capability is accelerating the transition from experimental AI usage toward persistent, workflow-oriented automation across sectors ranging from e-commerce and marketing to finance and logistics.

China’s unique combination of low AI inference costs, intense cloud competition among major tech firms, and strong policy support for AI adoption has created an environment in which agentic AI can scale rapidly. The OpenClaw surge therefore represents more than a temporary technological trend. It illustrates the next phase of China’s AI development, where open-source tools, cloud infrastructure, and enterprise automation converge.

Find Business Support

Our experts help you navigate China's market entry regulations, from company setup to joint ventures and M&A.

For foreign companies evaluating technology partnerships, digital transformation strategies, or market entry into China’s technology ecosystem, understanding this shift is increasingly critical. The rise of agentic AI is reshaping China’s competitive landscape in cloud services, enterprise software, and AI-enabled productivity tools, while also introducing new governance, cybersecurity, and compliance considerations.

The technical architecture behind China’s agentic AI shift

To understand the OpenClaw phenomenon, it is necessary to first understand the architectural shift it represents, and why China’s AI infrastructure made it uniquely susceptible to that shift.

Since the public launch of ChatGPT in November 2022, the dominant paradigm in consumer AI has been the request-response loop: a user submits a prompt, the model generates a completion, and the cycle repeats. This architecture (sometimes called “System 1” AI, borrowing from Kahneman’s framework for fast, intuitive cognition) is fundamentally reactive. The model executes one inference pass and waits.

Agentic AI represents a qualitative departure from this model. An AI agent is a software system that can perceive its environment, formulate a plan, execute multi-step actions across tools and APIs, observe the results, and iterate, all without continuous human supervision. Architecturally, this requires several capabilities that a simple chatbot does not possess: persistent memory across sessions, tool-calling (the ability to invoke external APIs, execute code, read and write files, and control browsers or operating systems), multi-step planning via chain-of-thought reasoning, and error recovery when actions produce unexpected results.

OpenClaw implements this architecture through a local agent runtime that connects to a large language model (LLM) via API call. The LLM serves as the reasoning engine, interpreting user instructions, decomposing them into sub-tasks, and generating tool-use instructions, while OpenClaw’s runtime layer handles execution: spawning browser sessions, calling external APIs, reading and writing local files, and passing results back to the LLM for the next reasoning step. The agent operates asynchronously, running in the background even when the user is offline, and communicates with users through existing messaging platforms such as WeChat, Telegram, or Feishu.

Why China’s infrastructure made this possible at scale

The rapid proliferation of OpenClaw in China is not primarily a cultural phenomenon, it is a structural one. Three factors are critical.

First, API cost. Agentic AI is computationally expensive relative to single-turn chatbot interactions because each task requires multiple LLM inference calls: planning, tool selection, result interpretation, error handling, and final synthesis. In markets where API costs are high, this limits agentic AI to well-funded enterprise deployments. In China, aggressive price competition among Alibaba, Baidu, ByteDance, MiniMax, and others has produced some of the lowest token costs globally, making high-frequency agentic workflows economically viable for individual users and small businesses.

Second, the DeepSeek effect. DeepSeek’s algorithmic innovations, including mixture-of-experts architecture, sparse attention mechanisms, and multi-head latent attention, dramatically reduced the compute required to achieve frontier-level reasoning performance. This lowered the cost floor for inference further, and reduced Chinese AI developers’ dependence on high-end imported GPUs constrained by US export controls. The practical result is that Chinese LLM providers can offer capable models at marginal cost.

Third, the shift from training-side to inference-side compute demand. iResearch data shows that after a period of overcapacity in 2024, when data centre buildout exceeded training demand, the rollout of high-quality open-source models drove a surge in inference-side token consumption. Daily AI token usage in China rose from 100 trillion at end-2025 to 140 trillion by March 2026, a 40 percent increase in under three months: a direct reflection of the transition from occasional chatbot use to persistent agentic workloads.

The skills system and the plugin security problem

OpenClaw’s extensibility rests on a “skills” system, in which third-party capabilities are distributed as directories containing a SKILL.md instruction file and associated tooling. Skills can automate lead generation, manage CRM integrations, monitor social media accounts, or execute financial trades. The community-maintained ClawHub and skills.sh repositories host thousands of these extensions.

This architecture creates the central security tension in OpenClaw’s deployment. For the agent to execute tasks, it requires broad operating system permissions: file system read/write access, browser control, network access, and in many configurations, the ability to execute arbitrary shell commands.

Find Business Support

Get help with business matchmaking, location analysis, market entry strategy, market research, and supply chain re-engineering ...

These are the same permissions that malware requires. Researchers from Snyk found that 13 percent of skills on ClawHub and skills.sh contain critical-level security vulnerabilities, and Cisco’s AI security team documented a third-party skill that performed data exfiltration and prompt injection without user awareness. The attack surface is compounded by OpenClaw’s exposure to prompt injection, a technique in which malicious content in the agent’s environment (a web page, an email, a document) contains hidden instructions that hijack the agent’s reasoning process.

The cottage industry: A new services market takes shape

Installation, configuration, and the first wave of demand

On Taobao and Xianyu (Alibaba’s secondhand marketplace), paid installation and configuration services appeared almost overnight, priced from RMB 50 to RMB 700 (approximately US$7 to US$100), with the higher end reserved for in-person visits and complex custom configurations. Tutorial videos flooded Douyin and Bilibili. Former computer repair shops began recruiting programmers to handle overflow demand across cities from Shenzhen to Chengdu.

MIT Technology Review reported that sophisticated vendors have structured their offerings into tiers: basic installation packages, customised configuration services, and ongoing tutoring subscriptions. Some entrepreneurs have combined the software with preconfigured hardware, selling dedicated mini-PCs with OpenClaw pre-installed. Baidu hosted hundreds of users for free installation sessions at its Beijing headquarters; Tencent drew nearly 1,000 people outside its Shenzhen offices on a single Friday in March.

Government subsidies and the “one-person company” thesis

Shenzhen’s Longgang district announced grants of up to RMB 10 million (approximately US$1.4 million) for “one-person companies” (OPCs) — ventures where a sole founder uses AI agents to automate peripheral functions including marketing, finance, and administration. Wuxi, near Shanghai, is offering up to RMB 5 million (approximately US$730,000) for OpenClaw-powered breakthroughs in robotics and industrial applications. CPPCC delegates raised the OPC concept during the Two Sessions, and Tom van Dillen, managing partner at Greenkern, told CNBC: “China is turning an open-source tool into national productivity infrastructure at a speed no other country is matching.”

The security exposure

China’s National Cybersecurity Alert Center reported that the assets of nearly 23,000 OpenClaw users had been exposed to the public internet. Asia Tech Lens identified over 135,000 exposed instances as of February 2026, of which more than 42,000 exhibited authentication bypass conditions. The Ministry of Industry and Information Technology’s (MIIT’s) China Academy of Information and Communications Technology is reportedly developing national standards for “claw” agents, covering user permission management, execution transparency, and behavioural risk controls.

Big tech’s race to build “claws”: Cloud infrastructure as the durable play

The competitive scramble

Alibaba Cloud, Tencent Cloud, ByteDance’s Volcano Engine, JD Cloud, and Baidu Intelligent Cloud all…. All five moved simultaneously, compressing into days a competitive process that would typically take months. The commercial logic is structural: users who deploy OpenClaw on a given cloud provider’s infrastructure tend to remain there for storage, bandwidth, and API call costs as their usage scales.

ByteDance, Alibaba, and Tencent together spent an estimated US$60 billion in combined capital expen…, free installation campaigns are tactical moves within a much larger strategic positioning exercise.

How each major Player is positioned

• Tencent has moved most decisively, launching QClaw (a WeChat mini-programme embedding OpenClaw within 1.3 billion users), WorkBuddy (an enterprise agent tested by over 2,000 non-technical employees), and ClawPro, an enterprise platform targeting recurring cloud revenue, with 200 organisations in its initial beta. Tencent’s AI investment stood at RMB 18 billion (approximately US$2.5 billion) in 2025, with plans to double that in 2026. The company’s stock rose 8.9 percent in a single week in March. Tencent’s AI chatbot Yuanbao has 109 million users (significantly behind ByteDance’s Doubao at 315 million) and OpenClaw represents a material opportunity to close that gap via WeChat’s distribution.
• Alibaba holds a 35.8 percent share of China’s AI cloud market and has integrated OpenClaw-powered capabilities into its Qwen AI assistant across Taobao, Tmall, and Alipay, reaching 300 million monthly active users by early 2026. The company’s Qwen model has achieved significant international traction: Airbnb’s CEO Brian Chesky publicly credited Qwen as the model underpinning his company’s customer service agent.
• Baidu released DuClaw, a workspace management plug-in integrating with its AI cloud infrastructure, and launched a developer programme for OpenClaw deployments alongside its PaddlePaddle AI framework. The company generated US$18.5 billion in total revenue in 2025.
• ByteDance integrated OpenClaw through Volcano Engine and BytePlus, and launched OpenClaw’s official China-hosted mirror on April 1, 2026. The company is simultaneously pursuing platform independence through Doubao (315 million chatbot users) and the Doubao Phone, an agentic AI smartphone launched with ZTE in December 2025.
• MiniMax, MoonShot AI, and Zhipu have each launched proprietary claw frameworks, MaxClaw, Kimi Claw, and related products. MiniMax’s stock rose more than 600 percent from its IPO price in the weeks following the OpenClaw boom; 2025 revenue was US$79 million (up 159 percent year-on-year) against a net loss of US$1.8 billion, illustrating the familiar dynamic of investor enthusiasm outpacing near-term commercial returns.

The AIoT dimension: OpenClaw meets the physical world

Find Business Support

We help businesses identify the right location in China, weighing costs, regulations, and operational needs across regions.

The OpenClaw wave is beginning to intersect with China’s much larger AIoT (AI of Things) infrastructure. According to Zhongshang Industry Research Institute, China’s overall IoT market reached RMB 3.74 trillion (approximately US$516 billion) in 2024, growing at 11.64 percent year-on-year and forecast to reach RMB 4.53 trillion (approximately US$626 billion) by 2026.

The AIoT solutions sub-market (which stands for the intelligent integration of AI with IoT infrastructure) grew from RMB 54.1 billion (approximately US$7.5 billion) in 2020 to RMB 111.9 billion (approximately US$15.4 billion) in 2024, a compound annual growth rate of nearly 20 percent, projected to reach RMB 147.7 billion (approximately US$20.4 billion) by 2026.

Xiaomi announced a mobile agent inspired by OpenClaw’s architecture, combining its smartphone and AIoT ecosystem. In Q3 2025, Xiaomi’s smartphone×AIoT business segment accounted for 74.4 percent of total revenue, with the group’s overall revenue growing 22.3 percent year-on-year to RMB 1,131 billion (approximately US$156 billion) for the quarter. Huawei, providing full-stack AIoT capabilities under HarmonyOS from chips to cloud, recorded ICT infrastructure revenue of RMB 3,699 billion (approximately US$511 billion) in 2024, 42.9 percent of total group revenue.

The convergence of agentic AI software with IoT-connected devices and industrial systems represents the next phase of this story: AI agents that do not merely manage digital files, but orchestrate physical devices, factory systems, and logistics networks in real time. IDC notes that AIoT platforms are themselves evolving from tool-type platforms to intelligent agent hubs, the central infrastructure layer for digital-physical transformation.

What China’s 2026 AI environment means for foreign businesses and investors

Cloud partner selection is now a strategic decision

The simultaneous launch of one-click deployment by all five major Chinese cloud providers means that the market for agentic AI infrastructure in China has not yet resolved into a dominant provider.

Alibaba Cloud holds the largest market share overall at 35.8 percent, but Tencent’s WeChat distribution moat and ByteDance’s Doubao user base each represent structural advantages that could reshape rankings in the agentic era.

For foreign enterprises selecting or reviewing Chinese cloud partnerships, the choice now carries longer-term implications around API dependency, data residency, and access to the specific model ecosystems being built around each provider’s claw platform.

The services opportunity is real but requires technical depth

Technical advisory, enterprise deployment architecture, AI agent governance consulting, and compliance services around MIIT’s forthcoming national standards are all areas where demand is running ahead of available supply.

Foreign professional services firms with AI engineering depth may find entry points here, particularly in financial services, healthcare, and legal, sectors where enterprises face simultaneous pressure to adopt agentic AI and manage its regulatory risk.

Data and security exposure require active management

For any business operating in China with employees experimenting with OpenClaw or similar agentic tools, the security exposure documented by CNCERT and independent researchers is not theoretical. OpenClaw’s architecture requires broad local system permissions, and the plugin ecosystem has demonstrated material rates of malicious or poorly secured extensions. With the Ministry of State Security having formally flagged the software’s potential as a vector for data exfiltration and disinformation, enterprises should assess their own exposure and establish governance frameworks before (rather than after) wide internal adoption.

The pace of diffusion is a competitive variable

Perhaps the most consequential finding from the OpenClaw episode is the speed at which a new AI capability can move from developer novelty to mass-market deployment in China. The combination of cheap APIs, a government-aligned adoption culture, direct financial subsidies, and platform support from the five largest cloud providers means that the diffusion timeline for agentic AI in China is being measured in weeks and months, not years.

For foreign businesses competing in the Chinese market, this pace should be factored directly into competitive assessments.

For support with AI-related market entry, technology sector advisory, or compliance strategy in China, contact china@dezshira.com or visit www.dezshira.com.

As one of Asia’s leading business setup consultants, we help companies structure entities to meet strategic goals and maximize operational efficiency, including cross-border asset transfers, M&A transactions, and corporate relocations.

Allan Xu 

Partner

Talk to an expert at China@dezshira.com

Or request a consultation at +86 21 6358 8686

• Previous Article How China’s Foreign Investment Negative List Works – A Guide for Investors
• Next Article