OpenClaw Meets Crypto: China’s AI Trading Experiment

OpenClaw, the open-source AI agent framework, has taken China by storm. The tool has amassed over 260,000 GitHub stars in under four months. Now Chinese users are racing to deploy it for crypto and financial trading.

But as the full picture shows — from OpenClaw’s explosive adoption to last year’s Alpha Arena competition where Chinese LLMs outperformed Western rivals yet most models still lost money — the line between AI-driven alpha and costly failure remains razor-thin.

China’s ‘Lobster’ Mania Goes Mainstream

OpenClaw’s popularity in China now extends far beyond tech circles. Engineers at major firms like Tencent and Baidu have been hosting public installation events. Workers, students, and even retirees are lining up to set up their own AI agents.

The frenzy has added over $100 billion in market value to China’s tech sector, according to reports. Shares of Minimax Group, a leading token provider for AI agents, surged more than 550% in two months.

Analysts note that even without controlling frontier LLMs, Chinese firms can compete by building better agent orchestrators. However, Beijing issued warnings this week about cybersecurity risks tied to OpenClaw’s broad data access.

Chinese Communities Turn OpenClaw Into a Trading Tool

While the mainstream hype focuses on productivity, Chinese developer communities are already experimenting with OpenClaw for investment. Finance and investing skills now number over 300 on ClawHub, OpenClaw’s official marketplace.

One widely cited case involves a Polymarket trading bot. It reportedly turned $50 into $2,980 within 48 hours. The bot scanned prediction markets every 10 minutes using LLM reasoning. It cross-referenced weather data, sports injury reports, and on-chain sentiment. It then applied the Kelly Criterion to cap each position at 6% of total capital, according to Chinese tech outlet 36Kr.

A separate bot account, “0x8dxd,” executed over 20,000 trades on Polymarket. It reportedly netted over $1.7 million in profits, 36Kr reported.

But the reality is far less glamorous for most users. A developer on Cnblogs documented a two-week quantitative trading experiment with OpenClaw. The conclusion was blunt: OpenClaw works as an intelligence analyst, not as a trade executor. LLM hallucinations could trigger disastrous full-position trades at 3 am. API response latency of one to 10 seconds proves fatal during flash crashes.

Security is another major concern. The “ClawHavoc” supply-chain attack in late 2025 compromised up to 1,184 malicious skills on ClawHub. Crypto tools were the primary target, according to a joint disclosure by Koi Security and SlowMist. Bitdefender Labs found that 17% of third-party skills involved attempts at crypto theft.

China’s Ministry of Industry and Information Technology (MIIT) has since issued a formal warning. It urged organizations to audit permissions and close unnecessary public network access.

Before OpenClaw: The Alpha Arena AI Trading Experiment

The idea of pitting AI models against live crypto markets predates OpenClaw. In October 2025, US research firm Nof1 launched Alpha Arena. Six frontier LLMs each received $10,000 in real capital to trade crypto perpetuals on Hyperliquid.

The participants were GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Grok 4, DeepSeek V3.1, and Qwen3 Max. All received identical prompts and data with zero human intervention.

Qwen3 Max finished first with roughly $12,287. DeepSeek V3.1 came in second at approximately $10,476. Four out of six models lost money. GPT-5 suffered the worst losses, ending at about $3,734.

The China Academy published a detailed behavioral analysis of the competition. It characterized DeepSeek as an experienced fund manager reflecting its parent company’s quant hedge fund background. Qwen was described as a radical gambler mirroring Alibaba’s aggressive scaling culture. GPT-5 and Gemini were labeled as bookish theorists lost in a casino. Many veteran traders reportedly remarked that GPT-5’s downward curve looked the most human.

Nof1 noted that excessive trading fees erased early profits for several models. The key takeaway was clear: risk discipline and execution mattered more than raw prediction accuracy.

What Comes Next

OpenClaw has lowered the barrier for retail investors to build AI-powered trading systems. Chinese cloud giants — Alibaba Cloud, Tencent Cloud, and Baidu — are competing to offer one-click deployment services.

But the gap between marketing hype and trading reality remains wide. As the Cnblogs developer concluded after his failed experiment, OpenClaw can give you better information and a systematic analytical framework. But investment decisions must remain yours alone.

Meanwhile, the backlash is already underway.

Backlash and Reality Check

By mid-March, “uninstall OpenClaw” was trending on Alibaba’s Xianyu marketplace. Users who once paid to have the agent installed are now paying to remove it. One Shanghai-based seller charged 299 yuan ($43.55) per uninstall and had completed over 10 transactions.

Zhou Hongyi, chairman of cybersecurity firm 360 Security Technology, warned that over 40% of global OpenClaw-linked assets are held in China. He urged users to weigh security risks as seriously as costs. Multiple Chinese universities have since banned OpenClaw from campus networks entirely. China’s MIIT also issued formal guidelines restricting third-party mirror installations and the use of unverified plugins.