How a Crayfish Bot Is Reshaping China's Tech Stack

Opening

Dear subscriber, on March 6th, nearly a thousand people lined up outside Tencent’s headquarters in Shenzhen, China. Developers clutching MacBooks, office workers carrying NAS drives, even a 9-year-old elementary school student. What they were waiting for wasn’t a new iPhone or a limited-edition figure. It was a single red crayfish.

It’s an open-source AI agent called OpenClaw. The phenomenon now shaking all of China under the name “quanmin yangxia” (全民养虾, “the whole nation raises crayfish”) reads, on the surface, like just another “AI tool goes viral” story. But what’s happening underneath is far more structural.

Today, let’s talk about how this one crayfish is simultaneously reshaping China’s cloud market, its messaging ecosystem, and its AI model competition.

What’s With the Crayfish: The Identity of OpenClaw

Before diving in, let’s clarify what this “crayfish” actually is. OpenClaw is an open-source AI agent framework released in late 2025 by Austrian developer Peter Steinberger. It was originally named Clawdbot, but because of its similarity to Anthropic’s Claude, it went through two renamings before settling on its current name.

There’s one decisive difference from existing AI tools like ChatGPT, Claude, or Gemini. Those tools “converse.” OpenClaw, as an agent, “works.” It installs directly on the user’s computer, reads files, sends emails, writes code, and operates browsers. When instructed through a messenger app like WhatsApp or Telegram, it functions like an AI employee working around the clock without rest.

Because the project’s icon is a red crayfish, and because installing and training it feels similar to feeding and raising a crayfish, it picked up the nickname “yangxia” (养虾, “raising crayfish”). Within 100 days of launch, it set an all-time record for GitHub stars, and of the 142,000 public instances confirmed on tracking platforms, nearly half came from China. Now, this crayfish is causing three major waves in China.

First Wave: Cloud Companies’ New Business

The most direct beneficiaries of the crayfish craze are the cloud companies.

Every time OpenClaw runs, it calls a large language model. Since it’s performing a “task” rather than a single conversation, its token¹ consumption is on a completely different scale from an ordinary chatbot. There are reports of 7 million tokens being consumed for a simple research task, 29 million tokens for a single crawler test, and cases of 50 million tokens being used when run all day. Using it properly for a month adds up to about 100 million tokens, which translates to roughly 7,000 yuan (about 1.3 million KRW) in cost.

For model companies and cloud providers, this means a structural explosion in token consumption. An agent burns tens of thousands to hundreds of thousands of tokens in a single task. The consumption pattern itself is completely different from a person chatting.

The results are showing up in the numbers. As of February 2026, daily token consumption for MiniMax’s M2 series models increased more than 6x compared to December 2025. Moonshot AI’s Kimi K2.5 saw cumulative revenue surpass its entire 2025 revenue within just 20 days of launch. As of February, MiniMax’s annualized recurring revenue (ARR) surpassed $150 million.

Tencent Cloud, Alibaba Cloud, Baidu AI Cloud, and even Volcano Engine (火山引擎, ByteDance’s cloud arm) all competitively rolled out one-click deployment services dedicated to OpenClaw. Tencent Cloud’s lightweight server Lighthouse surpassed 100,000 OpenClaw cloud users, and Alibaba Cloud launched a coding plan starting at 7.9 yuan (about 1,500 KRW) per month.

Let’s go one step deeper here. Until now, the AI billing model was simple: pay-as-you-go based on tokens used. But in the agent era, this structure is being shaken. Even for the same “process my email” task, a simple summary might take a few thousand tokens, but analyzing attachments and drafting a reply can burn through hundreds of thousands. From the user’s perspective, it’s “just one email handled,” yet the cost can differ by 100x. As this gap widens, the very basis of billing has to change. We’re moving past an era of paying for how much AI is used, toward an era of paying based on what kind of work AI handles. Cloud companies competitively rolling out separate “coding plans” and “agent plans” is an early signal of this shift.

This isn’t unique to China. The same logic applies to global cloud providers like AWS and Cloudflare. Agents have begun to require “agent-centric infrastructure” that goes beyond human-centric infrastructure. More calls, longer sessions, more compute — for cloud companies, this is clearly new business.

Second Wave: Cracks in WeChat’s Impregnable Fortress

China’s mobile ecosystem has operated solidly for over a decade around the super-app WeChat (微信). With over 1 billion monthly active users, it’s a platform that encompasses messaging, payments, social networking, and commerce all at once. It’s far beyond the position Korea’s KakaoTalk holds — effectively, it’s the operating system of digital life in China. If you’ve ever traveled to China, or have Chinese friends, you’d know this app’s influence firsthand. But the OpenClaw craze is now creating cracks in this structure.

OpenClaw’s core interface is the messenger app. The structure is that you give instructions to the AI through a messenger app you’re already using — WhatsApp, Telegram, Slack, Discord. The problem is that in China, access to these foreign messenger apps is difficult due to the Great Firewall (GFW)².

To fill this gap, Tencent’s QQ was the first to support OpenClaw. QQ is a platform with a more mature bot ecosystem than WeChat. Following that, the business messenger WeCom (企业微信) and ByteDance’s Feishu (飞书) also began releasing official OpenClaw plugins. There were also reports that Tencent released a product called QClaw, testing a way to remotely control OpenClaw across both WeChat and QQ.

Here’s an interesting reversal. Recall what happened in December 2025 when ByteDance launched its Doubao (豆包) phone assistant. When this agent tried to manipulate WeChat, WeChat forcibly logged users out within 48 hours. Taobao triggered CAPTCHA challenges, and financial apps issued security warnings. ByteDance ultimately disabled the WeChat manipulation feature and backed off. (And this was despite both being Chinese companies!)

Yet three months later, when OpenClaw appeared, the reaction of China’s tech ecosystem was the exact opposite. People lined up in person outside Tencent’s headquarters to help with installation. Companies actively integrated it into their own platforms. Why? Because OpenClaw isn’t a specific company’s product — it’s an open-source community project. Since no company can claim it as “theirs,” no company has grounds to exclude it either.

As a result, a situation has emerged where platforms like QQ, Feishu, and DingTalk (钉钉) are rising up around a new competitive axis called “agent interface” within what was once a WeChat-centered, single-messaging ecosystem. According to multiple local sources, this reignition of messenger competition within China’s mobile ecosystem is highly unusual.

Third Wave: Global Spotlight on Chinese AI Models

The third wave may carry the most long-term significance.

The data from OpenRouter (the world’s largest LLM API aggregation platform) is striking. As of February 24, 2026, Chinese models accounted for 61% of total token consumption among the platform’s top 10 models. All top 3 models were Chinese. MiniMax M2.5 ranked first with 2.45 trillion tokens per week, Moonshot AI’s Kimi K2.5 ranked second with 1.21 trillion tokens, and Zhipu AI’s (智谱) GLM-5 ranked third with 780 billion tokens. This overturned, all at once, a position American models had held for a long time.

The key factor that made this possible is price. MiniMax M2.5’s input token cost is $0.3 per million tokens. Anthropic’s Claude Opus 4.6 costs $5 for the same measure — a difference of about 16.7x. The gap widens further for output tokens: MiniMax M2.5 costs $1.1, while Claude Opus 4.6 costs $25, nearly a 23x difference.

Ironically, behind this price competitiveness lies US technology restrictions against China. Due to export limits on cutting-edge GPUs, Chinese AI companies had no choice but to focus on installable, on-device, and on-premise models rather than cloud-API-centric ones. MiniMax’s M2.5 has 229 billion total parameters, but uses a Mixture of Experts (MoE)³ architecture where only about 10 billion are actually activated during inference. It’s a method for extracting high performance from less computation.

As a result, with the agent economy now in full swing, price has become the decisive variable in agent workflows that run autonomously for long periods, consuming tens of thousands to hundreds of thousands of tokens. OpenRouter COO Chris Clark stated that Chinese open-weight models are taking “a disproportionately large share of the agent flows run by US companies.” a16z partner Martin Casado estimated that about 80% of startups using the open-source AI stack are running Chinese models. For more details, please refer to our previous newsletter, <Silicon Valley Is Quietly Using Chinese AI>.

On the SWE-Bench Verified benchmark, MiniMax M2.5 scores 80.2% and Claude Opus 4.6 scores 80.8%. A performance gap of just 0.6 percentage points, against a price gap of 16-23x. In an era where agents autonomously call APIs hundreds of times, this price gap translates directly into market share.

Oz’s Lens

Honestly, when I first saw this phenomenon, I also thought, “I guess AI agents are trending again.” But the more I dug into the data, the more I realized this isn’t just a trend — it’s a moment revealing how the agent economy actually works. Whenever a new technology emerges, early on there’s always the imagination of “what can we do with this?” But what actually changes the market isn’t the technology itself, but the cost structure and distribution channels that technology creates.

OpenClaw demonstrates three things. First, agents burn incomparably more tokens than chatbots — this fundamentally changes the revenue models of cloud and model companies. Second, the agent’s interface is the messenger app — this shifts the axis of platform competition from “number of users” to “agent compatibility.” Third, in the agent era, it’s not the cost of a single call but the cumulative cost of tens of thousands of calls that determines model choice — this means Chinese models’ price competitiveness could be a structural advantage, not simple dumping.

Of course, there are clearly overheated aspects to the current craze. Security vulnerabilities are already surfacing one after another, and the Ministry of Industry and Information Technology (MIIT) has issued security warnings twice. There are reports of 824 malicious skills being identified, and 135,000 OpenClaw instances exposed on the open internet. It’s a classic pattern of “technology diffusion outpacing security.”

But even after the bubble deflates, something will remain. What this craze has proven is that demand for agents is real. And the infrastructure competition to meet that demand has only just begun. In fact, NVIDIA has released NemoClaw, its own open-source foundation model with completed optimizations, specifically to join this competition.

Closing

To summarize: China’s crayfish craze isn’t simply a case of one tool going viral — it’s a signal of a structural transition where the agent economy is simultaneously reshaping three axes at once: cloud, messaging, and AI models.

The key question in this transition isn’t “which agent is better?” It’s “what infrastructure, what interface, and what cost structure does an agent need in order to function?” And the answer to that question is being written in real time in China right now.

For us watching this phenomenon from Korea, there are implications too. Where will KakaoTalk stand as an agent interface? Is Korea’s cloud infrastructure ready to handle agent demand? What level of agent compatibility do Korean AI models have? Now is the time to mull over these three questions.

References & Further Reading

• “OpenClaw龙虾：为何火爆？谁最受益？”, 36Kr, March 2026: The article that best summarizes the changing revenue structures of model companies and cloud providers.
• “OpenClaw Conquered China in 100 Days”, HelloChinaTech, March 2026: A structural analysis contrasting ByteDance’s Doubao episode with OpenClaw’s reception.
• “Chinese AI Models Capture 61% of Global Token Volume”, Wealthari, February 2026: An article summarizing Chinese models’ global share based on OpenRouter data.
• “Chinese Models Top OpenRouter Token Rankings”, Pandaily, February 2026: An analysis of how agent scenarios are changing token consumption patterns.
• “Tencent Moves to Bring OpenClaw AI Assistant Into WeChat”, Caixin Global, March 2026: Covers Tencent’s QClaw strategy and the background of its WeChat integration.
• “MiniMax Stock Surges Past Baidu in Market Cap”, CTOL Digital Solutions, March 2026: Analyzes the background of MiniMax surpassing Baidu’s market cap and the valuation logic of the agent economy.
• “China issues second warning on OpenClaw risks amid adoption frenzy”, South China Morning Post, March 2026: SCMP’s coverage of security risks and government response.

You might ask how I came to know this story — the Go-to-Market consulting firm I run, OBF, has Chinese clients including ByteDance and MiniMax. I’ve heard a lot from these companies through casual conversations, and I think this issue is bigger than expected, yet it isn’t being covered domestically. If Korean media or interested readers reach out, I’m happy to share more in-depth stories, so feel free to contact me. (Interviews, coverage, and PR with these companies are also possible — I can make the connections.)

I try to visit places like Shanghai and Shenzhen, and deliberately check the trending posts on Bilibili, Douyin, and Weibo (think of them as YouTube, TikTok, and Twitter respectively), turning on translation as I go. The pace of change is faster than expected, and there’s a time lag before it reaches us. If you want to see fast-moving tech trends, it seems like now is the time to widen your view beyond the US to include China as well.

Going forward, stories like this will be released for membership subscribers. As a reference, membership costs 3,000 KRW per month, and gives unlimited access to articles older than 30 days. Of course, free subscribers still have no trouble accessing the latest articles.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Token: the basic unit that an AI model uses to process text. A single word, part of a word, or a punctuation mark can each be one token. Roughly speaking, for Korean text, 1 million tokens can process about 500–750 pages of A4-sized text. ↩

2. Great Firewall (GFW): an internet censorship system operated by the Chinese government. It blocks access within China to foreign services like Google, Facebook, WhatsApp, and Telegram. ↩

3. Mixture of Experts (MoE): a type of AI model architecture in which only some “expert” networks among the total parameters are selectively activated. While the overall model size is large, the actual computation used during inference is small, allowing large-model performance at a lower cost. ↩