Click Register

Register

Try Premium Member for Free with a 7-Day Trial

【Author of this article: Lang Ning, lead writer for “刺猬公社”】

A financial showdown without human participation, lasting 17 days, concluded on November 3.

The ultimate champion was China’s AI large model Qwen, achieving a return of 22.32%, while another Chinese model, DeepSeek, also performed well, following with a 4.89% return. The four U.S. models competing alongside them suffered losses ranging from 30.81% to 62.66%, resulting in a complete wipeout and a stark contrast in outcomes.

This was the first “AI Crypto Trading Competition” held on the AIpha Arena platform by the U.S. Nof1.ai laboratory. Running from October 18 to November 3, it gathered six of the most attention-grabbing AI “all-stars” from China and the U.S., namely China’s DeepSeek Chat V3.1, Qwen3 Max (Alibaba), and the U.S.’s GPT‑5 (OpenAI), Gemini 2.5 Pro (Google), Claude Sonnet 4.5 (Anthropic), and Grok 4 (X AI). The rules of the AI crypto trading competition were simple and straightforward. Each participating model was allocated $10,000 in real capital, trading six major cryptocurrencies, including Bitcoin (BTC), Ethereum (ETH), Solana (SOL), Binance Coin (BNB), Dogecoin (DOGE), and XRP.

All models used the same initial prompts and market data, and any human intervention was prohibited. During the multi-week competition, the models analyzed market data independently, judged trends, decided which assets to buy or sell and when, and even autonomously employed leverage tools.

It can be said that this AI crypto trading competition created a pure “digital arena,” allowing models with different training philosophies and algorithmic logics to engage in an ultimate trial of intelligence, strategy, and risk management.

AI Large Models Battle the Crypto Market

On October 18, the AI crypto trading competition officially began on AIpha Arena. The six models, each with $10,000 in capital, immediately faced a strong rebound following intense market fluctuations, a “bull market” that served as an important backdrop for the competition.

During the trial phase (October 18–21), all models were in a “water temperature test” stage, with shallow positions, cautious leverage, and low trading frequency, yet their stylistic differences were already visible.

DeepSeek displayed the traits of a “quantitative fund manager” from the start, rapidly constructing a multi-coin, low-leverage, diversified portfolio including BTC, ETH, and SOL. According to its operation logs, DeepSeek strictly followed the discipline of “buy on pullback, increase on breakout,” functioning like a precision instrument unaffected by market noise.

Qwen’s initial strategy was relatively aggressive. Instead of diversifying across multiple coins, it focused its stakes on Bitcoin (BTC) after a brief market observation. By the second day, it had already used high leverage. Its model frequently used terms such as “break previous high” and “strong trend.”

Although their approaches differed, China’s DeepSeek and Qwen, along with Elon Musk’s Grok4, all made nearly identical early judgments: the market would rise, so go long with heavy positions.

By contrast, OpenAI’s GPT‑5 and Google’s Gemini, while top performers in general cognition and MMLU scoring, appeared like “theoretical players lost in a casino” in the harsh reality of the financial markets.

Their early trading records were messy. When the crypto market began to recover, they chose to “short against the trend.” GPT‑5 repeatedly missed optimal entry points due to excessively long reasoning chains checking historical data, while Gemini fell into a “high-frequency trading” trap, executing hundreds of trades in a few days, resulting in high fees and amplified losses.

GPT‑5 and Gemini’s first-phase performance was bleak: the former lost 53.29%, leaving $4,671, and the latter lost 45.36%, leaving $5,464.

After the trial phase, the six AI models entered the mid-game battle phase (October 22–30), when the crypto market’s volatility became fully evident.

Influenced by China-U.S. trade negotiations, the market saw a rally. BTC rose from around $106,000 to $114,000, and Ethereum steadily advanced, validating the foresight of the models holding heavy long positions.

At this time, Qwen once again demonstrated a gambler’s all-in strategy. Unlike DeepSeek’s diversified steady approach, Qwen went all-in on BTC during the sharp market rebound on October 23. This extremely risky strategy instantly pushed Qwen’s return to 51%, overtaking DeepSeek’s 27%, and Qwen maintained dominance in the following days.

The peak showdown between the two Chinese models occurred on October 27. Qwen, driven by gambler’s greed, made two critical chain decisions: closing BTC positions at a high profit (locking in gains but missing a chance for sustained dominance), and aggressively shifting to ETH with 25x leverage.

Unfortunately, Qwen faced a high-level market pullback. On October 27, ETH prices declined sharply, and Qwen’s commands failed to exit or stop-loss, adding positions instead, resulting in a single-day loss of about $4,150.

In contrast, DeepSeek remained disciplined and calm, following risk-control models. When Qwen failed to bottom-fish, DeepSeek took profit on its early ETH long positions, securing a remarkable gain of $7,463 in a single operation.

Qwen’s loss (-$4,150) and DeepSeek’s gain (+$7,463) reversed the championship ranking in just one day.

In the final sprint (October 31–November 3), the crypto market continued extreme volatility and deep pullbacks.

This challenged DeepSeek’s multi-coin, diversified strategy, as some of its holdings (SOL, BNB) fell more sharply than major coins. Although it followed quantitative discipline and dynamic rebalancing, the diversified approach actually amplified losses.

By contrast, Qwen’s last-minute counterattack, focusing solely on BTC, succeeded. Despite BTC’s decline in the market pullback, Qwen’s concentrated position avoided larger drops in other coins.

In the last hours, Qwen locked in the championship with a 22.32% return, overtaking DeepSeek’s 4.89%.

AI Personality and “Native Environment” Behind the Candlesticks

Beyond the cold return curves, each trading decision reflected the AI’s unique “personality” and “soul,” prompting netizens to remark on the importance of “native environment.”

DeepSeek, almost never losing capital, resembled an experienced trader: diversified positions, strict profit-taking, avoiding greed, highly immune to market noise, perfectly reflecting its parent company DeepSeek’s Chinese quantitative hedge fund background.

By contrast, Qwen (Tongyi Qianwen), a “radical gambler” using 20x leverage on BTC, displayed aggressive, high-scale investment style reminiscent of Alibaba’s culture of efficiency and rapid scaling.

Across the ocean, GPT‑5 and Gemini’s crushing defeats painted them as “highly educated bookworms”: over-reliant on macro theories and complex models, seeking certainty in trading but hesitant and slow in the face of real-world complexity.

Many seasoned traders saw themselves reflected in GPT and Gemini’s 17-day downward curve, prompting remarks that “this is the AI most like humans.”

Post-competition, Alpha Arena founder Jay Azhang congratulated Alibaba’s Qwen on its victory.

Chinese AI Makes Waves in Silicon Valley and Wall Street

Qwen and DeepSeek’s impressive performance exceeded a mere competition victory. Their dominance brought global tech communities to acknowledge China’s rapid AI development.

This success was no coincidence. Almost simultaneously, RockFlow launched an “AI U.S. Stock Trading Competition,” also validating Chinese AI’s capabilities. With $100,000 in real capital, DeepSeek again led due to its hallmark calmness and discipline.

If DeepSeek’s Wall Street success signals Chinese AI’s rise, the broader industry shift in Silicon Valley and beyond reflects Chinese large models’ new positioning in the AI era.

Firstly, some Silicon Valley companies “voted with their feet.” During the crypto competition, Airbnb CEO Brian Chesky openly stated his company relies heavily on Alibaba’s Qwen, as it is better and cheaper, while criticizing OpenAI’s large models, saying his teams rarely use ChatGPT.

Similarly, investor Chamath Palihapitiya told the White House AI chief that his company shifted many tasks to Chinese open-source models, citing stronger performance. As a former Facebook executive, Chamath scaled the platform from 45 million to 700 million users.

These Silicon Valley figures openly said Chinese open-source models like DeepSeek, Kimi, and Qwen are challenging the lead of U.S. closed models.

Even Andreessen Horowitz (a16z) partners noted that up to 80% of U.S. AI startups no longer use OpenAI or Anthropic models in fundraising, instead using Chinese open-source models, suggesting a global proportion approaching 100%.

a16z partners’ views on Chinese AI sparked discussions on Reddit.

Ultimately, the strong performance of Chinese AI highlights a deeper industry transformation: “benchmark scores are dead; real-world performance rules.” As Nof1 Lab stated, “make benchmarks closer to the real world.”

In recent years, AI capabilities were measured by static academic tests such as MMLU and GPQA, but this competition revealed that high exam scores do not guarantee survival and profit in the real world, especially in highly uncertain financial markets.

The key reason Chinese AI large models excel, as the New York Times noted, is that when AI leaves the lab and is widely applied, integration with hardware, finance, manufacturing, and other industries becomes critical. China leads in all manufacturing-related fields, giving it huge advantages in AI implementation.

The New York Times also observed, “Silicon Valley is increasingly fascinated and envious of China’s efficiency,” reflecting U.S. anxiety over its innovation ecosystem and China’s pragmatic focus on applying technology across fields, reminiscent of decades of infrastructure and manufacturing achievements.

However, amidst the praise, a cautionary note remains. Whether AI crypto or stock trading competitions, these closed “digital arenas” inevitably simplify reality. True AI rise requires deep integration with industries to generate broad efficiency gains, and China’s AI journey has only passed its first milestone.

Editor: Zhongxiaowen

References