When a technology generates its own folk vocabulary, adoption has passed the early-adopter phase. In China, the term is “养虾” — raising a lobster — and it refers to running an always-on OpenClaw agent on a dedicated second computer. The phrase, now mainstream enough for Xinhua coverage, captures something the software hype skipped: autonomous AI agents need their own hardware, and millions of Chinese users figured that out before the industry did.

The Slang and What It Describes

The term emerged organically in March 2026 as OpenClaw adoption surged in China. “Raising a lobster” describes a specific setup: one machine for your normal work, another machine running your agent 24/7. The second machine — nicknamed a “龙虾窝” (lobster nest) — handles the agent’s web browsing, file manipulation, app control, and continuous background processes while keeping those high-permission operations isolated from the user’s primary workstation and personal data.

A PandaYoo analysis published March 29 compiled reporting from Xinhua, China News Service, NBD, The Paper, and 36Kr to argue that this behavior represents a structural shift: the AI execution revolution has outpaced consumer hardware, creating a market gap for purpose-built “agent boxes.” The analysis notes that while the popular narrative focused on productivity and autonomy, the lived reality was more concrete — two computers, recurring token bills, and growing infrastructure costs.

Hardware Demand: Selling Out in Shenzhen and Beijing

The dedicated-hardware pattern is creating real supply pressure. The South China Morning Post reported that Mac Minis were selling out across China, with Beijing electronics sellers in Zhongguancun charging mark-ups of at least 500 yuan (US$73) above the standard 4,499-yuan price. Wait times at official Apple Stores expanded to at least a month. In Shenzhen’s Huaqiangbei — the world’s largest electronics wholesale market — vendors had largely run out of stock, according to state-owned Securities Times reporting cited by SCMP.

Chinese startups moved fast to fill the gap. Sina News reported that domestic teams launched pre-installed OpenClaw mini PCs — purpose-built “lobster nests” — and sold nearly 100 units within days. Device refurbishment company YiDianYun positioned its remanufactured hardware as the cost-effective substrate for enterprise AI agent deployments, with the company’s DefectGPT system running 81 automated checks per device at a claimed 99.9% outgoing quality rate.

The Cost Problem Driving Hardware Isolation

The hardware demand is partly a security calculation and partly an economic one. Xinhua’s reporting, originally published via China Youth Daily, documented the cost realities that users are confronting:

  • A 31-year-old software engineer named Zheng Nan reported spending approximately $1,000 just to optimize business-specific skills for his agent
  • Monthly compute costs run in the hundreds of yuan even for moderate use
  • OpenClaw’s architecture carries heavy system-prompt context on every model call — the agent’s persona files, behavior rules, and skill descriptions ship with each request, making even simple tasks token-expensive
  • The platform’s heartbeat mechanism triggers periodic wake-ups to check for pending tasks, consuming tokens even when the user has assigned no work

360 founder Zhou Hongyi, speaking at a March 12 media briefing cited in the same Xinhua report, framed the distinction: “If DeepSeek helped the public understand what large models are, OpenClaw helped them understand what agents are.” He argued that the agent layer represents the correct path for AI deployment because large models have hit scaling-law bottlenecks, while agents bypass those limits through tool composition and multi-agent collaboration.

Why the Vocabulary Matters

The emergence of “养虾” as a mainstream term signals that OpenClaw adoption in China has crossed from developer experimentation into mass consumer behavior. The Xinhua report notes GitHub daily downloads exceeding 200,000, and lists ByteDance (ArkClaw), Tencent (WorkBuddy), Alibaba (CoPaw), Zhipu (AutoClaw), and Moonshot (KimiClaw) as major firms rushing to ship competing agent products.

But the slang also encodes a problem. “Raising” implies ongoing cost and maintenance. Chinese users discovered that running an autonomous agent is closer to maintaining livestock than installing an app — it requires feeding (tokens), housing (dedicated hardware), and constant upkeep (skill optimization). The “first batch of lobster-raisers paying to uninstall” became a trending topic on Chinese social media, with China’s Ministry of Industry and Information Technology issuing warnings that even updated versions of OpenClaw do not eliminate security risks, according to the Xinhua report.

PandaYoo’s analysis argues this pattern will repeat globally as agent adoption scales: software excitement collides with hardware constraints, and the market shifts from “which model is best?” to “where does this run, who controls it, and what hardware posture makes it survivable?” The dedicated agent computer — cheap, quiet, always-on, isolated from personal data — may be the first new consumer hardware category that AI agents have created.